Clear supported_models list to keep only supported models

* Update pillow

* relaxation of pillow update

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -39,7 +39,7 @@ body:
     attributes:
       label: FastEmbed version
       description: What version of FastEmbed are you running? python -c "import fastembed; print(fastembed.__version__)". If you're not on the latest, please upgrade and see if the problem persists.
-      placeholder: v0.4.2
+      placeholder: v0.5.1
     validations:
       required: true
   - type: dropdown

.github/workflows/python-tests.yml

@@ -14,11 +14,11 @@ jobs:
     strategy:
       matrix:
         python-version:
-          - '3.8.x'
           - '3.9.x'
           - '3.10.x'
           - '3.11.x'
           - '3.12.x'
+          - '3.13.x'
         os:
           - ubuntu-latest
           - macos-latest
@@ -38,7 +38,7 @@ jobs:
         run: |
           python -m pip install poetry
           poetry config virtualenvs.create false
-          poetry install --no-interaction --no-ansi --without docs
+          poetry install --no-interaction --no-ansi --without dev,docs
 
       - name: Run pytest
         run: |

NOTICE

@@ -7,6 +7,8 @@ This distribution includes the following Jina AI models, each with its respectiv
     - License: cc-by-nc-4.0
 - jinaai/jina-reranker-v2-base-multilingual
     - License: cc-by-nc-4.0
+- jinaai/jina-embeddings-v3
+    - License: cc-by-nc-4.0
 
 These models are developed by Jina (https://jina.ai/) and are subject to Jina AI's licensing terms.
 

README.md

@@ -28,10 +28,10 @@ pip install fastembed-gpu
 
 ```python
 from fastembed import TextEmbedding
-from typing import List
+
 
 # Example list of documents
-documents: List[str] = [
+documents: list[str] = [
     "This is built to be faster and lighter than other embedding libraries e.g. Transformers, Sentence-Transformers, etc.",
     "fastembed is supported by and maintained by Qdrant.",
 ]
@@ -139,11 +139,10 @@ embeddings = list(model.embed(images))
 
 ### 🔄 Rerankers
 ```python
-from typing import List
 from fastembed.rerank.cross_encoder import TextCrossEncoder
 
 query = "Who is maintaining Qdrant?"
-documents: List[str] = [
+documents: list[str] = [
     "This is built to be faster and lighter than other embedding libraries e.g. Transformers, Sentence-Transformers, etc.",
     "fastembed is supported by and maintained by Qdrant.",
 ]

docs/Getting Started.ipynb

@@ -65,15 +65,13 @@
     }
    ],
    "source": [
-    "from typing import List\n",
-    "\n",
     "import numpy as np\n",
     "\n",
     "from fastembed import TextEmbedding\n",
     "\n",
     "\n",
     "# Example list of documents\n",
-    "documents: List[str] = [\n",
+    "documents: list[str] = [\n",
     "    \"This is built to be faster and lighter than other embedding libraries e.g. Transformers, Sentence-Transformers, etc.\",\n",
     "    \"fastembed is supported by and maintained by Qdrant.\",\n",
     "]\n",

docs/examples/ColBERT_with_FastEmbed.ipynb

@@ -54,7 +54,14 @@
     },
     {
      "data": {
-      "text/plain": "[{'model': 'colbert-ir/colbertv2.0',\n  'dim': 128,\n  'description': 'Late interaction model',\n  'size_in_GB': 0.44,\n  'sources': {'hf': 'colbert-ir/colbertv2.0'},\n  'model_file': 'model.onnx'}]"
+      "text/plain": [
+       "[{'model': 'colbert-ir/colbertv2.0',\n",
+       "  'dim': 128,\n",
+       "  'description': 'Late interaction model',\n",
+       "  'size_in_GB': 0.44,\n",
+       "  'sources': {'hf': 'colbert-ir/colbertv2.0'},\n",
+       "  'model_file': 'model.onnx'}]"
+      ]
      },
      "execution_count": 1,
      "metadata": {},
@@ -212,7 +219,9 @@
    "outputs": [
     {
      "data": {
-      "text/plain": "((26, 128), (32, 128))"
+      "text/plain": [
+       "((26, 128), (32, 128))"
+      ]
      },
      "execution_count": 18,
      "metadata": {},
@@ -271,7 +280,9 @@
     "import numpy as np\n",
     "\n",
     "\n",
-    "def compute_relevance_scores(query_embedding: np.array, document_embeddings: np.array, k: int):\n",
+    "def compute_relevance_scores(\n",
+    "    query_embedding: np.array, document_embeddings: np.array, k: int\n",
+    ") -> list[int]:\n",
     "    \"\"\"\n",
     "    Compute relevance scores for top-k documents given a query.\n",
     "\n",

docs/examples/FastEmbed_GPU.ipynb

@@ -388,8 +388,6 @@
     }
    ],
    "source": [
-    "from typing import List\n",
-    "\n",
     "import numpy as np\n",
     "\n",
     "from fastembed import TextEmbedding\n",
@@ -407,9 +405,7 @@
     "id": "iPtoHf7GeV-i"
    },
    "outputs": [],
-   "source": [
-    "documents: List[str] = list(np.repeat(\"Demonstrating GPU acceleration in fastembed\", 500))"
-   ]
+   "source": "documents: list[str] = list(np.repeat(\"Demonstrating GPU acceleration in fastembed\", 500))"
   },
   {
    "cell_type": "code",

docs/examples/FastEmbed_Multi_GPU.ipynb

@@ -0,0 +1,88 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Fastembed Multi-GPU Tutorial\n",
+    "This tutorial demonstrates how to leverage multi-GPU support in Fastembed. Fastembed supports embedding text and images utilizing modern GPUs for acceleration. Let's explore how to use Fastembed with multiple GPUs step by step."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "#### Prerequisites\n",
+    "To get started, ensure you have the following installed:\n",
+    "- Python 3.9 or later\n",
+    "- Fastembed (`pip install fastembed-gpu`)\n",
+    "- Refer to [this](https://github.com/qdrant/fastembed/blob/main/docs/examples/FastEmbed_GPU.ipynb) tutorial if you have issues with GPU dependencies\n",
+    "- Access to a multi-GPU server"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Multi-GPU using cuda argument with TextEmbedding Model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from fastembed import TextEmbedding\n",
+    "\n",
+    "# define the documents to embed\n",
+    "docs = [\"hello world\", \"flag embedding\"] * 100\n",
+    "\n",
+    "# define gpu ids\n",
+    "device_ids = [0, 1]\n",
+    "\n",
+    "if __name__ == \"__main__\":\n",
+    "    # initialize a TextEmbedding model using CUDA\n",
+    "    text_model = TextEmbedding(\n",
+    "        model_name=\"sentence-transformers/all-MiniLM-L6-v2\",\n",
+    "        cuda=True,\n",
+    "        device_ids=device_ids,\n",
+    "        lazy_load=True,\n",
+    "    )\n",
+    "\n",
+    "    # generate embeddings\n",
+    "    text_embeddings = list(text_model.embed(docs, batch_size=2, parallel=len(device_ids)))\n",
+    "    print(text_embeddings)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this snippet:\n",
+    "- `cuda=True` enables GPU acceleration.\n",
+    "- `device_ids=[0, 1]` specifies GPUs to use. Replace `[0, 1]` with available GPU IDs.\n",
+    "- `lazy_load=True`\n",
+    "\n",
+    "**NOTE**: When using multi-GPU settings, it is important to configure `parallel` and `lazy_load` properly to avoid inefficiencies:\n",
+    "\n",
+    "`parallel`: This parameter enables multi-GPU support by spawning child processes for each GPU specified in device_ids. To ensure proper utilization, the value of `parallel` must match the number of GPUs in device_ids. If using a single GPU, this parameter is not necessary.\n",
+    "\n",
+    "`lazy_load`: Enabling `lazy_load` prevents redundant memory usage. Without `lazy_load`, the model is initially loaded into the memory of the first GPU by the main process. When child processes are spawned for each GPU, the model is reloaded on the first GPU, causing redundant memory consumption and inefficiencies."
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": ".venv",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python",
+   "version": "3.10.15"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

docs/examples/FastEmbed_vs_HF_Comparison.ipynb

@@ -36,7 +36,7 @@
    "outputs": [],
    "source": [
     "import time\n",
-    "from typing import Callable, List, Tuple\n",
+    "from typing import Callable\n",
     "\n",
     "import matplotlib.pyplot as plt\n",
     "import torch.nn.functional as F\n",
@@ -64,6 +64,7 @@
    ],
    "source": [
     "import fastembed\n",
+    "\n",
     "fastembed.__version__"
    ]
   },
@@ -98,7 +99,7 @@
     }
    ],
    "source": [
-    "documents: List[str] = [\n",
+    "documents: list[str] = [\n",
     "    \"Chandrayaan-3 is India's third lunar mission\",\n",
     "    \"It aimed to land a rover on the Moon's surface - joining the US, China and Russia\",\n",
     "    \"The mission is a follow-up to Chandrayaan-2, which had partial success\",\n",
@@ -151,11 +152,11 @@
     "    HuggingFace Transformer implementation of FlagEmbedding\n",
     "    \"\"\"\n",
     "\n",
-    "    def __init__(self, model_id: str):\n",
+    "    def __init__(self, model_id: str) -> None:\n",
     "        self.model = AutoModel.from_pretrained(model_id)\n",
     "        self.tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
     "\n",
-    "    def embed(self, texts: List[str]):\n",
+    "    def embed(self, texts: list[str]):\n",
     "        encoded_input = self.tokenizer(\n",
     "            texts, max_length=512, padding=True, truncation=True, return_tensors=\"pt\"\n",
     "        )\n",
@@ -254,7 +255,7 @@
     "\n",
     "def calculate_time_stats(\n",
     "    embed_func: Callable, documents: list, k: int\n",
-    ") -> Tuple[float, float, float]:\n",
+    ") -> tuple[float, float, float]:\n",
     "    times = []\n",
     "    for _ in range(k):\n",
     "        # Timing the embed_func call\n",
@@ -309,7 +310,7 @@
    ],
    "source": [
     "def plot_character_per_second_comparison(\n",
-    "    hf_stats: Tuple[float, float, float], fst_stats: Tuple[float, float, float], documents: list\n",
+    "    hf_stats: tuple[float, float, float], fst_stats: tuple[float, float, float], documents: list\n",
     "):\n",
     "    # Calculating total characters in documents\n",
     "    total_characters = sum(len(doc) for doc in documents)\n",

docs/examples/Hindi_Tamil_RAG_with_Navarasa7B.ipynb

@@ -44,7 +44,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 21,
+   "execution_count": null,
    "metadata": {
     "ExecuteTime": {
      "end_time": "2024-03-30T00:45:24.814968Z",
@@ -58,8 +58,6 @@
    },
    "outputs": [],
    "source": [
-    "from typing import List\n",
-    "\n",
     "import numpy as np\n",
     "from datasets import load_dataset\n",
     "from peft import AutoPeftModelForCausalLM\n",
@@ -72,11 +70,11 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 23,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
-    "hf_token = <YOUR_HF_TOKEN_HERE>  # Get your token from https://huggingface.co/settings/token, needed for Gemma weights"
+    "hf_token = \"<YOUR_HF_TOKEN_HERE>\"  # Get your token from https://huggingface.co/settings/token, needed for Gemma weights"
    ]
   },
   {
@@ -246,7 +244,7 @@
    },
    "outputs": [],
    "source": [
-    "context_embeddings: List[np.ndarray] = list(\n",
+    "context_embeddings: list[np.ndarray] = list(\n",
     "    embedding_model.embed(contexts)\n",
     ")  # Note the list() call - this is a generator"
    ]

docs/examples/Hybrid_Search.ipynb

@@ -50,7 +50,6 @@
    "outputs": [],
    "source": [
     "import json\n",
-    "from typing import List, Tuple\n",
     "\n",
     "import numpy as np\n",
     "import pandas as pd\n",
@@ -489,11 +488,11 @@
     }
    ],
    "source": [
-    "def make_sparse_embedding(texts: List[str]):\n",
+    "def make_sparse_embedding(texts: list[str]) -> list[SparseEmbedding]:\n",
     "    return list(sparse_model.embed(texts, batch_size=32))\n",
     "\n",
     "\n",
-    "sparse_embedding: List[SparseEmbedding] = make_sparse_embedding(\n",
+    "sparse_embedding: list[SparseEmbedding] = make_sparse_embedding(\n",
     "    [\"Fastembed is a great library for text embeddings!\"]\n",
     ")\n",
     "sparse_embedding"
@@ -616,7 +615,7 @@
     }
    ],
    "source": [
-    "def get_tokens_and_weights(sparse_embedding, model_name):\n",
+    "def get_tokens_and_weights(sparse_embedding, model_name) -> dict[str, float]:\n",
     "    # Find the tokenizer for the model\n",
     "    tokenizer_source = None\n",
     "    for model_info in SparseTextEmbedding.list_supported_models():\n",
@@ -627,7 +626,7 @@
     "            raise ValueError(f\"Model {model_name} not found in the supported models.\")\n",
     "\n",
     "    tokenizer = AutoTokenizer.from_pretrained(tokenizer_source)\n",
-    "    token_weight_dict = {}\n",
+    "    token_weight_dict: dict[str, float] = {}\n",
     "    for i in range(len(sparse_embedding.indices)):\n",
     "        token = tokenizer.decode([sparse_embedding.indices[i]])\n",
     "        weight = sparse_embedding.values[i]\n",
@@ -662,7 +661,7 @@
    },
    "outputs": [],
    "source": [
-    "def make_dense_embedding(texts: List[str]):\n",
+    "def make_dense_embedding(texts: list[str]):\n",
     "    return list(dense_model.embed(texts))\n",
     "\n",
     "\n",
@@ -872,7 +871,7 @@
    },
    "outputs": [],
    "source": [
-    "def make_points(df: pd.DataFrame) -> List[PointStruct]:\n",
+    "def make_points(df: pd.DataFrame) -> list[PointStruct]:\n",
     "    sparse_vectors = df[\"sparse_embedding\"].tolist()\n",
     "    product_texts = df[\"combined_text\"].tolist()\n",
     "    dense_vectors = df[\"dense_embedding\"].tolist()\n",
@@ -899,7 +898,7 @@
     "    return points\n",
     "\n",
     "\n",
-    "points: List[PointStruct] = make_points(df)"
+    "points: list[PointStruct] = make_points(df)"
    ]
   },
   {
@@ -942,8 +941,8 @@
    "source": [
     "def search(query_text: str):\n",
     "    # # Compute sparse and dense vectors\n",
-    "    query_sparse_vectors: List[SparseEmbedding] = make_sparse_embedding([query_text])\n",
-    "    query_dense_vector: List[np.ndarray] = make_dense_embedding([query_text])\n",
+    "    query_sparse_vectors: list[SparseEmbedding] = make_sparse_embedding([query_text])\n",
+    "    query_dense_vector: list[np.ndarray] = make_dense_embedding([query_text])\n",
     "\n",
     "    search_results = client.search_batch(\n",
     "        collection_name=collection_name,\n",
@@ -1075,7 +1074,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "def rank_list(search_result: List[ScoredPoint]):\n",
+    "def rank_list(search_result: list[ScoredPoint]):\n",
     "    return [(point.id, rank + 1) for rank, point in enumerate(search_result)]\n",
     "\n",
     "\n",
@@ -1149,7 +1148,7 @@
    ],
    "source": [
     "def find_point_by_id(\n",
-    "    client: QdrantClient, collection_name: str, rrf_rank_list: List[Tuple[int, float]]\n",
+    "    client: QdrantClient, collection_name: str, rrf_rank_list: list[tuple[int, float]]\n",
     "):\n",
     "    return client.retrieve(\n",
     "        collection_name=collection_name, ids=[item[0] for item in rrf_rank_list]\n",

docs/examples/SPLADE_with_FastEmbed.ipynb

@@ -47,7 +47,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "metadata": {
     "ExecuteTime": {
      "end_time": "2024-03-30T00:49:20.516644Z",
@@ -56,8 +56,7 @@
    },
    "outputs": [],
    "source": [
-    "from fastembed import SparseTextEmbedding, SparseEmbedding\n",
-    "from typing import List"
+    "from fastembed import SparseTextEmbedding, SparseEmbedding"
    ]
   },
   {
@@ -134,7 +133,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": null,
    "metadata": {
     "ExecuteTime": {
      "end_time": "2024-03-30T00:49:28.624109Z",
@@ -143,7 +142,7 @@
    },
    "outputs": [],
    "source": [
-    "documents: List[str] = [\n",
+    "documents: list[str] = [\n",
     "    \"Chandrayaan-3 is India's third lunar mission\",\n",
     "    \"It aimed to land a rover on the Moon's surface - joining the US, China and Russia\",\n",
     "    \"The mission is a follow-up to Chandrayaan-2, which had partial success\",\n",
@@ -157,7 +156,7 @@
     "    \"Chandrayaan-3 was launched from the Satish Dhawan Space Centre in Sriharikota\",\n",
     "    \"Chandrayaan-3 was launched earlier in the year 2023\",\n",
     "]\n",
-    "sparse_embeddings_list: List[SparseEmbedding] = list(\n",
+    "sparse_embeddings_list: list[SparseEmbedding] = list(\n",
     "    model.embed(documents, batch_size=6)\n",
     ")  # batch_size is optional, notice the generator"
    ]
@@ -235,7 +234,9 @@
    "source": [
     "# Let's print the first 5 features and their weights for better understanding.\n",
     "for i in range(5):\n",
-    "    print(f\"Token at index {sparse_embeddings_list[0].indices[i]} has weight {sparse_embeddings_list[0].values[i]}\")"
+    "    print(\n",
+    "        f\"Token at index {sparse_embeddings_list[0].indices[i]} has weight {sparse_embeddings_list[0].values[i]}\"\n",
+    "    )"
    ]
   },
   {
@@ -261,7 +262,9 @@
     "import json\n",
     "from transformers import AutoTokenizer\n",
     "\n",
-    "tokenizer = AutoTokenizer.from_pretrained(SparseTextEmbedding.list_supported_models()[0][\"sources\"][\"hf\"])"
+    "tokenizer = AutoTokenizer.from_pretrained(\n",
+    "    SparseTextEmbedding.list_supported_models()[0][\"sources\"][\"hf\"]\n",
+    ")"
    ]
   },
   {
@@ -326,7 +329,9 @@
     "        token_weight_dict[token] = weight\n",
     "\n",
     "    # Sort the dictionary by weights\n",
-    "    token_weight_dict = dict(sorted(token_weight_dict.items(), key=lambda item: item[1], reverse=True))\n",
+    "    token_weight_dict = dict(\n",
+    "        sorted(token_weight_dict.items(), key=lambda item: item[1], reverse=True)\n",
+    "    )\n",
     "    return token_weight_dict\n",
     "\n",
     "\n",

docs/index.md

@@ -26,14 +26,14 @@ pip install fastembed
 ```python
 from fastembed import TextEmbedding
 
-documents: List[str] = [
+documents: list[str] = [
     "passage: Hello, World!",
     "query: Hello, World!",
     "passage: This is an example passage.",
     "fastembed is supported by and maintained by Qdrant."
 ]
 embedding_model = TextEmbedding()
-embeddings: List[np.ndarray] = embedding_model.embed(documents)
+embeddings: list[np.ndarray] = embedding_model.embed(documents)
 ```
 
 ## Usage with Qdrant

docs/qdrant/Retrieval_with_FastEmbed.ipynb

@@ -41,7 +41,6 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from typing import List\n",
     "import numpy as np\n",
     "from fastembed import TextEmbedding"
    ]
@@ -71,7 +70,7 @@
    ],
    "source": [
     "# Example list of documents\n",
-    "documents: List[str] = [\n",
+    "documents: list[str] = [\n",
     "    \"Maharana Pratap was a Rajput warrior king from Mewar\",\n",
     "    \"He fought against the Mughal Empire led by Akbar\",\n",
     "    \"The Battle of Haldighati in 1576 was his most famous battle\",\n",
@@ -87,7 +86,7 @@
     "embedding_model = TextEmbedding(model_name=\"BAAI/bge-small-en\")\n",
     "\n",
     "# We'll use the passage_embed method to get the embeddings for the documents\n",
-    "embeddings: List[np.ndarray] = list(\n",
+    "embeddings: list[np.ndarray] = list(\n",
     "    embedding_model.passage_embed(documents)\n",
     ")  # notice that we are casting the generator to a list\n",
     "\n",

docs/qdrant/Usage_With_Qdrant.ipynb

@@ -46,7 +46,6 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from typing import List\n",
     "from qdrant_client import QdrantClient"
    ]
   },
@@ -67,7 +66,7 @@
    "outputs": [],
    "source": [
     "# Example list of documents\n",
-    "documents: List[str] = [\n",
+    "documents: list[str] = [\n",
     "    \"Maharana Pratap was a Rajput warrior king from Mewar\",\n",
     "    \"He fought against the Mughal Empire led by Akbar\",\n",
     "    \"The Battle of Haldighati in 1576 was his most famous battle\",\n",
@@ -199,7 +198,9 @@
     }
    ],
    "source": [
-    "search_result = client.query(collection_name=\"demo_collection\", query_text=\"This is a query document\")\n",
+    "search_result = client.query(\n",
+    "    collection_name=\"demo_collection\", query_text=\"This is a query document\"\n",
+    ")\n",
     "print(search_result)"
    ]
   },

experiments/01_ONNX_Port.ipynb

@@ -19,7 +19,7 @@
    "outputs": [],
    "source": [
     "from pathlib import Path\n",
-    "from typing import List, Tuple, Any\n",
+    "from typing import Any\n",
     "\n",
     "import numpy as np\n",
     "import time\n",
@@ -91,9 +91,11 @@
     "    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]\n",
     "\n",
     "\n",
-    "def hf_embed(model_id: str, inputs: List[str]):\n",
+    "def hf_embed(model_id: str, inputs: list[str]):\n",
     "    # Tokenize the input texts\n",
-    "    batch_dict = hf_tokenizer(inputs, max_length=512, padding=True, truncation=True, return_tensors=\"pt\")\n",
+    "    batch_dict = hf_tokenizer(\n",
+    "        inputs, max_length=512, padding=True, truncation=True, return_tensors=\"pt\"\n",
+    "    )\n",
     "\n",
     "    outputs = hf_model(**batch_dict)\n",
     "    embeddings = average_pool(outputs.last_hidden_state, batch_dict[\"attention_mask\"])\n",
@@ -133,7 +135,9 @@
     "optimization_config = AutoOptimizationConfig.O4()\n",
     "optimizer = ORTOptimizer.from_pretrained(model)\n",
     "\n",
-    "optimizer.optimize(save_dir=save_dir, optimization_config=optimization_config, use_external_data_format=True)\n",
+    "optimizer.optimize(\n",
+    "    save_dir=save_dir, optimization_config=optimization_config, use_external_data_format=True\n",
+    ")\n",
     "model = ORTModelForFeatureExtraction.from_pretrained(save_dir)\n",
     "\n",
     "tokenizer.save_pretrained(save_dir)\n",
@@ -171,7 +175,9 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "def measure_pipeline_time(pipeline, input_texts: List[str], num_runs=10, **kwargs: Any) -> Tuple[float, float]:\n",
+    "def measure_pipeline_time(\n",
+    "    pipeline, input_texts: list[str], num_runs=10, **kwargs: Any\n",
+    ") -> tuple[float, float]:\n",
     "    \"\"\"Measures the time it takes to run the pipeline on the input texts.\"\"\"\n",
     "    times = []\n",
     "    total_chars = sum(len(text) for text in input_texts)\n",

fastembed/common/model_management.py

@@ -1,29 +1,37 @@
 import os
 import time
+import json
 import shutil
 import tarfile
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Optional
 
 import requests
-from huggingface_hub import snapshot_download
-from huggingface_hub.utils import RepositoryNotFoundError
+from huggingface_hub import snapshot_download, model_info, list_repo_tree
+from huggingface_hub.hf_api import RepoFile
+from huggingface_hub.utils import (
+    RepositoryNotFoundError,
+    disable_progress_bars,
+    enable_progress_bars,
+)
 from loguru import logger
 from tqdm import tqdm
 
 
 class ModelManagement:
+    METADATA_FILE = "files_metadata.json"
+
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         raise NotImplementedError()
 
     @classmethod
-    def _get_model_description(cls, model_name: str) -> Dict[str, Any]:
+    def _get_model_description(cls, model_name: str) -> dict[str, Any]:
         """
         Gets the model description from the model_name.
 
@@ -34,7 +42,7 @@ class ModelManagement:
             ValueError: If the model_name is not supported.
 
         Returns:
-            Dict[str, Any]: The model description.
+            dict[str, Any]: The model description.
         """
         for model in cls.list_supported_models():
             if model_name.lower() == model["model"].lower():
@@ -93,22 +101,74 @@ class ModelManagement:
     def download_files_from_huggingface(
         cls,
         hf_source_repo: str,
-        cache_dir: Optional[str] = None,
-        extra_patterns: Optional[List[str]] = None,
+        cache_dir: str,
+        extra_patterns: list[str],
         local_files_only: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ) -> str:
         """
         Downloads a model from HuggingFace Hub.
         Args:
             hf_source_repo (str): Name of the model on HuggingFace Hub, e.g. "qdrant/all-MiniLM-L6-v2-onnx".
             cache_dir (Optional[str]): The path to the cache directory.
-            extra_patterns (Optional[List[str]]): extra patterns to allow in the snapshot download, typically
+            extra_patterns (list[str]): extra patterns to allow in the snapshot download, typically
                 includes the required model files.
             local_files_only (bool, optional): Whether to only use local files. Defaults to False.
+            specific_model_path (Optional[str], optional): The path to the model dir already pooled from external source
         Returns:
             Path: The path to the model directory.
         """
+
+        def _verify_files_from_metadata(
+            model_dir: Path, stored_metadata: dict[str, Any], repo_files: list[RepoFile]
+        ) -> bool:
+            try:
+                for rel_path, meta in stored_metadata.items():
+                    file_path = model_dir / rel_path
+
+                    if not file_path.exists():
+                        return False
+
+                    if repo_files:  # online verification
+                        file_info = next((f for f in repo_files if f.path == file_path.name), None)
+                        if (
+                            not file_info
+                            or file_info.size != meta["size"]
+                            or file_info.blob_id != meta["blob_id"]
+                        ):
+                            return False
+
+                    else:  # offline verification
+                        if file_path.stat().st_size != meta["size"]:
+                            return False
+                return True
+            except (OSError, KeyError) as e:
+                logger.error(f"Error verifying files: {str(e)}")
+                return False
+
+        def _collect_file_metadata(
+            model_dir: Path, repo_files: list[RepoFile]
+        ) -> dict[str, dict[str, int]]:
+            meta = {}
+            file_info_map = {f.path: f for f in repo_files}
+            for file_path in model_dir.rglob("*"):
+                if file_path.is_file() and file_path.name != cls.METADATA_FILE:
+                    repo_file = file_info_map.get(file_path.name)
+                    if repo_file:
+                        meta[str(file_path.relative_to(model_dir))] = {
+                            "size": repo_file.size,
+                            "blob_id": repo_file.blob_id,
+                        }
+            return meta
+
+        def _save_file_metadata(model_dir: Path, meta: dict[str, dict[str, int]]) -> None:
+            try:
+                if not model_dir.exists():
+                    model_dir.mkdir(parents=True, exist_ok=True)
+                (model_dir / cls.METADATA_FILE).write_text(json.dumps(meta))
+            except (OSError, ValueError) as e:
+                logger.warning(f"Error saving metadata: {str(e)}")
+
         allow_patterns = [
             "config.json",
             "tokenizer.json",
@@ -116,10 +176,59 @@ class ModelManagement:
             "special_tokens_map.json",
             "preprocessor_config.json",
         ]
-        if extra_patterns is not None:
-            allow_patterns.extend(extra_patterns)
 
-        return snapshot_download(
+        allow_patterns.extend(extra_patterns)
+
+        snapshot_dir = Path(cache_dir) / f"models--{hf_source_repo.replace('/', '--')}"
+        metadata_file = snapshot_dir / cls.METADATA_FILE
+
+        if local_files_only:
+            disable_progress_bars()
+            if metadata_file.exists():
+                metadata = json.loads(metadata_file.read_text())
+                verified = _verify_files_from_metadata(snapshot_dir, metadata, repo_files=[])
+                if not verified:
+                    logger.warning(
+                        "Local file sizes do not match the metadata."
+                    )  # do not raise, still make an attempt to load the model
+            else:
+                logger.warning(
+                    "Metadata file not found. Proceeding without checking local files."
+                )  # if users have downloaded models from hf manually, or they're updating from previous versions of
+                # fastembed
+            result = snapshot_download(
+                repo_id=hf_source_repo,
+                allow_patterns=allow_patterns,
+                cache_dir=cache_dir,
+                local_files_only=local_files_only,
+                **kwargs,
+            )
+            return result
+
+        repo_revision = model_info(hf_source_repo).sha
+        repo_tree = list(list_repo_tree(hf_source_repo, revision=repo_revision, repo_type="model"))
+
+        allowed_extensions = {".json", ".onnx", ".txt"}
+        repo_files = (
+            [
+                f
+                for f in repo_tree
+                if isinstance(f, RepoFile) and Path(f.path).suffix in allowed_extensions
+            ]
+            if repo_tree
+            else []
+        )
+
+        verified_metadata = False
+
+        if snapshot_dir.exists() and metadata_file.exists():
+            metadata = json.loads(metadata_file.read_text())
+            verified_metadata = _verify_files_from_metadata(snapshot_dir, metadata, repo_files)
+
+        if verified_metadata:
+            disable_progress_bars()
+
+        result = snapshot_download(
             repo_id=hf_source_repo,
             allow_patterns=allow_patterns,
             cache_dir=cache_dir,
@@ -127,8 +236,26 @@ class ModelManagement:
             **kwargs,
         )
 
+        if (
+            not verified_metadata
+        ):  # metadata is not up-to-date, update it and check whether the files have been
+            # downloaded correctly
+            metadata = _collect_file_metadata(snapshot_dir, repo_files)
+
+            download_successful = _verify_files_from_metadata(
+                snapshot_dir, metadata, repo_files=[]
+            )  # offline verification
+            if not download_successful:
+                raise ValueError(
+                    "Files have been corrupted during downloading process. "
+                    "Please check your internet connection and try again."
+                )
+            _save_file_metadata(snapshot_dir, metadata)
+
+        return result
+
     @classmethod
-    def decompress_to_cache(cls, targz_path: str, cache_dir: str):
+    def decompress_to_cache(cls, targz_path: str, cache_dir: str) -> str:
         """
         Decompresses a .tar.gz file to a cache directory.
 
@@ -211,13 +338,13 @@ class ModelManagement:
 
     @classmethod
     def download_model(
-        cls, model: Dict[str, Any], cache_dir: Path, retries: int = 3, **kwargs
+        cls, model: dict[str, Any], cache_dir: Path, retries: int = 3, **kwargs: Any
     ) -> Path:
         """
         Downloads a model from HuggingFace Hub or Google Cloud Storage.
 
         Args:
-            model (Dict[str, Any]): The model description.
+            model (dict[str, Any]): The model description.
                 Example:
                 ```
                 {
@@ -238,6 +365,9 @@ class ModelManagement:
             Path: The path to the downloaded model directory.
         """
         local_files_only = kwargs.get("local_files_only", False)
+        specific_model_path: Optional[str] = kwargs.pop("specific_model_path", None)
+        if specific_model_path:
+            return Path(specific_model_path)
         retries = 1 if local_files_only else retries
         hf_source = model.get("sources", {}).get("hf")
         url_source = model.get("sources", {}).get("url")
@@ -265,6 +395,8 @@ class ModelManagement:
                             f"Could not download model from HuggingFace: {e} "
                             "Falling back to other sources."
                         )
+                finally:
+                    enable_progress_bars()
             if url_source or local_files_only:
                 try:
                     return cls.retrieve_model_gcs(

fastembed/common/onnx_model.py

@@ -1,7 +1,7 @@
 import warnings
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Any, Dict, Generic, Iterable, Optional, Sequence, Tuple, Type, TypeVar
+from typing import Any, Generic, Iterable, Optional, Sequence, Type, TypeVar
 
 import numpy as np
 import onnxruntime as ort
@@ -33,8 +33,8 @@ class OnnxModel(Generic[T]):
         self.tokenizer = None
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
@@ -52,6 +52,13 @@ class OnnxModel(Generic[T]):
         model_path = model_dir / model_file
         # List of Execution Providers: https://onnxruntime.ai/docs/execution-providers
 
+        if cuda and providers is not None:
+            warnings.warn(
+                f"`cuda` and `providers` are mutually exclusive parameters, cuda: {cuda}, providers: {providers}",
+                category=UserWarning,
+                stacklevel=6,
+            )
+
         if providers is not None:
             onnx_providers = list(providers)
         elif cuda:
@@ -96,7 +103,7 @@ class OnnxModel(Generic[T]):
     def load_onnx_model(self) -> None:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def onnx_embed(self, *args, **kwargs) -> OnnxOutputContext:
+    def onnx_embed(self, *args: Any, **kwargs: Any) -> OnnxOutputContext:
         raise NotImplementedError("Subclasses must implement this method")
 
 
@@ -105,7 +112,7 @@ class EmbeddingWorker(Worker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ) -> OnnxModel:
         raise NotImplementedError()
 
@@ -113,7 +120,7 @@ class EmbeddingWorker(Worker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model = self.init_embedding(model_name, cache_dir, **kwargs)
 
@@ -121,5 +128,5 @@ class EmbeddingWorker(Worker):
     def start(cls, model_name: str, cache_dir: str, **kwargs: Any) -> "EmbeddingWorker":
         return cls(model_name=model_name, cache_dir=cache_dir, **kwargs)
 
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
         raise NotImplementedError("Subclasses must implement this method")

fastembed/common/preprocessor_utils.py

@@ -1,6 +1,6 @@
 import json
 from pathlib import Path
-from typing import Tuple
+
 from tokenizers import AddedToken, Tokenizer
 
 from fastembed.image.transform.operators import Compose
@@ -17,7 +17,7 @@ def load_special_tokens(model_dir: Path) -> dict:
     return tokens_map
 
 
-def load_tokenizer(model_dir: Path) -> Tuple[Tokenizer, dict]:
+def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict]:
     config_path = model_dir / "config.json"
     if not config_path.exists():
         raise ValueError(f"Could not find config.json in {model_dir}")

fastembed/common/types.py

@@ -1,7 +1,7 @@
-import os
+from pathlib import Path
 import sys
 from PIL import Image
-from typing import Any, Dict, Iterable, Tuple, Union
+from typing import Any, Iterable, Union
 
 if sys.version_info >= (3, 10):
     from typing import TypeAlias
@@ -9,8 +9,8 @@ else:
     from typing_extensions import TypeAlias
 
 
-PathInput: TypeAlias = Union[str, os.PathLike]
+PathInput: TypeAlias = Union[str, Path]
 PilInput: TypeAlias = Union[Image.Image, Iterable[Image.Image]]
 ImageInput: TypeAlias = Union[PathInput, Iterable[PathInput], PilInput]
 
-OnnxProvider: TypeAlias = Union[str, Tuple[str, Dict[Any, Any]]]
+OnnxProvider: TypeAlias = Union[str, tuple[str, dict[Any, Any]]]

fastembed/common/utils.py

@@ -1,16 +1,18 @@
 import os
-import tempfile
-from itertools import islice
-from pathlib import Path
-from typing import Generator, Iterable, Optional, Union
-import unicodedata
 import sys
-import numpy as np
 import re
-from typing import Set
+import tempfile
+import unicodedata
+from pathlib import Path
+from itertools import islice
+from typing import Iterable, Optional, TypeVar
+
+import numpy as np
+
+T = TypeVar("T")
 
 
-def normalize(input_array, p=2, dim=1, eps=1e-12) -> np.ndarray:
+def normalize(input_array: np.ndarray, p: int = 2, dim: int = 1, eps: float = 1e-12) -> np.ndarray:
     # Calculate the Lp norm along the specified dimension
     norm = np.linalg.norm(input_array, ord=p, axis=dim, keepdims=True)
     norm = np.maximum(norm, eps)  # Avoid division by zero
@@ -18,7 +20,7 @@ def normalize(input_array, p=2, dim=1, eps=1e-12) -> np.ndarray:
     return normalized_array
 
 
-def iter_batch(iterable: Union[Iterable, Generator], size: int) -> Iterable:
+def iter_batch(iterable: Iterable[T], size: int) -> Iterable[list[T]]:
     """
     >>> list(iter_batch([1,2,3,4,5], 3))
     [[1, 2, 3], [4, 5]]
@@ -45,7 +47,7 @@ def define_cache_dir(cache_dir: Optional[str] = None) -> Path:
     return cache_path
 
 
-def get_all_punctuation() -> Set[str]:
+def get_all_punctuation() -> set[str]:
     return set(
         chr(i) for i in range(sys.maxunicode) if unicodedata.category(chr(i)).startswith("P")
     )

fastembed/embedding.py

@@ -1,4 +1,4 @@
-from typing import Optional
+from typing import Optional, Any
 
 from loguru import logger
 
@@ -19,6 +19,6 @@ class JinaEmbedding(TextEmbedding):
         model_name: str = "jinaai/jina-embeddings-v2-base-en",
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)

fastembed/image/image_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type
+from typing import Any, Iterable, Optional, Sequence, Type
 
 import numpy as np
 
@@ -8,15 +8,15 @@ from fastembed.image.onnx_embedding import OnnxImageEmbedding
 
 
 class ImageEmbedding(ImageEmbeddingBase):
-    EMBEDDINGS_REGISTRY: List[Type[ImageEmbeddingBase]] = [OnnxImageEmbedding]
+    EMBEDDINGS_REGISTRY: list[Type[ImageEmbeddingBase]] = [OnnxImageEmbedding]
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
 
             Example:
                 ```
@@ -47,9 +47,9 @@ class ImageEmbedding(ImageEmbeddingBase):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         for EMBEDDING_MODEL_TYPE in self.EMBEDDINGS_REGISTRY:
@@ -77,7 +77,7 @@ class ImageEmbedding(ImageEmbeddingBase):
         images: ImageInput,
         batch_size: int = 16,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of documents into list of embeddings.

fastembed/image/image_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional
+from typing import Iterable, Optional, Any
 
 import numpy as np
 
@@ -12,7 +12,7 @@ class ImageEmbeddingBase(ModelManagement):
         model_name: str,
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
@@ -24,7 +24,7 @@ class ImageEmbeddingBase(ModelManagement):
         images: ImageInput,
         batch_size: int = 16,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Embeds a list of images into a list of embeddings.

fastembed/image/onnx_embedding.py

@@ -1,7 +1,6 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type
+from typing import Any, Iterable, Optional, Sequence, Type
 
 import numpy as np
-
 from fastembed.common import ImageInput, OnnxProvider
 from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import define_cache_dir, normalize
@@ -53,6 +52,17 @@ supported_onnx_models = [
         },
         "model_file": "model.onnx",
     },
+    {
+        "model": "jinaai/jina-clip-v1",
+        "dim": 768,
+        "description": "Image embeddings, Multimodal (text&image), 2024 year",
+        "license": "apache-2.0",
+        "size_in_GB": 0.34,
+        "sources": {
+            "hf": "jinaai/jina-clip-v1",
+        },
+        "model_file": "onnx/vision_model.onnx",
+    },
 ]
 
 
@@ -64,10 +74,11 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -80,11 +91,12 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
@@ -109,7 +121,10 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = define_cache_dir(cache_dir)
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         if not self.lazy_load:
@@ -129,12 +144,12 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
         )
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_onnx_models
 
@@ -143,7 +158,7 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
         images: ImageInput,
         batch_size: int = 16,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of images into list of embeddings.
@@ -178,8 +193,8 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
         return OnnxImageEmbeddingWorker
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
@@ -191,7 +206,7 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[np.ndarray]):
 
 
 class OnnxImageEmbeddingWorker(ImageEmbeddingWorker):
-    def init_embedding(self, model_name: str, cache_dir: str, **kwargs) -> OnnxImageEmbedding:
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> OnnxImageEmbedding:
         return OnnxImageEmbedding(
             model_name=model_name,
             cache_dir=cache_dir,

fastembed/image/onnx_image_model.py

@@ -2,7 +2,7 @@ import contextlib
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Type
+from typing import Any, Iterable, Optional, Sequence, Type
 
 import numpy as np
 from PIL import Image
@@ -24,13 +24,13 @@ class OnnxImageModel(OnnxModel[T]):
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def __init__(self) -> None:
+    def __init__(self):
         super().__init__()
         self.processor = None
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
@@ -58,10 +58,10 @@ class OnnxImageModel(OnnxModel[T]):
     def load_onnx_model(self) -> None:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _build_onnx_input(self, encoded: np.ndarray) -> Dict[str, np.ndarray]:
+    def _build_onnx_input(self, encoded: np.ndarray) -> dict[str, np.ndarray]:
         return {node.name: encoded for node in self.model.get_inputs()}
 
-    def onnx_embed(self, images: List[ImageInput], **kwargs) -> OnnxOutputContext:
+    def onnx_embed(self, images: list[ImageInput], **kwargs: Any) -> OnnxOutputContext:
         with contextlib.ExitStack():
             image_files = [
                 Image.open(image) if not isinstance(image, Image.Image) else image
@@ -83,8 +83,8 @@ class OnnxImageModel(OnnxModel[T]):
         parallel: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
-        **kwargs,
+        device_ids: Optional[list[int]] = None,
+        **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
 
@@ -125,7 +125,7 @@ class OnnxImageModel(OnnxModel[T]):
 
 
 class ImageEmbeddingWorker(EmbeddingWorker):
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
             embeddings = self.model.onnx_embed(batch)
             yield idx, embeddings

fastembed/image/transform/functional.py

@@ -1,4 +1,4 @@
-from typing import Sized, Tuple, Union
+from typing import Sized, Union
 
 import numpy as np
 from PIL import Image
@@ -14,7 +14,7 @@ def convert_to_rgb(image: Image.Image) -> Image.Image:
 
 def center_crop(
     image: Union[Image.Image, np.ndarray],
-    size: Tuple[int, int],
+    size: tuple[int, int],
 ) -> np.ndarray:
     if isinstance(image, np.ndarray):
         _, orig_height, orig_width = image.shape
@@ -62,8 +62,8 @@ def center_crop(
 
 def normalize(
     image: np.ndarray,
-    mean=Union[float, np.ndarray],
-    std=Union[float, np.ndarray],
+    mean: Union[float, np.ndarray],
+    std: Union[float, np.ndarray],
 ) -> np.ndarray:
     if not isinstance(image, np.ndarray):
         raise ValueError("image must be a numpy array")
@@ -96,10 +96,10 @@ def normalize(
 
 
 def resize(
-    image: Image,
-    size: Union[int, Tuple[int, int]],
-    resample: Image.Resampling = Image.Resampling.BILINEAR,
-) -> Image:
+    image: Image.Image,
+    size: Union[int, tuple[int, int]],
+    resample: Union[int, Image.Resampling] = Image.Resampling.BILINEAR,
+) -> Image.Image:
     if isinstance(size, tuple):
         return image.resize(size, resample)
 
@@ -114,11 +114,37 @@ def resize(
     return image.resize(new_size, resample)
 
 
-def rescale(image: np.ndarray, scale: float, dtype=np.float32) -> np.ndarray:
+def rescale(image: np.ndarray, scale: float, dtype: type = np.float32) -> np.ndarray:
     return (image * scale).astype(dtype)
 
 
-def pil2ndarray(image: Union[Image.Image, np.ndarray]):
+def pil2ndarray(image: Union[Image.Image, np.ndarray]) -> np.ndarray:
     if isinstance(image, Image.Image):
         return np.asarray(image).transpose((2, 0, 1))
     return image
+
+
+def pad2square(
+    image: Image.Image,
+    size: int,
+    fill_color: Union[str, int, tuple[int, ...]] = 0,
+) -> Image.Image:
+    height, width = image.height, image.width
+
+    left, right = 0, width
+    top, bottom = 0, height
+
+    crop_required = False
+    if width > size:
+        left = (width - size) // 2
+        right = left + size
+        crop_required = True
+
+    if height > size:
+        top = (height - size) // 2
+        bottom = top + size
+        crop_required = True
+
+    new_image = Image.new(mode="RGB", size=(size, size), color=fill_color)
+    new_image.paste(image.crop((left, top, right, bottom)) if crop_required else image)
+    return new_image

fastembed/image/transform/operators.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, List, Tuple, Union
+from typing import Any, Union, Optional
 
 import numpy as np
 from PIL import Image
@@ -10,93 +10,111 @@ from fastembed.image.transform.functional import (
     pil2ndarray,
     rescale,
     resize,
+    pad2square,
 )
 
 
 class Transform:
-    def __call__(self, images: List) -> Union[List[Image.Image], List[np.ndarray]]:
+    def __call__(self, images: list) -> Union[list[Image.Image], list[np.ndarray]]:
         raise NotImplementedError("Subclasses must implement this method")
 
 
 class ConvertToRGB(Transform):
-    def __call__(self, images: List[Image.Image]) -> List[Image.Image]:
+    def __call__(self, images: list[Image.Image]) -> list[Image.Image]:
         return [convert_to_rgb(image=image) for image in images]
 
 
 class CenterCrop(Transform):
-    def __init__(self, size: Tuple[int, int]):
+    def __init__(self, size: tuple[int, int]):
         self.size = size
 
-    def __call__(self, images: List[Image.Image]) -> List[np.ndarray]:
+    def __call__(self, images: list[Image.Image]) -> list[np.ndarray]:
         return [center_crop(image=image, size=self.size) for image in images]
 
 
 class Normalize(Transform):
-    def __init__(self, mean: Union[float, List[float]], std: Union[float, List[float]]):
+    def __init__(self, mean: Union[float, list[float]], std: Union[float, list[float]]):
         self.mean = mean
         self.std = std
 
-    def __call__(self, images: List[np.ndarray]) -> List[np.ndarray]:
+    def __call__(self, images: list[np.ndarray]) -> list[np.ndarray]:
         return [normalize(image, mean=self.mean, std=self.std) for image in images]
 
 
 class Resize(Transform):
     def __init__(
         self,
-        size: Union[int, Tuple[int, int]],
+        size: Union[int, tuple[int, int]],
         resample: Image.Resampling = Image.Resampling.BICUBIC,
     ):
         self.size = size
         self.resample = resample
 
-    def __call__(self, images: List[Image.Image]) -> List[Image.Image]:
-        return [
-            resize(image, size=self.size, resample=self.resample) for image in images
-        ]
+    def __call__(self, images: list[Image.Image]) -> list[Image.Image]:
+        return [resize(image, size=self.size, resample=self.resample) for image in images]
 
 
 class Rescale(Transform):
     def __init__(self, scale: float = 1 / 255):
         self.scale = scale
 
-    def __call__(self, images: List[np.ndarray]) -> List[np.ndarray]:
+    def __call__(self, images: list[np.ndarray]) -> list[np.ndarray]:
         return [rescale(image, scale=self.scale) for image in images]
 
 
 class PILtoNDarray(Transform):
-    def __call__(
-        self, images: List[Union[Image.Image, np.ndarray]]
-    ) -> List[np.ndarray]:
+    def __call__(self, images: list[Union[Image.Image, np.ndarray]]) -> list[np.ndarray]:
         return [pil2ndarray(image) for image in images]
 
 
+class PadtoSquare(Transform):
+    def __init__(
+        self,
+        size: int,
+        fill_color: Optional[Union[str, int, tuple[int, ...]]] = None,
+    ):
+        self.size = size
+        self.fill_color = fill_color
+
+    def __call__(self, images: list[Image.Image]) -> list[Image.Image]:
+        return [
+            pad2square(image=image, size=self.size, fill_color=self.fill_color) for image in images
+        ]
+
+
 class Compose:
-    def __init__(self, transforms: List[Transform]):
+    def __init__(self, transforms: list[Transform]):
         self.transforms = transforms
 
     def __call__(
-        self, images: Union[List[Image.Image], List[np.ndarray]]
-    ) -> Union[List[np.ndarray], List[Image.Image]]:
+        self, images: Union[list[Image.Image], list[np.ndarray]]
+    ) -> Union[list[np.ndarray], list[Image.Image]]:
         for transform in self.transforms:
             images = transform(images)
         return images
 
     @classmethod
-    def from_config(cls, config: Dict[str, Any]) -> "Compose":
+    def from_config(cls, config: dict[str, Any]) -> "Compose":
         """Creates processor from a config dict.
         Args:
-            config (Dict[str, Any]): Configuration dictionary.
+            config (dict[str, Any]): Configuration dictionary.
 
                 Valid keys:
                     - do_resize
+                    - resize_mode
                     - size
+                    - fill_color
                     - do_center_crop
                     - crop_size
                     - do_rescale
                     - rescale_factor
                     - do_normalize
                     - image_mean
+                    - mean
                     - image_std
+                    - std
+                    - resample
+                    - interpolation
                 Valid size keys (nested):
                     - {"height", "width"}
                     - {"shortest_edge"}
@@ -107,6 +125,7 @@ class Compose:
         transforms = []
         cls._get_convert_to_rgb(transforms, config)
         cls._get_resize(transforms, config)
+        cls._get_pad2square(transforms, config)
         cls._get_center_crop(transforms, config)
         cls._get_pil2ndarray(transforms, config)
         cls._get_rescale(transforms, config)
@@ -114,11 +133,11 @@ class Compose:
         return cls(transforms=transforms)
 
     @staticmethod
-    def _get_convert_to_rgb(transforms: List[Transform], config: Dict[str, Any]):
+    def _get_convert_to_rgb(transforms: list[Transform], config: dict[str, Any]) -> None:
         transforms.append(ConvertToRGB())
 
-    @staticmethod
-    def _get_resize(transforms: List[Transform], config: Dict[str, Any]):
+    @classmethod
+    def _get_resize(cls, transforms: list[Transform], config: dict[str, Any]) -> None:
         mode = config.get("image_processor_type", "CLIPImageProcessor")
         if mode == "CLIPImageProcessor":
             if config.get("do_resize", False):
@@ -161,9 +180,27 @@ class Compose:
                         resample=config.get("resample", Image.Resampling.BICUBIC),
                     )
                 )
+        elif mode == "JinaCLIPImageProcessor":
+            interpolation = config.get("interpolation")
+            if isinstance(interpolation, str):
+                resample = cls._interpolation_resolver(interpolation)
+            else:
+                resample = interpolation or Image.Resampling.BICUBIC
+
+            if "size" in config:
+                resize_mode = config.get("resize_mode", "shortest")
+                if resize_mode == "shortest":
+                    transforms.append(
+                        Resize(
+                            size=config["size"],
+                            resample=resample,
+                        )
+                    )
+        else:
+            raise ValueError(f"Preprocessor {mode} is not supported")
 
     @staticmethod
-    def _get_center_crop(transforms: List[Transform], config: Dict[str, Any]):
+    def _get_center_crop(transforms: list[Transform], config: dict[str, Any]) -> None:
         mode = config.get("image_processor_type", "CLIPImageProcessor")
         if mode == "CLIPImageProcessor":
             if config.get("do_center_crop", False):
@@ -177,22 +214,55 @@ class Compose:
                 transforms.append(CenterCrop(size=crop_size))
         elif mode == "ConvNextFeatureExtractor":
             pass
+        elif mode == "JinaCLIPImageProcessor":
+            pass
         else:
             raise ValueError(f"Preprocessor {mode} is not supported")
 
     @staticmethod
-    def _get_pil2ndarray(transforms: List[Transform], config: Dict[str, Any]):
+    def _get_pil2ndarray(transforms: list[Transform], config: dict[str, Any]) -> None:
         transforms.append(PILtoNDarray())
 
     @staticmethod
-    def _get_rescale(transforms: List[Transform], config: Dict[str, Any]):
+    def _get_rescale(transforms: list[Transform], config: dict[str, Any]) -> None:
         if config.get("do_rescale", True):
             rescale_factor = config.get("rescale_factor", 1 / 255)
             transforms.append(Rescale(scale=rescale_factor))
 
     @staticmethod
-    def _get_normalize(transforms: List[Transform], config: Dict[str, Any]):
+    def _get_normalize(transforms: list[Transform], config: dict[str, Any]) -> None:
         if config.get("do_normalize", False):
+            transforms.append(Normalize(mean=config["image_mean"], std=config["image_std"]))
+        elif "mean" in config and "std" in config:
+            transforms.append(Normalize(mean=config["mean"], std=config["std"]))
+
+    @staticmethod
+    def _get_pad2square(transforms: list[Transform], config: dict[str, Any]) -> None:
+        mode = config.get("image_processor_type", "CLIPImageProcessor")
+        if mode == "CLIPImageProcessor":
+            pass
+        elif mode == "ConvNextFeatureExtractor":
+            pass
+        elif mode == "JinaCLIPImageProcessor":
             transforms.append(
-                Normalize(mean=config["image_mean"], std=config["image_std"])
+                PadtoSquare(
+                    size=config["size"],
+                    fill_color=config.get("fill_color", 0),
+                )
             )
+
+    @staticmethod
+    def _interpolation_resolver(resample: Optional[str] = None) -> Image.Resampling:
+        interpolation_map = {
+            "nearest": Image.Resampling.NEAREST,
+            "lanczos": Image.Resampling.LANCZOS,
+            "bilinear": Image.Resampling.BILINEAR,
+            "bicubic": Image.Resampling.BICUBIC,
+            "box": Image.Resampling.BOX,
+            "hamming": Image.Resampling.HAMMING,
+        }
+
+        if resample and (method := interpolation_map.get(resample.lower())):
+            return method
+
+        raise ValueError(f"Unknown interpolation method: {resample}")

fastembed/late_interaction/colbert.py

@@ -1,5 +1,5 @@
 import string
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 from tokenizers import Encoding
@@ -68,21 +68,21 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
         return output.model_output.astype(np.float32)
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], is_doc: bool = True, **kwargs: Any
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], is_doc: bool = True, **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         marker_token = self.DOCUMENT_MARKER_TOKEN_ID if is_doc else self.QUERY_MARKER_TOKEN_ID
         onnx_input["input_ids"] = np.insert(onnx_input["input_ids"], 1, marker_token, axis=1)
         onnx_input["attention_mask"] = np.insert(onnx_input["attention_mask"], 1, 1, axis=1)
         return onnx_input
 
-    def tokenize(self, documents: List[str], is_doc: bool = True, **kwargs: Any) -> List[Encoding]:
+    def tokenize(self, documents: list[str], is_doc: bool = True, **kwargs: Any) -> list[Encoding]:
         return (
             self._tokenize_documents(documents=documents)
             if is_doc
             else self._tokenize_query(query=next(iter(documents)))
         )
 
-    def _tokenize_query(self, query: str) -> List[Encoding]:
+    def _tokenize_query(self, query: str) -> list[Encoding]:
         encoded = self.tokenizer.encode_batch([query])
         # colbert authors recommend to pad queries with [MASK] tokens for query augmentation to improve performance
         if len(encoded[0].ids) < self.MIN_QUERY_LENGTH:
@@ -101,16 +101,16 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
                 self.tokenizer.enable_padding(**prev_padding)
         return encoded
 
-    def _tokenize_documents(self, documents: List[str]) -> List[Encoding]:
+    def _tokenize_documents(self, documents: list[str]) -> list[Encoding]:
         encoded = self.tokenizer.encode_batch(documents)
         return encoded
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_colbert_models
 
@@ -121,10 +121,11 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -137,11 +138,12 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
@@ -167,7 +169,10 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
         self.cache_dir = define_cache_dir(cache_dir)
 
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
         self.mask_token_id = None
         self.pad_token_id = None
@@ -191,13 +196,16 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
             self.tokenizer.encode(symbol, add_special_tokens=False).ids[0]
             for symbol in string.punctuation
         }
+        current_max_length = self.tokenizer.truncation["max_length"]
+        # ensure not to overflow after adding document-marker
+        self.tokenizer.enable_truncation(max_length=current_max_length - 1)
 
     def embed(
         self,
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of documents into list of embeddings.
@@ -226,7 +234,7 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
             **kwargs,
         )
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[np.ndarray]:
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[np.ndarray]:
         if isinstance(query, str):
             query = [query]
 
@@ -244,7 +252,7 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[np.ndarray]):
 
 
 class ColbertEmbeddingWorker(TextEmbeddingWorker):
-    def init_embedding(self, model_name: str, cache_dir: str, **kwargs) -> Colbert:
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> Colbert:
         return Colbert(
             model_name=model_name,
             cache_dir=cache_dir,

fastembed/late_interaction/jina_colbert.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, List, Type
+from typing import Any, Type
 
 import numpy as np
 
@@ -33,17 +33,17 @@ class JinaColbert(Colbert):
         return JinaColbertEmbeddingWorker
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_jina_colbert_models
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], is_doc: bool = True, **kwargs: Any
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], is_doc: bool = True, **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         onnx_input = super()._preprocess_onnx_input(onnx_input, is_doc)
 
         # the attention mask for jina-colbert-v2 is always 1 in queries
@@ -53,7 +53,7 @@ class JinaColbert(Colbert):
 
 
 class JinaColbertEmbeddingWorker(TextEmbeddingWorker):
-    def init_embedding(self, model_name: str, cache_dir: str, **kwargs) -> JinaColbert:
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> JinaColbert:
         return JinaColbert(
             model_name=model_name,
             cache_dir=cache_dir,

fastembed/late_interaction/late_interaction_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Union
+from typing import Iterable, Optional, Union, Any
 
 import numpy as np
 
@@ -11,7 +11,7 @@ class LateInteractionTextEmbeddingBase(ModelManagement):
         model_name: str,
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
@@ -23,11 +23,11 @@ class LateInteractionTextEmbeddingBase(ModelManagement):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         raise NotImplementedError()
 
-    def passage_embed(self, texts: Iterable[str], **kwargs) -> Iterable[np.ndarray]:
+    def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[np.ndarray]:
         """
         Embeds a list of text passages into a list of embeddings.
 
@@ -42,9 +42,7 @@ class LateInteractionTextEmbeddingBase(ModelManagement):
         # This is model-specific, so that different models can have specialized implementations
         yield from self.embed(texts, **kwargs)
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs
-    ) -> Iterable[np.ndarray]:
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[np.ndarray]:
         """
         Embeds queries
 

fastembed/late_interaction/late_interaction_text_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 
@@ -11,15 +11,15 @@ from fastembed.late_interaction.late_interaction_embedding_base import (
 
 
 class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
-    EMBEDDINGS_REGISTRY: List[Type[LateInteractionTextEmbeddingBase]] = [Colbert, JinaColbert]
+    EMBEDDINGS_REGISTRY: list[Type[LateInteractionTextEmbeddingBase]] = [Colbert, JinaColbert]
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
 
             Example:
                 ```
@@ -50,9 +50,9 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         for EMBEDDING_MODEL_TYPE in self.EMBEDDINGS_REGISTRY:
@@ -80,7 +80,7 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of documents into list of embeddings.
@@ -99,7 +99,7 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[np.ndarray]:
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[np.ndarray]:
         """
         Embeds queries
 

fastembed/parallel_processor.py

@@ -1,14 +1,15 @@
 import logging
 import os
 from collections import defaultdict
+from copy import deepcopy
 from enum import Enum
 from multiprocessing import Queue, get_context
 from multiprocessing.context import BaseContext
 from multiprocessing.process import BaseProcess
 from multiprocessing.sharedctypes import Synchronized as BaseValue
 from queue import Empty
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Type
-from copy import deepcopy
+from typing import Any, Iterable, Optional, Type
+
 
 # Single item should be processed in less than:
 processing_timeout = 10 * 60  # seconds
@@ -27,7 +28,7 @@ class Worker:
     def start(cls, *args: Any, **kwargs: Any) -> "Worker":
         raise NotImplementedError()
 
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
         raise NotImplementedError()
 
 
@@ -37,7 +38,7 @@ def _worker(
     output_queue: Queue,
     num_active_workers: BaseValue,
     worker_id: int,
-    kwargs: Optional[Dict[str, Any]] = None,
+    kwargs: Optional[dict[str, Any]] = None,
 ) -> None:
     """
     A worker that pulls data pints off the input queue, and places the execution result on the output queue.
@@ -93,7 +94,7 @@ class ParallelWorkerPool:
         num_workers: int,
         worker: Type[Worker],
         start_method: Optional[str] = None,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         cuda: bool = False,
     ):
         self.worker_class = worker
@@ -101,7 +102,7 @@ class ParallelWorkerPool:
         self.input_queue: Optional[Queue] = None
         self.output_queue: Optional[Queue] = None
         self.ctx: BaseContext = get_context(start_method)
-        self.processes: List[BaseProcess] = []
+        self.processes: list[BaseProcess] = []
         self.queue_size = self.num_workers * max_internal_batch_size
         self.emergency_shutdown = False
         self.device_ids = device_ids
@@ -150,7 +151,7 @@ class ParallelWorkerPool:
 
     def semi_ordered_map(
         self, stream: Iterable[Any], *args: Any, **kwargs: Any
-    ) -> Iterable[Tuple[int, Any]]:
+    ) -> Iterable[tuple[int, Any]]:
         try:
             self.start(**kwargs)
 

fastembed/rerank/cross_encoder/onnx_text_cross_encoder.py

@@ -1,11 +1,15 @@
-from typing import List, Iterable, Dict, Any, Sequence, Optional
+from typing import Any, Iterable, Optional, Sequence, Type
 
 from loguru import logger
 
 from fastembed.common import OnnxProvider
-from fastembed.rerank.cross_encoder.onnx_text_model import OnnxCrossEncoderModel
-from fastembed.rerank.cross_encoder.text_cross_encoder_base import TextCrossEncoderBase
+from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.utils import define_cache_dir
+from fastembed.rerank.cross_encoder.onnx_text_model import (
+    OnnxCrossEncoderModel,
+    TextRerankerWorker,
+)
+from fastembed.rerank.cross_encoder.text_cross_encoder_base import TextCrossEncoderBase
 
 supported_onnx_models = [
     {
@@ -73,11 +77,11 @@ supported_onnx_models = [
 
 class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_onnx_models
 
@@ -88,10 +92,11 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -104,11 +109,12 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. Xenova/ms-marco-MiniLM-L-6-v2.
@@ -138,7 +144,10 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = define_cache_dir(cache_dir)
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         if not self.lazy_load:
@@ -159,7 +168,7 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         query: str,
         documents: Iterable[str],
         batch_size: int = 64,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[float]:
         """Reranks documents based on their relevance to a given query.
 
@@ -175,3 +184,44 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
         yield from self._rerank_documents(
             query=query, documents=documents, batch_size=batch_size, **kwargs
         )
+
+    def rerank_pairs(
+        self,
+        pairs: Iterable[tuple[str, str]],
+        batch_size: int = 64,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[float]:
+        yield from self._rerank_pairs(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            pairs=pairs,
+            batch_size=batch_size,
+            parallel=parallel,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            **kwargs,
+        )
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[TextRerankerWorker]:
+        return TextCrossEncoderWorker
+
+    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+        return (float(elem) for elem in output.model_output)
+
+
+class TextCrossEncoderWorker(TextRerankerWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        **kwargs: Any,
+    ) -> OnnxTextCrossEncoder:
+        return OnnxTextCrossEncoder(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/rerank/cross_encoder/onnx_text_model.py

@@ -1,16 +1,29 @@
-from typing import Sequence, Optional, List, Dict, Iterable
+import os
+from multiprocessing import get_all_start_methods
 from pathlib import Path
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
+from numpy.typing import NDArray
 from tokenizers import Encoding
 
-from fastembed.common.onnx_model import OnnxModel, OnnxProvider, OnnxOutputContext
+from fastembed.common.onnx_model import (
+    EmbeddingWorker,
+    OnnxModel,
+    OnnxOutputContext,
+    OnnxProvider,
+)
 from fastembed.common.preprocessor_utils import load_tokenizer
 from fastembed.common.utils import iter_batch
+from fastembed.parallel_processor import ParallelWorkerPool
 
 
-class OnnxCrossEncoderModel(OnnxModel):
-    ONNX_OUTPUT_NAMES: Optional[List[str]] = None
+class OnnxCrossEncoderModel(OnnxModel[float]):
+    ONNX_OUTPUT_NAMES: Optional[list[str]] = None
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["TextRerankerWorker"]:
+        raise NotImplementedError("Subclasses must implement this method")
 
     def _load_onnx_model(
         self,
@@ -31,40 +44,110 @@ class OnnxCrossEncoderModel(OnnxModel):
         )
         self.tokenizer, _ = load_tokenizer(model_dir=model_dir)
 
-    def tokenize(self, query: str, documents: List[str], **kwargs) -> List[Encoding]:
-        return self.tokenizer.encode_batch([(query, doc) for doc in documents])
-
-    def onnx_embed(self, query: str, documents: List[str], **kwargs) -> OnnxOutputContext:
-        tokenized_input = self.tokenize(query, documents, **kwargs)
+    def tokenize(self, pairs: list[tuple[str, str]], **_: Any) -> list[Encoding]:
+        return self.tokenizer.encode_batch(pairs)
 
+    def _build_onnx_input(
+        self, tokenized_input
+    ) -> dict[str, NDArray[Union[np.float32, np.int64]]]:
+        input_names = {node.name for node in self.model.get_inputs()}
         inputs = {
             "input_ids": np.array([enc.ids for enc in tokenized_input], dtype=np.int64),
-            "attention_mask": np.array(
-                [enc.attention_mask for enc in tokenized_input], dtype=np.int64
-            ),
         }
-        input_names = {node.name for node in self.model.get_inputs()}
         if "token_type_ids" in input_names:
             inputs["token_type_ids"] = np.array(
                 [enc.type_ids for enc in tokenized_input], dtype=np.int64
             )
+        if "attention_mask" in input_names:
+            inputs["attention_mask"] = np.array(
+                [enc.attention_mask for enc in tokenized_input], dtype=np.int64
+            )
+        return inputs
 
+    def onnx_embed(self, query: str, documents: list[str], **kwargs: Any) -> OnnxOutputContext:
+        pairs = [(query, doc) for doc in documents]
+        return self.onnx_embed_pairs(pairs, **kwargs)
+
+    def onnx_embed_pairs(self, pairs: list[tuple[str, str]], **kwargs: Any) -> OnnxOutputContext:
+        tokenized_input = self.tokenize(pairs, **kwargs)
+        inputs = self._build_onnx_input(tokenized_input)
         onnx_input = self._preprocess_onnx_input(inputs, **kwargs)
         outputs = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)
-        return OnnxOutputContext(model_output=outputs[0][:, 0].tolist())
+        relevant_output = outputs[0]
+        scores = relevant_output[:, 0]
+        return OnnxOutputContext(model_output=scores)
 
     def _rerank_documents(
-        self, query: str, documents: Iterable[str], batch_size: int, **kwargs
+        self, query: str, documents: Iterable[str], batch_size: int, **kwargs: Any
     ) -> Iterable[float]:
         if not hasattr(self, "model") or self.model is None:
             self.load_onnx_model()
         for batch in iter_batch(documents, batch_size):
-            yield from self.onnx_embed(query, batch, **kwargs).model_output
+            yield from self._post_process_onnx_output(self.onnx_embed(query, batch, **kwargs))
+
+    def _rerank_pairs(
+        self,
+        model_name: str,
+        cache_dir: str,
+        pairs: Iterable[tuple[str, str]],
+        batch_size: int,
+        parallel: Optional[int] = None,
+        providers: Optional[Sequence[OnnxProvider]] = None,
+        cuda: bool = False,
+        device_ids: Optional[list[int]] = None,
+        **kwargs: Any,
+    ) -> Iterable[float]:
+        is_small = False
+
+        if isinstance(pairs, tuple):
+            pairs = [pairs]
+            is_small = True
+
+        if isinstance(pairs, list):
+            if len(pairs) < batch_size:
+                is_small = True
+
+        if parallel is None or is_small:
+            if not hasattr(self, "model") or self.model is None:
+                self.load_onnx_model()
+            for batch in iter_batch(pairs, batch_size):
+                yield from self._post_process_onnx_output(self.onnx_embed_pairs(batch, **kwargs))
+        else:
+            if parallel == 0:
+                parallel = os.cpu_count()
+
+            start_method = "forkserver" if "forkserver" in get_all_start_methods() else "spawn"
+            params = {
+                "model_name": model_name,
+                "cache_dir": cache_dir,
+                "providers": providers,
+                **kwargs,
+            }
+
+            pool = ParallelWorkerPool(
+                num_workers=parallel or 1,
+                worker=self._get_worker_class(),
+                cuda=cuda,
+                device_ids=device_ids,
+                start_method=start_method,
+            )
+            for batch in pool.ordered_map(iter_batch(pairs, batch_size), **params):
+                yield from self._post_process_onnx_output(batch)
+
+    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+        raise NotImplementedError("Subclasses must implement this method")
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
         return onnx_input
+
+
+class TextRerankerWorker(EmbeddingWorker):
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
+        for idx, batch in items:
+            onnx_output = self.model.onnx_embed_pairs(batch)
+            yield idx, onnx_output

fastembed/rerank/cross_encoder/text_cross_encoder.py

@@ -1,21 +1,21 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type
+from typing import Any, Iterable, Optional, Sequence, Type
 
-from fastembed.rerank.cross_encoder.text_cross_encoder_base import TextCrossEncoderBase
-from fastembed.rerank.cross_encoder.onnx_text_cross_encoder import OnnxTextCrossEncoder
 from fastembed.common import OnnxProvider
+from fastembed.rerank.cross_encoder.onnx_text_cross_encoder import OnnxTextCrossEncoder
+from fastembed.rerank.cross_encoder.text_cross_encoder_base import TextCrossEncoderBase
 
 
 class TextCrossEncoder(TextCrossEncoderBase):
-    CROSS_ENCODER_REGISTRY: List[Type[TextCrossEncoderBase]] = [
+    CROSS_ENCODER_REGISTRY: list[Type[TextCrossEncoderBase]] = [
         OnnxTextCrossEncoder,
     ]
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
 
             Example:
                 ```
@@ -45,9 +45,9 @@ class TextCrossEncoder(TextCrossEncoderBase):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
 
@@ -72,7 +72,7 @@ class TextCrossEncoder(TextCrossEncoderBase):
         )
 
     def rerank(
-        self, query: str, documents: Iterable[str], batch_size: int = 64, **kwargs
+        self, query: str, documents: Iterable[str], batch_size: int = 64, **kwargs: Any
     ) -> Iterable[float]:
         """Rerank a list of documents based on a query.
 
@@ -85,3 +85,36 @@ class TextCrossEncoder(TextCrossEncoderBase):
             Iterable of scores for each document
         """
         yield from self.model.rerank(query, documents, batch_size=batch_size, **kwargs)
+
+    def rerank_pairs(
+        self,
+        pairs: Iterable[tuple[str, str]],
+        batch_size: int = 64,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[float]:
+        """
+        Rerank a list of query-document pairs.
+
+        Args:
+            pairs (Iterable[tuple[str, str]]): An iterable of tuples, where each tuple contains a query and a document
+                to be scored together.
+            batch_size (int, optional): The number of query-document pairs to process in a single batch. Defaults to 64.
+            parallel (Optional[int], optional): The number of parallel processes to use for reranking.
+                If None, parallelization is disabled. Defaults to None.
+            **kwargs (Any): Additional arguments to pass to the underlying reranking model.
+
+        Returns:
+            Iterable[float]: An iterable of scores corresponding to each query-document pair in the input.
+            Higher scores indicate a stronger match between the query and the document.
+
+        Example:
+            >>> encoder = TextCrossEncoder("Xenova/ms-marco-MiniLM-L-6-v2")
+            >>> pairs = [("What is AI?", "Artificial intelligence is ..."), ("What is ML?", "Machine learning is ...")]
+            >>> scores = list(encoder.rerank_pairs(pairs))
+            >>> print(list(map(lambda x: round(x, 2), scores)))
+            [-1.24, -10.6]
+        """
+        yield from self.model.rerank_pairs(
+            pairs, batch_size=batch_size, parallel=parallel, **kwargs
+        )

fastembed/rerank/cross_encoder/text_cross_encoder_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional
+from typing import Any, Iterable, Optional
 
 from fastembed.common.model_management import ModelManagement
 
@@ -9,7 +9,7 @@ class TextCrossEncoderBase(ModelManagement):
         model_name: str,
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
@@ -21,9 +21,9 @@ class TextCrossEncoderBase(ModelManagement):
         query: str,
         documents: Iterable[str],
         batch_size: int = 64,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[float]:
-        """Reranks a list of documents given a query.
+        """Rerank a list of documents given a query.
 
         Args:
             query (str): The query to rerank the documents.
@@ -32,6 +32,27 @@ class TextCrossEncoderBase(ModelManagement):
             **kwargs: Additional keyword argument to pass to the rerank method.
 
         Yields:
-            Iterable[float]: The scores of reranked the documents.
+            Iterable[float]: The scores of the reranked the documents.
+        """
+        raise NotImplementedError("This method should be overridden by subclasses")
+
+    def rerank_pairs(
+        self,
+        pairs: Iterable[tuple[str, str]],
+        batch_size: int = 64,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[float]:
+        """Rerank query-document pairs.
+        Args:
+            pairs (Iterable[tuple[str, str]]): Query-document pairs to rerank
+            batch_size (int): The batch size to use for reranking.
+            parallel: parallel:
+                If > 1, data-parallel encoding will be used, recommended for offline encoding of large datasets.
+                If 0, use all available cores.
+                If None, don't use data-parallel processing, use default onnxruntime threading instead.
+            **kwargs: Additional keyword argument to pass to the rerank method.
+        Yields:
+            Iterable[float]: Scores for each individual pair
         """
         raise NotImplementedError("This method should be overridden by subclasses")

fastembed/sparse/bm25.py

@@ -2,7 +2,7 @@ import os
 from collections import defaultdict
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Dict, Iterable, List, Optional, Tuple, Type, Union
+from typing import Any, Iterable, Optional, Type, Union
 
 import mmh3
 import numpy as np
@@ -92,6 +92,8 @@ class Bm25(SparseTextEmbeddingBase):
         b (float, optional): The b parameter in the BM25 formula. Defines the importance of the document length.
             Defaults to 0.75.
         avg_len (float, optional): The average length of the documents in the corpus. Defaults to 256.0.
+        language (str): Specifies the language for the stemmer.
+        disable_stemmer (bool): Disable the stemmer.
     Raises:
         ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
     """
@@ -105,7 +107,9 @@ class Bm25(SparseTextEmbeddingBase):
         avg_len: float = 256.0,
         language: str = "english",
         token_max_length: int = 40,
-        **kwargs,
+        disable_stemmer: bool = False,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, **kwargs)
 
@@ -122,27 +126,36 @@ class Bm25(SparseTextEmbeddingBase):
         self.cache_dir = define_cache_dir(cache_dir)
 
         self._model_dir = self.download_model(
-            model_description, self.cache_dir, local_files_only=self._local_files_only
+            model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         self.token_max_length = token_max_length
         self.punctuation = set(get_all_punctuation())
-        self.stopwords = set(self._load_stopwords(self._model_dir, self.language))
+        self.disable_stemmer = disable_stemmer
+
+        if disable_stemmer:
+            self.stopwords = set()
+            self.stemmer = None
+        else:
+            self.stopwords = set(self._load_stopwords(self._model_dir, self.language))
+            self.stemmer = SnowballStemmer(language)
 
-        self.stemmer = SnowballStemmer(language)
         self.tokenizer = SimpleTokenizer
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_bm25_models
 
     @classmethod
-    def _load_stopwords(cls, model_dir: Path, language: str) -> List[str]:
+    def _load_stopwords(cls, model_dir: Path, language: str) -> list[str]:
         stopwords_path = model_dir / f"{language}.txt"
         if not stopwords_path.exists():
             return []
@@ -182,6 +195,9 @@ class Bm25(SparseTextEmbeddingBase):
                 "k": self.k,
                 "b": self.b,
                 "avg_len": self.avg_len,
+                "language": self.language,
+                "token_max_length": self.token_max_length,
+                "disable_stemmer": self.disable_stemmer,
             }
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
@@ -197,7 +213,7 @@ class Bm25(SparseTextEmbeddingBase):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
         Encode a list of documents into list of embeddings.
@@ -222,19 +238,21 @@ class Bm25(SparseTextEmbeddingBase):
             parallel=parallel,
         )
 
-    def _stem(self, tokens: List[str]) -> List[str]:
+    def _stem(self, tokens: list[str]) -> list[str]:
         stemmed_tokens = []
         for token in tokens:
+            lower_token = token.lower()
+
             if token in self.punctuation:
                 continue
 
-            if token.lower() in self.stopwords:
+            if lower_token in self.stopwords:
                 continue
 
             if len(token) > self.token_max_length:
                 continue
 
-            stemmed_token = self.stemmer.stem_word(token.lower())
+            stemmed_token = self.stemmer.stem_word(lower_token) if self.stemmer else lower_token
 
             if stemmed_token:
                 stemmed_tokens.append(stemmed_token)
@@ -242,8 +260,8 @@ class Bm25(SparseTextEmbeddingBase):
 
     def raw_embed(
         self,
-        documents: List[str],
-    ) -> List[SparseEmbedding]:
+        documents: list[str],
+    ) -> list[SparseEmbedding]:
         embeddings = []
         for document in documents:
             document = remove_non_alphanumeric(document)
@@ -253,7 +271,7 @@ class Bm25(SparseTextEmbeddingBase):
             embeddings.append(SparseEmbedding.from_dict(token_id2value))
         return embeddings
 
-    def _term_frequency(self, tokens: List[str]) -> Dict[int, float]:
+    def _term_frequency(self, tokens: list[str]) -> dict[int, float]:
         """Calculate the term frequency part of the BM25 formula.
 
         (
@@ -263,10 +281,10 @@ class Bm25(SparseTextEmbeddingBase):
         )
 
         Args:
-            tokens (List[str]): The list of tokens in the document.
+            tokens (list[str]): The list of tokens in the document.
 
         Returns:
-            Dict[int, float]: The token_id to term frequency mapping.
+            dict[int, float]: The token_id to term frequency mapping.
         """
         tf_map = {}
         counter = defaultdict(int)
@@ -287,7 +305,9 @@ class Bm25(SparseTextEmbeddingBase):
     def compute_token_id(cls, token: str) -> int:
         return abs(mmh3.hash(token))
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[SparseEmbedding]:
+    def query_embed(
+        self, query: Union[str, Iterable[str]], **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         """To emulate BM25 behaviour, we don't need to use weights in the query, and
         it's enough to just hash the tokens and assign a weight of 1.0 to them.
         """
@@ -315,7 +335,7 @@ class Bm25Worker(Worker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model = self.init_embedding(model_name, cache_dir, **kwargs)
 
@@ -323,11 +343,11 @@ class Bm25Worker(Worker):
     def start(cls, model_name: str, cache_dir: str, **kwargs: Any) -> "Bm25Worker":
         return cls(model_name=model_name, cache_dir=cache_dir, **kwargs)
 
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
             onnx_output = self.model.raw_embed(batch)
             yield idx, onnx_output
 
     @staticmethod
-    def init_embedding(model_name: str, cache_dir: str, **kwargs) -> Bm25:
+    def init_embedding(model_name: str, cache_dir: str, **kwargs: Any) -> Bm25:
         return Bm25(model_name=model_name, cache_dir=cache_dir, **kwargs)

fastembed/sparse/bm42.py

@@ -1,7 +1,7 @@
 import math
 import string
 from pathlib import Path
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import mmh3
 import numpy as np
@@ -63,10 +63,11 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         alpha: float = 0.5,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -81,11 +82,12 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
                 It is recommended to only change this parameter based on training data for a specific dataset.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
@@ -111,7 +113,10 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.cache_dir = define_cache_dir(cache_dir)
 
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         self.invert_vocab = {}
@@ -141,7 +146,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.special_tokens_ids = set(self.special_token_to_id.values())
         self.stopwords = set(self._load_stopwords(self._model_dir))
 
-    def _filter_pair_tokens(self, tokens: List[Tuple[str, Any]]) -> List[Tuple[str, Any]]:
+    def _filter_pair_tokens(self, tokens: list[tuple[str, Any]]) -> list[tuple[str, Any]]:
         result = []
         for token, value in tokens:
             if token in self.stopwords or token in self.punctuation:
@@ -149,7 +154,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             result.append((token, value))
         return result
 
-    def _stem_pair_tokens(self, tokens: List[Tuple[str, Any]]) -> List[Tuple[str, Any]]:
+    def _stem_pair_tokens(self, tokens: list[tuple[str, Any]]) -> list[tuple[str, Any]]:
         result = []
         for token, value in tokens:
             processed_token = self.stemmer.stem_word(token)
@@ -158,8 +163,8 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
     @classmethod
     def _aggregate_weights(
-        cls, tokens: List[Tuple[str, List[int]]], weights: List[float]
-    ) -> List[Tuple[str, float]]:
+        cls, tokens: list[tuple[str, list[int]]], weights: list[float]
+    ) -> list[tuple[str, float]]:
         result = []
         for token, idxs in tokens:
             sum_weight = sum(weights[idx] for idx in idxs)
@@ -167,8 +172,8 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         return result
 
     def _reconstruct_bpe(
-        self, bpe_tokens: Iterable[Tuple[int, str]]
-    ) -> List[Tuple[str, List[int]]]:
+        self, bpe_tokens: Iterable[tuple[int, str]]
+    ) -> list[tuple[str, list[int]]]:
         result = []
         acc = ""
         acc_idx = []
@@ -195,7 +200,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
         return result
 
-    def _rescore_vector(self, vector: Dict[str, float]) -> Dict[int, float]:
+    def _rescore_vector(self, vector: dict[str, float]) -> dict[int, float]:
         """
         Orders all tokens in the vector by their importance and generates a new score based on the importance order.
         So that the scoring doesn't depend on absolute values assigned by the model, but on the relative importance.
@@ -246,16 +251,16 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             yield SparseEmbedding.from_dict(rescored)
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_bm42_models
 
     @classmethod
-    def _load_stopwords(cls, model_dir: Path) -> List[str]:
+    def _load_stopwords(cls, model_dir: Path) -> list[str]:
         stopwords_path = model_dir / "stopwords.txt"
         if not stopwords_path.exists():
             return []
@@ -268,7 +273,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
         Encode a list of documents into list of embeddings.
@@ -298,14 +303,16 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         )
 
     @classmethod
-    def _query_rehash(cls, tokens: Iterable[str]) -> Dict[int, float]:
+    def _query_rehash(cls, tokens: Iterable[str]) -> dict[int, float]:
         result = {}
         for token in tokens:
             token_id = abs(mmh3.hash(token))
             result[token_id] = 1.0
         return result
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[SparseEmbedding]:
+    def query_embed(
+        self, query: Union[str, Iterable[str]], **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         """
         To emulate BM25 behaviour, we don't need to use smart weights in the query, and
         it's enough to just hash the tokens and assign a weight of 1.0 to them.
@@ -332,7 +339,7 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
 
 class Bm42TextEmbeddingWorker(TextEmbeddingWorker):
-    def init_embedding(self, model_name: str, cache_dir: str, **kwargs) -> Bm42:
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> Bm42:
         return Bm42(
             model_name=model_name,
             cache_dir=cache_dir,

fastembed/sparse/sparse_embedding_base.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass
-from typing import Dict, Iterable, Optional, Union
+from typing import Iterable, Optional, Union, Any
 
 import numpy as np
 
@@ -11,17 +11,17 @@ class SparseEmbedding:
     values: np.ndarray
     indices: np.ndarray
 
-    def as_object(self) -> Dict[str, np.ndarray]:
+    def as_object(self) -> dict[str, np.ndarray]:
         return {
             "values": self.values,
             "indices": self.indices,
         }
 
-    def as_dict(self) -> Dict[int, float]:
+    def as_dict(self) -> dict[int, float]:
         return {i: v for i, v in zip(self.indices, self.values)}
 
     @classmethod
-    def from_dict(cls, data: Dict[int, float]) -> "SparseEmbedding":
+    def from_dict(cls, data: dict[int, float]) -> "SparseEmbedding":
         if len(data) == 0:
             return cls(values=np.array([]), indices=np.array([]))
         indices, values = zip(*data.items())
@@ -34,7 +34,7 @@ class SparseTextEmbeddingBase(ModelManagement):
         model_name: str,
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
@@ -46,13 +46,11 @@ class SparseTextEmbeddingBase(ModelManagement):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         raise NotImplementedError()
 
-    def passage_embed(
-        self, texts: Iterable[str], **kwargs
-    ) -> Iterable[SparseEmbedding]:
+    def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[SparseEmbedding]:
         """
         Embeds a list of text passages into a list of embeddings.
 
@@ -68,7 +66,7 @@ class SparseTextEmbeddingBase(ModelManagement):
         yield from self.embed(texts, **kwargs)
 
     def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs
+        self, query: Union[str, Iterable[str]], **kwargs: Any
     ) -> Iterable[SparseEmbedding]:
         """
         Embeds queries

fastembed/sparse/sparse_text_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 from fastembed.common import OnnxProvider
 from fastembed.sparse.bm25 import Bm25
@@ -12,15 +12,15 @@ import warnings
 
 
 class SparseTextEmbedding(SparseTextEmbeddingBase):
-    EMBEDDINGS_REGISTRY: List[Type[SparseTextEmbeddingBase]] = [SpladePP, Bm42, Bm25]
+    EMBEDDINGS_REGISTRY: list[Type[SparseTextEmbeddingBase]] = [SpladePP, Bm42, Bm25]
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
 
             Example:
                 ```
@@ -50,9 +50,9 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
         if model_name == "prithvida/Splade_PP_en_v1":
@@ -89,7 +89,7 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
         Encode a list of documents into list of embeddings.
@@ -108,7 +108,9 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
 
-    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[SparseEmbedding]:
+    def query_embed(
+        self, query: Union[str, Iterable[str]], **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         """
         Embeds queries
 

fastembed/sparse/splade_pp.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 from fastembed.common import OnnxProvider
@@ -55,11 +55,11 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             yield SparseEmbedding(values=scores, indices=indices)
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_splade_models
 
@@ -70,10 +70,11 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -86,11 +87,12 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
@@ -115,7 +117,10 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.cache_dir = define_cache_dir(cache_dir)
 
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         if not self.lazy_load:
@@ -136,7 +141,7 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[SparseEmbedding]:
         """
         Encode a list of documents into list of embeddings.
@@ -171,7 +176,7 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
 
 class SpladePPEmbeddingWorker(TextEmbeddingWorker):
-    def init_embedding(self, model_name: str, cache_dir: str, **kwargs) -> SpladePP:
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> SpladePP:
         return SpladePP(
             model_name=model_name,
             cache_dir=cache_dir,

fastembed/sparse/utils/tokenizer.py

@@ -1,12 +1,11 @@
 # This code is a modified copy of the `NLTKWordTokenizer` class from `NLTK` library.
 
 import re
-from typing import List
 
 
 class SimpleTokenizer:
     @staticmethod
-    def tokenize(text: str) -> List[str]:
+    def tokenize(text: str) -> list[str]:
         text = re.sub(r"[^\w]", " ", text.lower())
         text = re.sub(r"\s+", " ", text)
 
@@ -81,7 +80,7 @@ class WordTokenizer:
     ]
 
     @classmethod
-    def tokenize(cls, text: str) -> List[str]:
+    def tokenize(cls, text: str) -> list[str]:
         """Return a tokenized copy of `text`.
 
         >>> s = '''Good muffins cost $3.88 (roughly 3,36 euros)\nin New York.'''

fastembed/text/clip_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Type
+from typing import Any, Iterable, Type
 
 import numpy as np
 
@@ -22,18 +22,24 @@ supported_clip_models = [
 
 
 class CLIPOnnxEmbedding(OnnxTextEmbedding):
+    supported_models = supported_clip_models
+
     @classmethod
     def _get_worker_class(cls) -> Type[TextEmbeddingWorker]:
         return CLIPEmbeddingWorker
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
-        return supported_clip_models
+        return cls.supported_models
+
+    @classmethod
+    def add_custom_model(cls, model_info: dict[str, Any]):
+        cls.supported_models.append(model_info)
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[np.ndarray]:
         return output.model_output
@@ -44,7 +50,7 @@ class CLIPEmbeddingWorker(OnnxTextEmbeddingWorker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ) -> OnnxTextEmbedding:
         return CLIPOnnxEmbedding(
             model_name=model_name,

fastembed/text/e5_onnx_embedding.py

@@ -1,72 +0,0 @@
-from typing import Any, Dict, List, Type
-
-import numpy as np
-
-from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
-from fastembed.text.onnx_text_model import TextEmbeddingWorker
-
-supported_multilingual_e5_models = [
-    {
-        "model": "intfloat/multilingual-e5-large",
-        "dim": 1024,
-        "description": "Text embeddings, Unimodal (text), Multilingual (~100 languages), 512 input tokens truncation, Prefixes for queries/documents: necessary, 2024 year.",
-        "license": "mit",
-        "size_in_GB": 2.24,
-        "sources": {
-            "url": "https://storage.googleapis.com/qdrant-fastembed/fast-multilingual-e5-large.tar.gz",
-            "hf": "qdrant/multilingual-e5-large-onnx",
-        },
-        "model_file": "model.onnx",
-        "additional_files": ["model.onnx_data"],
-    },
-    {
-        "model": "sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
-        "dim": 768,
-        "description": "Text embeddings, Unimodal (text), Multilingual (~50 languages), 384 input tokens truncation, Prefixes for queries/documents: not necessary, 2021 year.",
-        "license": "apache-2.0",
-        "size_in_GB": 1.00,
-        "sources": {
-            "hf": "xenova/paraphrase-multilingual-mpnet-base-v2",
-        },
-        "model_file": "onnx/model.onnx",
-    },
-]
-
-
-class E5OnnxEmbedding(OnnxTextEmbedding):
-    @classmethod
-    def _get_worker_class(cls) -> Type["TextEmbeddingWorker"]:
-        return E5OnnxEmbeddingWorker
-
-    @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
-        """Lists the supported models.
-
-        Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
-        """
-        return supported_multilingual_e5_models
-
-    def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
-        """
-        Preprocess the onnx input.
-        """
-        onnx_input.pop("token_type_ids", None)
-        return onnx_input
-
-
-class E5OnnxEmbeddingWorker(OnnxTextEmbeddingWorker):
-    def init_embedding(
-        self,
-        model_name: str,
-        cache_dir: str,
-        **kwargs,
-    ) -> E5OnnxEmbedding:
-        return E5OnnxEmbedding(
-            model_name=model_name,
-            cache_dir=cache_dir,
-            threads=1,
-            **kwargs,
-        )

fastembed/text/multitask_embedding.py

@@ -0,0 +1,103 @@
+from enum import Enum
+from typing import Any, Type, Iterable, Union, Optional
+
+import numpy as np
+
+from fastembed.text.pooled_normalized_embedding import PooledNormalizedEmbedding
+from fastembed.text.onnx_embedding import OnnxTextEmbeddingWorker
+from fastembed.text.onnx_text_model import TextEmbeddingWorker
+
+supported_multitask_models = [
+    {
+        "model": "jinaai/jina-embeddings-v3",
+        "dim": 1024,
+        "tasks": {
+            "retrieval.query": 0,
+            "retrieval.passage": 1,
+            "separation": 2,
+            "classification": 3,
+            "text-matching": 4,
+        },
+        "description": "Multi-task unimodal (text) embedding model, multi-lingual (~100), 1024 tokens truncation, and 8192 sequence length. Prefixes for queries/documents: not necessary, 2024 year.",
+        "license": "cc-by-nc-4.0",
+        "size_in_GB": 2.29,
+        "sources": {
+            "hf": "jinaai/jina-embeddings-v3",
+        },
+        "model_file": "onnx/model.onnx",
+        "additional_files": ["onnx/model.onnx_data"],
+    },
+]
+
+
+class Task(int, Enum):
+    RETRIEVAL_QUERY = 0
+    RETRIEVAL_PASSAGE = 1
+    SEPARATION = 2
+    CLASSIFICATION = 3
+    TEXT_MATCHING = 4
+
+
+class JinaEmbeddingV3(PooledNormalizedEmbedding):
+    PASSAGE_TASK = Task.RETRIEVAL_PASSAGE
+    QUERY_TASK = Task.RETRIEVAL_QUERY
+    supported_models = supported_multitask_models
+
+    def __init__(self, *args: Any, **kwargs: Any):
+        super().__init__(*args, **kwargs)
+        self._current_task_id = self.PASSAGE_TASK
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["TextEmbeddingWorker"]:
+        return JinaEmbeddingV3Worker
+
+    @classmethod
+    def list_supported_models(cls) -> list[dict[str, Any]]:
+        return cls.supported_models
+
+    @classmethod
+    def add_custom_model(cls, model_info: dict[str, Any]):
+        cls.supported_models.append(model_info)
+
+    def _preprocess_onnx_input(
+        self, onnx_input: dict[str, np.ndarray], **kwargs
+    ) -> dict[str, np.ndarray]:
+        onnx_input["task_id"] = np.array(self._current_task_id, dtype=np.int64)
+        return onnx_input
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        task_id: int = PASSAGE_TASK,
+        **kwargs,
+    ) -> Iterable[np.ndarray]:
+        self._current_task_id = task_id
+        kwargs["task_id"] = task_id
+        yield from super().embed(documents, batch_size, parallel, **kwargs)
+
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs) -> Iterable[np.ndarray]:
+        self._current_task_id = self.QUERY_TASK
+        yield from super().embed(query, **kwargs)
+
+    def passage_embed(self, texts: Iterable[str], **kwargs) -> Iterable[np.ndarray]:
+        self._current_task_id = self.PASSAGE_TASK
+        yield from super().embed(texts, **kwargs)
+
+
+class JinaEmbeddingV3Worker(OnnxTextEmbeddingWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+        **kwargs,
+    ) -> JinaEmbeddingV3:
+        model = JinaEmbeddingV3(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )
+        model._current_task_id = kwargs["task_id"]
+        return model

fastembed/text/onnx_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 
@@ -16,6 +16,7 @@ supported_onnx_models = [
         "license": "mit",
         "size_in_GB": 0.42,
         "sources": {
+            "hf": "Qdrant/fast-bge-base-en",
             "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz",
         },
         "model_file": "model_optimized.onnx",
@@ -50,6 +51,7 @@ supported_onnx_models = [
         "license": "mit",
         "size_in_GB": 0.13,
         "sources": {
+            "hf": "Qdrant/bge-small-en",
             "url": "https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz",
         },
         "model_file": "model_optimized.onnx",
@@ -72,21 +74,11 @@ supported_onnx_models = [
         "license": "mit",
         "size_in_GB": 0.09,
         "sources": {
+            "hf": "Qdrant/bge-small-zh-v1.5",
             "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-zh-v1.5.tar.gz",
         },
         "model_file": "model_optimized.onnx",
     },
-    {
-        "model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
-        "dim": 384,
-        "description": "Text embeddings, Unimodal (text), Multilingual (~50 languages), 512 input tokens truncation, Prefixes for queries/documents: not necessary, 2019 year.",
-        "license": "apache-2.0",
-        "size_in_GB": 0.22,
-        "sources": {
-            "hf": "qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q",
-        },
-        "model_file": "model_optimized.onnx",
-    },
     {
         "model": "thenlper/gte-large",
         "dim": 1024,
@@ -164,21 +156,38 @@ supported_onnx_models = [
         },
         "model_file": "onnx/model.onnx",
     },
+    {
+        "model": "jinaai/jina-clip-v1",
+        "dim": 768,
+        "description": "Text embeddings, Multimodal (text&image), English, Prefixes for queries/documents: not necessary, 2024 year",
+        "license": "apache-2.0",
+        "size_in_GB": 0.55,
+        "sources": {
+            "hf": "jinaai/jina-clip-v1",
+        },
+        "model_file": "onnx/text_model.onnx",
+    },
 ]
 
 
 class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
     """Implementation of the Flag Embedding model."""
 
+    supported_models = supported_onnx_models
+
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
-        return supported_onnx_models
+        return cls.supported_models
+
+    @classmethod
+    def add_custom_model(cls, model_info: dict[str, Any]):
+        cls.supported_models.append(model_info)
 
     def __init__(
         self,
@@ -187,10 +196,11 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
         device_id: Optional[int] = None,
-        **kwargs,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
     ):
         """
         Args:
@@ -203,11 +213,12 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
                 Mutually exclusive with the `cuda` and `device_ids` arguments. Defaults to None.
             cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
                 Defaults to False.
-            device_ids (Optional[List[int]], optional): The list of device ids to use for data parallel processing in
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
                 workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
             lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
                 Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
             device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
 
         Raises:
             ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
@@ -231,7 +242,10 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = define_cache_dir(cache_dir)
         self._model_dir = self.download_model(
-            self.model_description, self.cache_dir, local_files_only=self._local_files_only
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=specific_model_path,
         )
 
         if not self.lazy_load:
@@ -242,7 +256,7 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of documents into list of embeddings.
@@ -276,8 +290,8 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
         return OnnxTextEmbeddingWorker
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
@@ -285,7 +299,13 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[np.ndarray]):
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[np.ndarray]:
         embeddings = output.model_output
-        return normalize(embeddings[:, 0]).astype(np.float32)
+        if embeddings.ndim == 3:  # (batch_size, seq_len, embedding_dim)
+            processed_embeddings = embeddings[:, 0]
+        elif embeddings.ndim == 2:  # (batch_size, embedding_dim)
+            processed_embeddings = embeddings
+        else:
+            raise ValueError(f"Unsupported embedding shape: {embeddings.shape}")
+        return normalize(processed_embeddings).astype(np.float32)
 
     def load_onnx_model(self) -> None:
         self._load_onnx_model(
@@ -303,7 +323,7 @@ class OnnxTextEmbeddingWorker(TextEmbeddingWorker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ) -> OnnxTextEmbedding:
         return OnnxTextEmbedding(
             model_name=model_name,

fastembed/text/onnx_text_model.py

@@ -1,7 +1,7 @@
 import os
 from multiprocessing import get_all_start_methods
 from pathlib import Path
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Type, Union
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
 from tokenizers import Encoding
@@ -14,7 +14,7 @@ from fastembed.parallel_processor import ParallelWorkerPool
 
 
 class OnnxTextModel(OnnxModel[T]):
-    ONNX_OUTPUT_NAMES: Optional[List[str]] = None
+    ONNX_OUTPUT_NAMES: Optional[list[str]] = None
 
     @classmethod
     def _get_worker_class(cls) -> Type["TextEmbeddingWorker"]:
@@ -23,14 +23,14 @@ class OnnxTextModel(OnnxModel[T]):
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def __init__(self) -> None:
+    def __init__(self):
         super().__init__()
         self.tokenizer = None
         self.special_token_to_id = {}
 
     def _preprocess_onnx_input(
-        self, onnx_input: Dict[str, np.ndarray], **kwargs
-    ) -> Dict[str, np.ndarray]:
+        self, onnx_input: dict[str, np.ndarray], **kwargs: Any
+    ) -> dict[str, np.ndarray]:
         """
         Preprocess the onnx input.
         """
@@ -44,6 +44,7 @@ class OnnxTextModel(OnnxModel[T]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_id: Optional[int] = None,
+        specific_model_path: Optional[str] = None,
     ) -> None:
         super()._load_onnx_model(
             model_dir=model_dir,
@@ -58,13 +59,13 @@ class OnnxTextModel(OnnxModel[T]):
     def load_onnx_model(self) -> None:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def tokenize(self, documents: List[str], **kwargs) -> List[Encoding]:
+    def tokenize(self, documents: list[str], **kwargs: Any) -> list[Encoding]:
         return self.tokenizer.encode_batch(documents)
 
     def onnx_embed(
         self,
-        documents: List[str],
-        **kwargs,
+        documents: list[str],
+        **kwargs: Any,
     ) -> OnnxOutputContext:
         encoded = self.tokenize(documents, **kwargs)
         input_ids = np.array([e.ids for e in encoded])
@@ -79,7 +80,6 @@ class OnnxTextModel(OnnxModel[T]):
             onnx_input["token_type_ids"] = np.array(
                 [np.zeros(len(e), dtype=np.int64) for e in input_ids], dtype=np.int64
             )
-
         onnx_input = self._preprocess_onnx_input(onnx_input, **kwargs)
 
         model_output = self.model.run(self.ONNX_OUTPUT_NAMES, onnx_input)
@@ -98,8 +98,8 @@ class OnnxTextModel(OnnxModel[T]):
         parallel: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
-        **kwargs,
+        device_ids: Optional[list[int]] = None,
+        **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
 
@@ -140,7 +140,7 @@ class OnnxTextModel(OnnxModel[T]):
 
 
 class TextEmbeddingWorker(EmbeddingWorker):
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, Any]]:
         for idx, batch in items:
             onnx_output = self.model.onnx_embed(batch)
             yield idx, onnx_output

fastembed/text/pooled_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Type
+from typing import Any, Iterable, Type
 
 import numpy as np
 
@@ -40,10 +40,47 @@ supported_pooled_models = [
         },
         "model_file": "onnx/model.onnx",
     },
+    {
+        "model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
+        "dim": 384,
+        "description": "Text embeddings, Unimodal (text), Multilingual (~50 languages), 512 input tokens truncation, Prefixes for queries/documents: not necessary, 2019 year.",
+        "license": "apache-2.0",
+        "size_in_GB": 0.22,
+        "sources": {
+            "hf": "qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q",
+        },
+        "model_file": "model_optimized.onnx",
+    },
+    {
+        "model": "sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
+        "dim": 768,
+        "description": "Text embeddings, Unimodal (text), Multilingual (~50 languages), 384 input tokens truncation, Prefixes for queries/documents: not necessary, 2021 year.",
+        "license": "apache-2.0",
+        "size_in_GB": 1.00,
+        "sources": {
+            "hf": "xenova/paraphrase-multilingual-mpnet-base-v2",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "intfloat/multilingual-e5-large",
+        "dim": 1024,
+        "description": "Text embeddings, Unimodal (text), Multilingual (~100 languages), 512 input tokens truncation, Prefixes for queries/documents: necessary, 2024 year.",
+        "license": "mit",
+        "size_in_GB": 2.24,
+        "sources": {
+            "url": "https://storage.googleapis.com/qdrant-fastembed/fast-multilingual-e5-large.tar.gz",
+            "hf": "qdrant/multilingual-e5-large-onnx",
+        },
+        "model_file": "model.onnx",
+        "additional_files": ["model.onnx_data"],
+    },
 ]
 
 
 class PooledEmbedding(OnnxTextEmbedding):
+    supported_models = supported_pooled_models
+
     @classmethod
     def _get_worker_class(cls) -> Type[TextEmbeddingWorker]:
         return PooledEmbeddingWorker
@@ -60,13 +97,17 @@ class PooledEmbedding(OnnxTextEmbedding):
         return pooled_embeddings
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
-        return supported_pooled_models
+        return cls.supported_models
+
+    @classmethod
+    def add_custom_model(cls, model_info: dict[str, Any]):
+        cls.supported_models.append(model_info)
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[np.ndarray]:
         if output.attention_mask is None:
@@ -82,7 +123,7 @@ class PooledEmbeddingWorker(OnnxTextEmbeddingWorker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ) -> OnnxTextEmbedding:
         return PooledEmbedding(
             model_name=model_name,

fastembed/text/pooled_normalized_embedding.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Type
+from typing import Any, Iterable, Type
 
 import numpy as np
 
@@ -57,22 +57,55 @@ supported_pooled_normalized_models = [
         "sources": {"hf": "jinaai/jina-embeddings-v2-base-code"},
         "model_file": "onnx/model.onnx",
     },
+    {
+        "model": "jinaai/jina-embeddings-v2-base-zh",
+        "dim": 768,
+        "description": "Text embeddings, Unimodal (text), supports mixed Chinese-English input text, 8192 input tokens truncation, Prefixes for queries/documents: not necessary, 2024 year.",
+        "license": "apache-2.0",
+        "size_in_GB": 0.64,
+        "sources": {"hf": "jinaai/jina-embeddings-v2-base-zh"},
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "jinaai/jina-embeddings-v2-base-es",
+        "dim": 768,
+        "description": "Text embeddings, Unimodal (text), supports mixed Spanish-English input text, 8192 input tokens truncation, Prefixes for queries/documents: not necessary, 2024 year.",
+        "license": "apache-2.0",
+        "size_in_GB": 0.64,
+        "sources": {"hf": "jinaai/jina-embeddings-v2-base-es"},
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "thenlper/gte-base",
+        "dim": 768,
+        "description": "General text embeddings, Unimodal (text), supports English only input text, 512 input tokens truncation, Prefixes for queries/documents: not necessary, 2024 year.",
+        "license": "mit",
+        "size_in_GB": 0.44,
+        "sources": {"hf": "thenlper/gte-base"},
+        "model_file": "onnx/model.onnx",
+    },
 ]
 
 
 class PooledNormalizedEmbedding(PooledEmbedding):
+    supported_models = supported_pooled_normalized_models
+
     @classmethod
     def _get_worker_class(cls) -> Type[TextEmbeddingWorker]:
         return PooledNormalizedEmbeddingWorker
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
         """
-        return supported_pooled_normalized_models
+        return cls.supported_models
+
+    @classmethod
+    def add_custom_model(cls, model_info: dict[str, Any]):
+        cls.supported_models.append(model_info)
 
     def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[np.ndarray]:
         if output.attention_mask is None:
@@ -88,7 +121,7 @@ class PooledNormalizedEmbeddingWorker(OnnxTextEmbeddingWorker):
         self,
         model_name: str,
         cache_dir: str,
-        **kwargs,
+        **kwargs: Any,
     ) -> OnnxTextEmbedding:
         return PooledNormalizedEmbedding(
             model_name=model_name,

fastembed/text/text_embedding.py

@@ -1,32 +1,32 @@
-from typing import Any, Dict, Iterable, List, Optional, Sequence, Type, Union
+import warnings
+from typing import Any, Iterable, Optional, Sequence, Type, Union
 
 import numpy as np
-
 from fastembed.common import OnnxProvider
 from fastembed.text.clip_embedding import CLIPOnnxEmbedding
-from fastembed.text.e5_onnx_embedding import E5OnnxEmbedding
 from fastembed.text.pooled_normalized_embedding import PooledNormalizedEmbedding
 from fastembed.text.pooled_embedding import PooledEmbedding
+from fastembed.text.multitask_embedding import JinaEmbeddingV3
 from fastembed.text.onnx_embedding import OnnxTextEmbedding
 from fastembed.text.text_embedding_base import TextEmbeddingBase
 
 
 class TextEmbedding(TextEmbeddingBase):
-    EMBEDDINGS_REGISTRY: List[Type[TextEmbeddingBase]] = [
+    EMBEDDINGS_REGISTRY: list[Type[TextEmbeddingBase]] = [
         OnnxTextEmbedding,
-        E5OnnxEmbedding,
         CLIPOnnxEmbedding,
         PooledNormalizedEmbedding,
         PooledEmbedding,
+        JinaEmbeddingV3,
     ]
 
     @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
+    def list_supported_models(cls) -> list[dict[str, Any]]:
         """
         Lists the supported models.
 
         Returns:
-            List[Dict[str, Any]]: A list of dictionaries containing the model information.
+            list[dict[str, Any]]: A list of dictionaries containing the model information.
 
             Example:
                 ```
@@ -50,6 +50,38 @@ class TextEmbedding(TextEmbeddingBase):
             result.extend(embedding.list_supported_models())
         return result
 
+    @classmethod
+    def add_custom_model(
+        cls, model_info: dict[str, Any], mean_pooling: bool = True, normalization: bool = False
+    ) -> None:
+        """
+        Register a custom model so that TextEmbedding(...) can find it later.
+
+        Args:
+            model_info: Dictionary describing the model, e.g.:
+                {
+                    "model": "alibaba/blablabla",
+                    "dim": 512,
+                    "description": "...",
+                    "license": "apache-2.0",
+                    "size_in_GB": 1.23,
+                    "sources": { ... }  # optional
+                }
+            mean_pooling: apply mean_pooling or not.
+            normalization: apply normalization or not.
+
+        Returns:
+            None
+        """
+        if mean_pooling and not normalization:
+            PooledEmbedding.add_custom_model(model_info)
+        elif mean_pooling and normalization:
+            PooledNormalizedEmbedding.add_custom_model(model_info)
+        elif "clip" in model_info["model"].lower():
+            CLIPOnnxEmbedding.add_custom_model(model_info)
+        else:
+            OnnxTextEmbedding.add_custom_model(model_info)
+
     def __init__(
         self,
         model_name: str = "BAAI/bge-small-en-v1.5",
@@ -57,11 +89,36 @@ class TextEmbedding(TextEmbeddingBase):
         threads: Optional[int] = None,
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
-        device_ids: Optional[List[int]] = None,
+        device_ids: Optional[list[int]] = None,
         lazy_load: bool = False,
-        **kwargs,
+        **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
+        if model_name == "nomic-ai/nomic-embed-text-v1.5-Q":
+            warnings.warn(
+                "The model 'nomic-ai/nomic-embed-text-v1.5-Q' has been updated on HuggingFace. "
+                "Please review the latest documentation and release notes to ensure compatibility with your workflow. ",
+                UserWarning,
+                stacklevel=2,
+            )
+        if model_name == "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2":
+            warnings.warn(
+                "The model 'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2' has been updated to "
+                "include a mean pooling layer. Please ensure your usage aligns with the new functionality. "
+                "Support for the previous version without mean pooling will be removed as of version 0.5.2.",
+                UserWarning,
+                stacklevel=2,
+            )
+        if model_name in {
+            "sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
+            "intfloat/multilingual-e5-large",
+        }:
+            warnings.warn(
+                f"{model_name} has been updated as of fastembed 0.5.2, outputs are now average pooled.",
+                UserWarning,
+                stacklevel=2,
+            )
+
         for EMBEDDING_MODEL_TYPE in self.EMBEDDINGS_REGISTRY:
             supported_models = EMBEDDING_MODEL_TYPE.list_supported_models()
             if any(model_name.lower() == model["model"].lower() for model in supported_models):
@@ -78,7 +135,7 @@ class TextEmbedding(TextEmbeddingBase):
                 return
 
         raise ValueError(
-            f"Model {model_name} is not supported in TextEmbedding."
+            f"Model {model_name} is not supported in TextEmbedding. "
             "Please check the supported models using `TextEmbedding.list_supported_models()`"
         )
 
@@ -87,7 +144,7 @@ class TextEmbedding(TextEmbeddingBase):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         """
         Encode a list of documents into list of embeddings.
@@ -105,3 +162,30 @@ class TextEmbedding(TextEmbeddingBase):
             List of embeddings, one per document
         """
         yield from self.model.embed(documents, batch_size, parallel, **kwargs)
+
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[np.ndarray]:
+        """
+        Embeds queries
+
+        Args:
+            query (Union[str, Iterable[str]]): The query to embed, or an iterable e.g. list of queries.
+
+        Returns:
+            Iterable[np.ndarray]: The embeddings.
+        """
+        # This is model-specific, so that different models can have specialized implementations
+        yield from self.model.query_embed(query, **kwargs)
+
+    def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[np.ndarray]:
+        """
+        Embeds a list of text passages into a list of embeddings.
+
+        Args:
+            texts (Iterable[str]): The list of texts to embed.
+            **kwargs: Additional keyword argument to pass to the embed method.
+
+        Yields:
+            Iterable[SparseEmbedding]: The sparse embeddings.
+        """
+        # This is model-specific, so that different models can have specialized implementations
+        yield from self.model.passage_embed(texts, **kwargs)

fastembed/text/text_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Iterable, Optional, Union
+from typing import Iterable, Optional, Union, Any
 
 import numpy as np
 
@@ -11,7 +11,7 @@ class TextEmbeddingBase(ModelManagement):
         model_name: str,
         cache_dir: Optional[str] = None,
         threads: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ):
         self.model_name = model_name
         self.cache_dir = cache_dir
@@ -23,11 +23,11 @@ class TextEmbeddingBase(ModelManagement):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        **kwargs,
+        **kwargs: Any,
     ) -> Iterable[np.ndarray]:
         raise NotImplementedError()
 
-    def passage_embed(self, texts: Iterable[str], **kwargs) -> Iterable[np.ndarray]:
+    def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[np.ndarray]:
         """
         Embeds a list of text passages into a list of embeddings.
 
@@ -42,9 +42,8 @@ class TextEmbeddingBase(ModelManagement):
         # This is model-specific, so that different models can have specialized implementations
         yield from self.embed(texts, **kwargs)
 
-    def query_embed(
-        self, query: Union[str, Iterable[str]], **kwargs
-    ) -> Iterable[np.ndarray]:
+    def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[np.ndarray]:
+
         """
         Embeds queries
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "fastembed"
-version = "0.4.2"
+version = "0.5.1"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["Qdrant Team <info@qdrant.tech>", "NirantK <nirant.bits@gmail.com>"]
 license = "Apache License"
@@ -11,39 +11,53 @@ repository = "https://github.com/qdrant/fastembed"
 keywords = ["vector", "embedding", "neural", "search", "qdrant", "sentence-transformers"]
 
 [tool.poetry.dependencies]
-python = ">=3.8.0,<3.13"
-onnx = "^1.15.0"
-onnxruntime = ">=1.17.0,<1.20.0"
+python = ">=3.9.0"
+numpy = [
+    { version = ">=1.21,<2.1.0", python = "<3.10" },
+    { version = ">=1.21", python = ">=3.10,<3.12" },
+    { version = ">=1.26", python = ">=3.12,<3.13" },
+    { version = ">=2.1.0", python = ">=3.13" }
+]
+onnxruntime = [
+    { version = ">=1.17.0,<1.20.0", python = "<3.10" },
+    { version = ">=1.17.0,!=1.20.0", python = ">=3.10,<3.13" },
+    { version = ">1.20.0", python = ">=3.13" }
+]
 tqdm = "^4.66"
 requests = "^2.31"
 tokenizers = ">=0.15,<1.0"
 huggingface-hub = ">=0.20,<1.0"
 loguru = "^0.7.2"
-numpy = [
-    { version = ">=1.21", python = "<3.12" },
-    { version = ">=1.26", python = ">=3.12" }
-]
-pillow = "^10.3.0"
+pillow = ">=10.3.0,<12.0.0"
 mmh3 = "^4.1.0"
 py-rust-stemmers = "^0.1.0"
 
-[tool.poetry.group.dev.dependencies]
+[tool.poetry.group.test.dependencies]
 pytest = "^7.4.2"
 ruff = ">=0.3.1,<1.0"
+
+[tool.poetry.group.dev.dependencies]
 notebook = ">=7.0.2"
-pre-commit = {version = "^3.6.2", python = ">=3.9,<3.12" }
+pre-commit = "^3.6.2"
+onnx = ">=1.15.0"
 
 [tool.poetry.group.docs.dependencies]
 mkdocs-material = "^9.5.10"
 mkdocstrings = "^0.24.0"
-pillow = "^10.2.0"
+pillow = ">=10.3.0,<12.0.0"
 cairosvg = "^2.7.1"
 mknotebooks = "^0.8.0"
 
+[tool.poetry.group.types.dependencies]
+pyright = ">=1.1.293"
+mypy = "^1.0.0"
 
 [build-system]
 requires = ["poetry-core"]
 build-backend = "poetry.core.masonry.api"
 
+[tool.pyright]
+typeCheckingMode = "strict"
+
 [tool.ruff]
 line-length = 99

tests/profiling.py

@@ -9,7 +9,7 @@
 
 # %%
 import time
-from typing import Callable, List, Tuple
+from typing import Callable
 
 import matplotlib.pyplot as plt
 import torch.nn.functional as F
@@ -23,7 +23,7 @@ from fastembed.embedding import DefaultEmbedding
 # data is a list of strings, each string is a document.
 
 # %%
-documents: List[str] = [
+documents: list[str] = [
     "Chandrayaan-3 is India's third lunar mission",
     "It aimed to land a rover on the Moon's surface - joining the US, China and Russia",
     "The mission is a follow-up to Chandrayaan-2, which had partial success",
@@ -56,7 +56,7 @@ class HF:
         self.model = AutoModel.from_pretrained(model_id)
         self.tokenizer = AutoTokenizer.from_pretrained(model_id)
 
-    def embed(self, texts: List[str]):
+    def embed(self, texts: list[str]):
         encoded_input = self.tokenizer(
             texts, max_length=512, padding=True, truncation=True, return_tensors="pt"
         )
@@ -88,7 +88,7 @@ embedding_model = DefaultEmbedding()
 # %%
 def calculate_time_stats(
     embed_func: Callable, documents: list, k: int
-) -> Tuple[float, float, float]:
+) -> tuple[float, float, float]:
     times = []
     for _ in range(k):
         # Timing the embed_func call
@@ -111,8 +111,8 @@ print(f"FastEmbed (Average, Max, Min): {fst_stats}")
 
 # %%
 def plot_character_per_second_comparison(
-    hf_stats: Tuple[float, float, float],
-    fst_stats: Tuple[float, float, float],
+    hf_stats: tuple[float, float, float],
+    fst_stats: tuple[float, float, float],
     documents: list,
 ):
     # Calculating total characters in documents

tests/test_add_custom_model.py

@@ -0,0 +1,115 @@
+import os
+import numpy as np
+import pytest
+
+from fastembed.text.text_embedding import TextEmbedding
+from tests.utils import delete_model_cache
+
+canonical_vectors = [
+    {
+        "model": "intfloat/multilingual-e5-small",
+        "mean_pooling": True,
+        "normalization": True,
+        "canonical_vector": [3.1317e-02, 3.0939e-02, -3.5117e-02, -6.7274e-02, 8.5084e-02],
+    },
+    {
+        "model": "intfloat/multilingual-e5-small",
+        "mean_pooling": True,
+        "normalization": False,
+        "canonical_vector": [1.4604e-01, 1.4428e-01, -1.6376e-01, -3.1372e-01, 3.9677e-01],
+    },
+    {
+        "model": "mixedbread-ai/mxbai-embed-xsmall-v1",
+        "mean_pooling": False,
+        "normalization": False,
+        "canonical_vector": [
+            2.49407589e-02,
+            1.00189969e-02,
+            1.07807154e-02,
+            3.63860987e-02,
+            -2.27128249e-02,
+        ],
+    },
+]
+
+DIMENSIONS = {
+    "intfloat/multilingual-e5-small": 384,
+    "mixedbread-ai/mxbai-embed-xsmall-v1": 384,
+}
+
+SOURCES = {
+    "intfloat/multilingual-e5-small": "intfloat/multilingual-e5-small",
+    "mixedbread-ai/mxbai-embed-xsmall-v1": "mixedbread-ai/mxbai-embed-xsmall-v1",
+}
+
+
+@pytest.mark.parametrize("scenario", canonical_vectors)
+def test_add_custom_model_variations(scenario):
+    is_ci = bool(os.getenv("CI", False))
+
+    base_model_name = scenario["model"]
+    mean_pooling = scenario["mean_pooling"]
+    normalization = scenario["normalization"]
+    cv = np.array(scenario["canonical_vector"], dtype=np.float32)
+
+    backup_supported_models = {}
+    for embedding_cls in TextEmbedding.EMBEDDINGS_REGISTRY:
+        backup_supported_models[embedding_cls] = embedding_cls.list_supported_models().copy()
+
+    suffixes = []
+    suffixes.append("mean" if mean_pooling else "no-mean")
+    suffixes.append("norm" if normalization else "no-norm")
+    suffix_str = "-".join(suffixes)
+
+    custom_model_name = f"{base_model_name}-{suffix_str}"
+
+    dim = DIMENSIONS[base_model_name]
+    hf_source = SOURCES[base_model_name]
+
+    model_info = {
+        "model": custom_model_name,
+        "dim": dim,
+        "description": f"{base_model_name} with {suffix_str}",
+        "license": "mit",
+        "size_in_GB": 0.13,
+        "sources": {
+            "hf": hf_source,
+        },
+        "model_file": "onnx/model.onnx",
+        "additional_files": [],
+    }
+
+    if is_ci and model_info["size_in_GB"] > 1.0:
+        pytest.skip(
+            f"Skipping {custom_model_name} on CI due to size_in_GB={model_info['size_in_GB']}"
+        )
+
+    try:
+        TextEmbedding.add_custom_model(
+            model_info=model_info, mean_pooling=mean_pooling, normalization=normalization
+        )
+
+        model = TextEmbedding(model_name=custom_model_name)
+
+        docs = ["hello world", "flag embedding"]
+        embeddings = list(model.embed(docs))
+        embeddings = np.stack(embeddings, axis=0)
+
+        assert embeddings.shape == (
+            2,
+            dim,
+        ), f"Expected shape (2, {dim}) for {custom_model_name}, but got {embeddings.shape}"
+
+        num_compare_dims = cv.shape[0]
+        assert np.allclose(
+            embeddings[0, :num_compare_dims], cv, atol=1e-3
+        ), f"Embedding mismatch for {custom_model_name} (first {num_compare_dims} dims)."
+
+        assert not np.allclose(embeddings[0, :], 0.0), "Embedding should not be all zeros."
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+    finally:
+        for embedding_cls, old_list in backup_supported_models.items():
+            embedding_cls.supported_models = old_list

tests/test_attention_embeddings.py

@@ -8,7 +8,7 @@ from tests.utils import delete_model_cache
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25"])
-def test_attention_embeddings(model_name):
+def test_attention_embeddings(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = SparseTextEmbedding(model_name=model_name)
 
@@ -71,7 +71,7 @@ def test_attention_embeddings(model_name):
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions", "Qdrant/bm25"])
-def test_parallel_processing(model_name):
+def test_parallel_processing(model_name: str) -> None:
     is_ci = os.getenv("CI")
 
     model = SparseTextEmbedding(model_name=model_name)
@@ -96,7 +96,7 @@ def test_parallel_processing(model_name):
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm25"])
-def test_multilanguage(model_name):
+def test_multilanguage(model_name: str) -> None:
     is_ci = os.getenv("CI")
 
     docs = ["Mangez-vous vraiment des grenouilles?", "Je suis au lit"]
@@ -122,7 +122,7 @@ def test_multilanguage(model_name):
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm25"])
-def test_special_characters(model_name):
+def test_special_characters(model_name: str) -> None:
     is_ci = os.getenv("CI")
 
     docs = [
@@ -145,7 +145,7 @@ def test_special_characters(model_name):
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/bm42-all-minilm-l6-v2-attentions"])
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str) -> None:
     model = SparseTextEmbedding(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")
     docs = ["hello world", "flag embedding"]

tests/test_image_onnx_embeddings.py

@@ -21,10 +21,13 @@ CANONICAL_VECTOR_VALUES = {
     "Qdrant/Unicom-ViT-B-32": np.array(
         [0.0418, 0.0550, 0.0003, 0.0253, -0.0185, 0.0016, -0.0368, -0.0402, -0.0891, -0.0186]
     ),
+    "jinaai/jina-clip-v1": np.array(
+        [-0.029, 0.0216, 0.0396, 0.0283, -0.0023, 0.0151, 0.011, -0.0235, 0.0251, -0.0343]
+    ),
 }
 
 
-def test_embedding():
+def test_embedding() -> None:
     is_ci = os.getenv("CI")
 
     for model_desc in ImageEmbedding.list_supported_models():
@@ -58,7 +61,7 @@ def test_embedding():
 
 
 @pytest.mark.parametrize("n_dims,model_name", [(512, "Qdrant/clip-ViT-B-32-vision")])
-def test_batch_embedding(n_dims, model_name):
+def test_batch_embedding(n_dims: int, model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = ImageEmbedding(model_name=model_name)
     n_images = 32
@@ -78,7 +81,7 @@ def test_batch_embedding(n_dims, model_name):
 
 
 @pytest.mark.parametrize("n_dims,model_name", [(512, "Qdrant/clip-ViT-B-32-vision")])
-def test_parallel_processing(n_dims, model_name):
+def test_parallel_processing(n_dims: int, model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = ImageEmbedding(model_name=model_name)
 
@@ -106,7 +109,7 @@ def test_parallel_processing(n_dims, model_name):
 
 
 @pytest.mark.parametrize("model_name", ["Qdrant/clip-ViT-B-32-vision"])
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = ImageEmbedding(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")

tests/test_late_interaction_embeddings.py

@@ -226,7 +226,7 @@ def test_parallel_processing():
     "model_name",
     ["colbert-ir/colbertv2.0"],
 )
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str):
     is_ci = os.getenv("CI")
 
     model = LateInteractionTextEmbedding(model_name=model_name, lazy_load=True)

tests/test_multi_gpu.py

@@ -1,4 +1,5 @@
 import pytest
+from typing import Optional
 from fastembed import (
     TextEmbedding,
     SparseTextEmbedding,
@@ -13,7 +14,7 @@ CACHE_DIR = "../model_cache"
 
 @pytest.mark.skip(reason="Requires a multi-gpu server")
 @pytest.mark.parametrize("device_id", [None, 0, 1])
-def test_gpu_via_providers(device_id):
+def test_gpu_via_providers(device_id: Optional[int]) -> None:
     docs = ["hello world", "flag embedding"]
 
     device_id = device_id if device_id is not None else 0
@@ -85,7 +86,7 @@ def test_gpu_via_providers(device_id):
 
 @pytest.mark.skip(reason="Requires a multi-gpu server")
 @pytest.mark.parametrize("device_ids", [None, [0], [1], [0, 1]])
-def test_gpu_cuda_device_ids(device_ids):
+def test_gpu_cuda_device_ids(device_ids: Optional[list[int]]) -> None:
     docs = ["hello world", "flag embedding"]
     device_id = device_ids[0] if device_ids else 0
     embedding_model = TextEmbedding(
@@ -170,7 +171,7 @@ def test_gpu_cuda_device_ids(device_ids):
 @pytest.mark.parametrize(
     "device_ids,parallel", [(None, None), (None, 2), ([1], None), ([1], 1), ([1], 2), ([0, 1], 2)]
 )
-def test_multi_gpu_parallel_inference(device_ids, parallel):
+def test_multi_gpu_parallel_inference(device_ids: Optional[list[int]], parallel: int) -> None:
     docs = ["hello world", "flag embedding"] * 100
     batch_size = 5
 

tests/test_sparse_embeddings.py

@@ -49,7 +49,7 @@ CANONICAL_COLUMN_VALUES = {
 docs = ["Hello World"]
 
 
-def test_batch_embedding():
+def test_batch_embedding() -> None:
     is_ci = os.getenv("CI")
     docs_to_embed = docs * 10
 
@@ -64,7 +64,7 @@ def test_batch_embedding():
             delete_model_cache(model.model._model_dir)
 
 
-def test_single_embedding():
+def test_single_embedding() -> None:
     is_ci = os.getenv("CI")
     for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
         model = SparseTextEmbedding(model_name=model_name)
@@ -80,7 +80,7 @@ def test_single_embedding():
             delete_model_cache(model.model._model_dir)
 
 
-def test_parallel_processing():
+def test_parallel_processing() -> None:
     is_ci = os.getenv("CI")
     model = SparseTextEmbedding(model_name="prithivida/Splade_PP_en_v1")
     docs = ["hello world", "flag embedding"] * 30
@@ -111,7 +111,7 @@ def test_parallel_processing():
 
 
 @pytest.fixture
-def bm25_instance():
+def bm25_instance() -> None:
     ci = os.getenv("CI", True)
     model = Bm25("Qdrant/bm25", language="english")
     yield model
@@ -119,7 +119,7 @@ def bm25_instance():
         delete_model_cache(model._model_dir)
 
 
-def test_stem_with_stopwords_and_punctuation(bm25_instance):
+def test_stem_with_stopwords_and_punctuation(bm25_instance: Bm25) -> None:
     # Setup
     bm25_instance.stopwords = {"the", "is", "a"}
     bm25_instance.punctuation = {".", ",", "!"}
@@ -135,7 +135,7 @@ def test_stem_with_stopwords_and_punctuation(bm25_instance):
     assert result == expected, f"Expected {expected}, but got {result}"
 
 
-def test_stem_case_insensitive_stopwords(bm25_instance):
+def test_stem_case_insensitive_stopwords(bm25_instance: Bm25) -> None:
     # Setup
     bm25_instance.stopwords = {"the", "is", "a"}
     bm25_instance.punctuation = {".", ",", "!"}
@@ -151,11 +151,32 @@ def test_stem_case_insensitive_stopwords(bm25_instance):
     assert result == expected, f"Expected {expected}, but got {result}"
 
 
+@pytest.mark.parametrize("disable_stemmer", [True, False])
+def test_disable_stemmer_behavior(disable_stemmer: bool) -> None:
+    # Setup
+    model = Bm25("Qdrant/bm25", language="english", disable_stemmer=disable_stemmer)
+    model.stopwords = {"the", "is", "a"}
+    model.punctuation = {".", ",", "!"}
+
+    # Test data
+    tokens = ["The", "quick", "brown", "fox", "is", "a", "test", "sentence", ".", "!"]
+
+    # Execute
+    result = model._stem(tokens)
+
+    # Assert
+    if disable_stemmer:
+        expected = ["quick", "brown", "fox", "test", "sentence"]  # no stemming, lower case only
+    else:
+        expected = ["quick", "brown", "fox", "test", "sentenc"]
+    assert result == expected, f"Expected {expected}, but got {result}"
+
+
 @pytest.mark.parametrize(
     "model_name",
     ["prithivida/Splade_PP_en_v1"],
 )
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = SparseTextEmbedding(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")

tests/test_text_cross_encoder.py

@@ -15,38 +15,45 @@ CANONICAL_SCORE_VALUES = {
     "jinaai/jina-reranker-v2-base-multilingual": np.array([1.6533, -1.6455]),
 }
 
-
-def test_rerank():
-    is_ci = os.getenv("CI")
-
-    for model_desc in TextCrossEncoder.list_supported_models():
-        if not is_ci and model_desc["size_in_GB"] > 1:
-            continue
-
-        model_name = model_desc["model"]
-        model = TextCrossEncoder(model_name=model_name)
-
-        query = "What is the capital of France?"
-        documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."]
-        scores = np.array(list(model.rerank(query, documents)))
-
-        canonical_scores = CANONICAL_SCORE_VALUES[model_name]
-        assert np.allclose(
-            scores, canonical_scores, atol=1e-3
-        ), f"Model: {model_name}, Scores: {scores}, Expected: {canonical_scores}"
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
+SELECTED_MODELS = {
+    "Xenova": "Xenova/ms-marco-MiniLM-L-6-v2",
+    "BAAI": "BAAI/bge-reranker-base",
+    "jinaai": "jinaai/jina-reranker-v1-tiny-en",
+}
 
 
 @pytest.mark.parametrize(
     "model_name",
-    [
-        model_desc["model"]
-        for model_desc in TextCrossEncoder.list_supported_models()
-        if model_desc["size_in_GB"] < 1 and model_desc["model"] in CANONICAL_SCORE_VALUES.keys()
-    ],
+    [model_name for model_name in CANONICAL_SCORE_VALUES],
 )
-def test_batch_rerank(model_name):
+def test_rerank(model_name: str) -> None:
+    is_ci = os.getenv("CI")
+
+    model = TextCrossEncoder(model_name=model_name)
+
+    query = "What is the capital of France?"
+    documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."]
+    scores = np.array(list(model.rerank(query, documents)))
+
+    pairs = [(query, doc) for doc in documents]
+    scores2 = np.array(list(model.rerank_pairs(pairs)))
+    assert np.allclose(
+        scores, scores2, atol=1e-5
+    ), f"Model: {model_name}, Scores: {scores}, Scores2: {scores2}"
+
+    canonical_scores = CANONICAL_SCORE_VALUES[model_name]
+    assert np.allclose(
+        scores, canonical_scores, atol=1e-3
+    ), f"Model: {model_name}, Scores: {scores}, Expected: {canonical_scores}"
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    [model_name for model_name in SELECTED_MODELS.values()],
+)
+def test_batch_rerank(model_name: str) -> None:
     is_ci = os.getenv("CI")
 
     model = TextCrossEncoder(model_name=model_name)
@@ -55,6 +62,12 @@ def test_batch_rerank(model_name):
     documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 50
     scores = np.array(list(model.rerank(query, documents, batch_size=10)))
 
+    pairs = [(query, doc) for doc in documents]
+    scores2 = np.array(list(model.rerank_pairs(pairs)))
+    assert np.allclose(
+        scores, scores2, atol=1e-5
+    ), f"Model: {model_name}, Scores: {scores}, Scores2: {scores2}"
+
     canonical_scores = np.tile(CANONICAL_SCORE_VALUES[model_name], 50)
 
     assert scores.shape == canonical_scores.shape, f"Unexpected shape for model {model_name}"
@@ -69,7 +82,7 @@ def test_batch_rerank(model_name):
     "model_name",
     ["Xenova/ms-marco-MiniLM-L-6-v2"],
 )
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextCrossEncoder(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")
@@ -80,3 +93,27 @@ def test_lazy_load(model_name):
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    [model_name for model_name in SELECTED_MODELS.values()],
+)
+def test_rerank_pairs_parallel(model_name: str) -> None:
+    is_ci = os.getenv("CI")
+
+    model = TextCrossEncoder(model_name=model_name)
+    query = "What is the capital of France?"
+    documents = ["Paris is the capital of France.", "Berlin is the capital of Germany."] * 10
+    pairs = [(query, doc) for doc in documents]
+    scores_parallel = np.array(list(model.rerank_pairs(pairs, parallel=2, batch_size=10)))
+    scores_sequential = np.array(list(model.rerank_pairs(pairs, batch_size=10)))
+    assert np.allclose(
+        scores_parallel, scores_sequential, atol=1e-5
+    ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Scores (Sequential): {scores_sequential}"
+    canonical_scores = CANONICAL_SCORE_VALUES[model_name]
+    assert np.allclose(
+        scores_parallel[: len(canonical_scores)], canonical_scores, atol=1e-3
+    ), f"Model: {model_name}, Scores (Parallel): {scores_parallel}, Expected: {canonical_scores}"
+    if is_ci:
+        delete_model_cache(model.model._model_dir)

tests/test_text_multitask_embeddings.py

@@ -0,0 +1,253 @@
+import os
+
+import numpy as np
+import pytest
+
+from fastembed import TextEmbedding
+from fastembed.text.multitask_embedding import Task
+from tests.utils import delete_model_cache
+
+
+CANONICAL_VECTOR_VALUES = {
+    "jinaai/jina-embeddings-v3": [
+        {
+            "task_id": Task.RETRIEVAL_QUERY,
+            "vectors": np.array(
+                [
+                    [0.0623, -0.0402, 0.1706, -0.0143, 0.0617],
+                    [-0.1064, -0.0733, 0.0353, 0.0096, 0.0667],
+                ]
+            ),
+        },
+        {
+            "task_id": Task.RETRIEVAL_PASSAGE,
+            "vectors": np.array(
+                [
+                    [0.0513, -0.0247, 0.1751, -0.0075, 0.0679],
+                    [-0.0987, -0.0786, 0.09, 0.0087, 0.0577],
+                ]
+            ),
+        },
+        {
+            "task_id": Task.SEPARATION,
+            "vectors": np.array(
+                [
+                    [0.094, -0.1065, 0.1305, 0.0547, 0.0556],
+                    [0.0315, -0.1468, 0.065, 0.0568, 0.0546],
+                ]
+            ),
+        },
+        {
+            "task_id": Task.CLASSIFICATION,
+            "vectors": np.array(
+                [
+                    [0.0606, -0.0877, 0.1384, 0.0065, 0.0722],
+                    [-0.0502, -0.119, 0.032, 0.0514, 0.0689],
+                ]
+            ),
+        },
+        {
+            "task_id": Task.TEXT_MATCHING,
+            "vectors": np.array(
+                [
+                    [0.0911, -0.0341, 0.1305, -0.026, 0.0576],
+                    [-0.1432, -0.05, 0.0133, 0.0464, 0.0789],
+                ]
+            ),
+        },
+    ]
+}
+docs = ["Hello World", "Follow the white rabbit."]
+
+
+def test_batch_embedding():
+    is_ci = os.getenv("CI")
+    docs_to_embed = docs * 10
+    default_task = Task.RETRIEVAL_PASSAGE
+
+    for model_desc in TextEmbedding.list_supported_models():
+        if not is_ci and model_desc["size_in_GB"] > 1:
+            continue
+
+        model_name = model_desc["model"]
+        dim = model_desc["dim"]
+
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
+
+        model = TextEmbedding(model_name=model_name)
+
+        print(f"evaluating {model_name} default task")
+
+        embeddings = list(model.embed(documents=docs_to_embed, batch_size=6))
+        embeddings = np.stack(embeddings, axis=0)
+
+        assert embeddings.shape == (len(docs_to_embed), dim)
+
+        canonical_vector = CANONICAL_VECTOR_VALUES[model_name][default_task]["vectors"]
+        assert np.allclose(
+            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+        ), model_desc["model"]
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+
+def test_single_embedding():
+    is_ci = os.getenv("CI")
+
+    for model_desc in TextEmbedding.list_supported_models():
+        if not is_ci and model_desc["size_in_GB"] > 1:
+            continue
+
+        model_name = model_desc["model"]
+        dim = model_desc["dim"]
+
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
+
+        model = TextEmbedding(model_name=model_name)
+
+        for task in CANONICAL_VECTOR_VALUES[model_name]:
+            print(f"evaluating {model_name} task_id: {task['task_id']}")
+
+            embeddings = list(model.embed(documents=docs, task_id=task["task_id"]))
+            embeddings = np.stack(embeddings, axis=0)
+
+            assert embeddings.shape == (len(docs), dim)
+
+            canonical_vector = task["vectors"]
+            assert np.allclose(
+                embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+            ), model_desc["model"]
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+
+def test_single_embedding_query():
+    is_ci = os.getenv("CI")
+    task_id = Task.RETRIEVAL_QUERY
+
+    for model_desc in TextEmbedding.list_supported_models():
+        if not is_ci and model_desc["size_in_GB"] > 1:
+            continue
+
+        model_name = model_desc["model"]
+        dim = model_desc["dim"]
+
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
+
+        model = TextEmbedding(model_name=model_name)
+
+        print(f"evaluating {model_name} query_embed task_id: {task_id}")
+
+        embeddings = list(model.query_embed(query=docs))
+        embeddings = np.stack(embeddings, axis=0)
+
+        assert embeddings.shape == (len(docs), dim)
+
+        canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
+        assert np.allclose(
+            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+        ), model_desc["model"]
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+
+def test_single_embedding_passage():
+    is_ci = os.getenv("CI")
+    task_id = Task.RETRIEVAL_PASSAGE
+
+    for model_desc in TextEmbedding.list_supported_models():
+        if not is_ci and model_desc["size_in_GB"] > 1:
+            continue
+
+        model_name = model_desc["model"]
+        dim = model_desc["dim"]
+
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
+
+        model = TextEmbedding(model_name=model_name)
+
+        print(f"evaluating {model_name} passage_embed task_id: {task_id}")
+
+        embeddings = list(model.passage_embed(texts=docs))
+        embeddings = np.stack(embeddings, axis=0)
+
+        assert embeddings.shape == (len(docs), dim)
+
+        canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
+        assert np.allclose(
+            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+        ), model_desc["model"]
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+
+def test_parallel_processing():
+    is_ci = os.getenv("CI")
+
+    docs = ["Hello World", "Follow the white rabbit."] * 10
+
+    model_name = "jinaai/jina-embeddings-v3"
+    dim = 1024
+
+    model = TextEmbedding(model_name=model_name)
+
+    task_id = Task.SEPARATION
+    embeddings_1 = list(model.embed(docs, batch_size=10, parallel=None, task_id=task_id))
+    embeddings_1 = np.stack(embeddings_1, axis=0)
+
+    embeddings_2 = list(model.embed(docs, batch_size=10, parallel=1, task_id=task_id))
+    embeddings_2 = np.stack(embeddings_2, axis=0)
+
+    assert embeddings_1.shape[0] == len(docs) and embeddings_1.shape[-1] == dim
+    assert np.allclose(embeddings_1, embeddings_2, atol=1e-4)
+
+    canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
+    assert np.allclose(embeddings_2[:2, : canonical_vector.shape[1]], canonical_vector, atol=1e-4)
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)
+
+
+def test_task_assignment():
+    is_ci = os.getenv("CI")
+
+    for model_desc in TextEmbedding.list_supported_models():
+        if not is_ci and model_desc["size_in_GB"] > 1:
+            continue
+
+        model_name = model_desc["model"]
+        if model_name not in CANONICAL_VECTOR_VALUES.keys():
+            continue
+
+        model = TextEmbedding(model_name=model_name)
+
+        for i, task_id in enumerate(Task):
+            _ = list(model.embed(documents=docs, batch_size=1, task_id=i))
+            assert model.model._current_task_id == task_id
+
+        if is_ci:
+            delete_model_cache(model.model._model_dir)
+
+
+@pytest.mark.parametrize(
+    "model_name",
+    ["jinaai/jina-embeddings-v3"],
+)
+def test_lazy_load(model_name: str):
+    is_ci = os.getenv("CI")
+    model = TextEmbedding(model_name=model_name, lazy_load=True)
+    assert not hasattr(model.model, "model")
+
+    list(model.embed(docs))
+    assert hasattr(model.model, "model")
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)

tests/test_text_onnx_embeddings.py

@@ -1,4 +1,5 @@
 import os
+import platform
 
 import numpy as np
 import pytest
@@ -31,22 +32,24 @@ CANONICAL_VECTOR_VALUES = {
         [-0.034478, 0.03102, 0.00673, 0.02611, -0.039362]
     ),
     "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2": np.array(
-        [0.0094, 0.0184, 0.0328, 0.0072, -0.0351]
+        [0.0361, 0.1862, 0.2776, 0.2461, -0.1904]
     ),
-    "intfloat/multilingual-e5-large": np.array([0.0098, 0.0045, 0.0066, -0.0354, 0.0070]),
+    "intfloat/multilingual-e5-large": np.array([0.4544, -0.0968, 0.1054, -1.3753, 0.1500]),
     "sentence-transformers/paraphrase-multilingual-mpnet-base-v2": np.array(
-        [-0.01341097, 0.0416553, -0.00480805, 0.02844842, 0.0505299]
+        [0.0047, 0.1334, -0.0102, 0.0714, 0.1930]
     ),
     "jinaai/jina-embeddings-v2-small-en": np.array([-0.0455, -0.0428, -0.0122, 0.0613, 0.0015]),
     "jinaai/jina-embeddings-v2-base-en": np.array([-0.0332, -0.0509, 0.0287, -0.0043, -0.0077]),
     "jinaai/jina-embeddings-v2-base-de": np.array([-0.0085, 0.0417, 0.0342, 0.0309, -0.0149]),
     "jinaai/jina-embeddings-v2-base-code": np.array([0.0145, -0.0164, 0.0136, -0.0170, 0.0734]),
+    "jinaai/jina-embeddings-v2-base-zh": np.array([0.0381, 0.0286, -0.0231, 0.0052, -0.0151]),
+    "jinaai/jina-embeddings-v2-base-es": np.array([-0.0108, -0.0092, -0.0373, 0.0171, -0.0301]),
     "nomic-ai/nomic-embed-text-v1": np.array([0.3708, 0.2031, -0.3406, -0.2114, -0.3230]),
     "nomic-ai/nomic-embed-text-v1.5": np.array(
         [-0.15407836, -0.03053198, -3.9138033, 0.1910364, 0.13224715]
     ),
     "nomic-ai/nomic-embed-text-v1.5-Q": np.array(
-        [-0.12525563, 0.38030425, -3.961622, 0.04176439, -0.0758301]
+        [0.0802303, 0.3700881, -4.3053818, 0.4431803, -0.271572]
     ),
     "thenlper/gte-large": np.array(
         [-0.01920587, 0.00113156, -0.00708992, -0.00632304, -0.04025577]
@@ -62,14 +65,23 @@ CANONICAL_VECTOR_VALUES = {
     ),
     "snowflake/snowflake-arctic-embed-l": np.array([0.0189, -0.0673, 0.0183, 0.0124, 0.0146]),
     "Qdrant/clip-ViT-B-32-text": np.array([0.0083, 0.0103, -0.0138, 0.0199, -0.0069]),
+    "thenlper/gte-base": np.array([0.0038, 0.0355, 0.0181, 0.0092, 0.0654]),
+    "jinaai/jina-clip-v1": np.array([-0.0862, -0.0101, -0.0056, 0.0375, -0.0472]),
 }
 
+MULTI_TASK_MODELS = ["jinaai/jina-embeddings-v3"]
 
-def test_embedding():
+
+def test_embedding() -> None:
     is_ci = os.getenv("CI")
+    is_mac = platform.system() == "Darwin"
 
     for model_desc in TextEmbedding.list_supported_models():
-        if not is_ci and model_desc["size_in_GB"] > 1:
+        if (
+            (not is_ci and model_desc["size_in_GB"] > 1)
+            or model_desc["model"] in MULTI_TASK_MODELS
+            or (is_mac and model_desc["model"] == "nomic-ai/nomic-embed-text-v1.5-Q")
+        ):
             continue
 
         dim = model_desc["dim"]
@@ -92,7 +104,7 @@ def test_embedding():
     "n_dims,model_name",
     [(384, "BAAI/bge-small-en-v1.5"), (768, "jinaai/jina-embeddings-v2-base-en")],
 )
-def test_batch_embedding(n_dims, model_name):
+def test_batch_embedding(n_dims: int, model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextEmbedding(model_name=model_name)
 
@@ -109,7 +121,7 @@ def test_batch_embedding(n_dims, model_name):
     "n_dims,model_name",
     [(384, "BAAI/bge-small-en-v1.5"), (768, "jinaai/jina-embeddings-v2-base-en")],
 )
-def test_parallel_processing(n_dims, model_name):
+def test_parallel_processing(n_dims: int, model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextEmbedding(model_name=model_name)
 
@@ -135,7 +147,7 @@ def test_parallel_processing(n_dims, model_name):
     "model_name",
     ["BAAI/bge-small-en-v1.5"],
 )
-def test_lazy_load(model_name):
+def test_lazy_load(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = TextEmbedding(model_name=model_name, lazy_load=True)
     assert not hasattr(model.model, "model")

tests/utils.py

@@ -2,7 +2,8 @@ import shutil
 import traceback
 
 from pathlib import Path
-from typing import Union
+from types import TracebackType
+from typing import Union, Callable, Any, Type
 
 
 def delete_model_cache(model_dir: Union[str, Path]) -> None:
@@ -16,7 +17,11 @@ def delete_model_cache(model_dir: Union[str, Path]) -> None:
         model_dir (Union[str, Path]): The path to the model cache directory.
     """
 
-    def on_error(func, path, exc_info):
+    def on_error(
+        func: Callable[..., Any],
+        path: str,
+        exc_info: tuple[Type[BaseException], BaseException, TracebackType],
+    ) -> None:
         print("Failed to remove: ", path)
         print("Exception: ", exc_info)
         traceback.print_exception(*exc_info)