debug: try out with markers

* feat: Quantized models

* refactor: use model_file for GCS

* refactoring: refactor model downloading (#209)

* refactoring: refactor model downloading

* refactor: update docstring

Co-authored-by: Anush <anushshetty90@gmail.com>

* Update fastembed/common/model_management.py

Co-authored-by: George <george.panchuk@qdrant.tech>

* fix: model_file for Snowflake models

---------

Co-authored-by: George <george.panchuk@qdrant.tech>

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -0,0 +1,56 @@
+name: Bug/New Model Request
+description: File a bug report/Request a new Model
+title: "[Bug/Model Request]: "
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+      value: "A bug happened!"
+    validations:
+      required: true
+  - type: textarea
+    id: Python version
+    attributes:
+      label: What Python version are you on? e.g. python --version
+      description: Also tell us, what package manager are you using e.g. conda, pip, poetry?
+      placeholder: Python3.10
+    validations:
+      required: true
+  - type: dropdown
+    id: version
+    attributes:
+      label: 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.
+      options:
+        - 0.2.6 (Latest)
+        - 0.2.5
+        - 0.2.4 
+        - 0.2.3 
+        - 0.2.2
+        - 0.2.1
+        - 0.1.x 
+      default: 0
+    validations:
+      required: true
+  - type: dropdown
+    id: os
+    attributes:
+      label: What os are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - MacOS
+        - Windows
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant stack traces and/or logs
+      description: Please copy and paste any relevant raised exceptions. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/config.yaml

@@ -0,0 +1,5 @@
+blank_issues_enabled: false
+contact_links:
+  - name: GitHub Community Support
+    url: https://github.com/qdrant/fastembed/discussions
+    about: Please ask and answer questions here.

.github/workflows/python-tests.yml

@@ -3,6 +3,8 @@ name: Tests
 on:
   push:
     branches: [ master, main ]
+  schedule:
+    - cron: 0 0 * * *
   pull_request:
 
 env:
@@ -29,18 +31,24 @@ jobs:
     name: Python ${{ matrix.python-version }} on ${{ matrix.os }} test
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v3
       - name: Set up Python
-        uses: actions/setup-python@v2
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python-version }}
       - name: Install dependencies
         run: |
           python -m pip install poetry
           poetry config virtualenvs.create false
-          poetry install --no-interaction --no-ansi
-      - name: Run tests
-        run: |
-          export IS_UBUNTU_CI=$(test "${{ matrix.os }}" = "ubuntu-latest" && echo "true" || echo "false")
-          pytest
-        shell: bash
+          poetry install --no-interaction --no-ansi --without docs
+
+      - name: Install Test Dependencies
+        run: pip install pytest pytest-md pytest-emoji
+
+      - name: Run pytest
+        uses: pavelzw/pytest-action@v2
+        with:
+            verbose: true
+            emoji: true
+            job-summary: true
+            report-title: 'FastEmbed Test Report'

.gitignore

@@ -85,28 +85,8 @@ ipython_config.py
 # pyenv
 #   For a library or package, you might want to ignore these files since the code is
 #   intended to run in multiple environments; otherwise, check them in:
-# .python-version
+.python-version
 
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
-# poetry
-#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
-#   This is especially recommended for binary packages to ensure reproducibility, and is more
-#   commonly ignored for libraries.
-#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
-#poetry.lock
-
-# pdm
-#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
-#pdm.lock
-#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
-#   in version control.
-#   https://pdm.fming.dev/#use-with-ide
 .pdm.toml
 
 # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
@@ -152,27 +132,11 @@ dmypy.json
 # Cython debug symbols
 cython_debug/
 
-# PyCharm
-#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
-#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
-#  and can be added to the global gitignore or merged into this file.  For a more nuclear
-#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
-#.idea/
+.idea/
 .DS_Store
-nbs/*.tar.gz
 *.tar.gz
-nbs/fast-*/*
-local_cache/*/*
-*/local_cache/*/*
-*/*/local_cache/*/*
+**/local_cache/
 docs/experimental/*.parquet
 docs/experimental/*.bin
 qdrant_storage/*
-fooling_around/fast-multilingual-e5-large/config.json
-fooling_around/fast-multilingual-e5-large/model_optimized.onnx
-fooling_around/fast-multilingual-e5-large/model_optimized.onnx.data
-fooling_around/fast-multilingual-e5-large/ort_config.json
-fooling_around/fast-multilingual-e5-large/sentencepiece.bpe.model
-fooling_around/fast-multilingual-e5-large/special_tokens_map.json
-fooling_around/fast-multilingual-e5-large/tokenizer_config.json
-fooling_around/fast-multilingual-e5-large/tokenizer.json
+experiments/models/*

.pre-commit-config.yaml

@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.1.13
+    rev: v0.3.4
     hooks:
       - id: ruff
         types_or: [ python, pyi, jupyter ]

CONTRIBUTING.md

@@ -0,0 +1,78 @@
+# Contributing to FastEmbed!
+
+:+1::tada: First off, thanks for taking the time to contribute! :tada::+1:
+
+The following is a set of guidelines for contributing to FastEmbed. These are mostly guidelines, not rules. Use your best judgment, and feel free to propose changes to this document in a pull request.
+
+## Table Of Contents
+
+[I don't want to read this whole thing, I just have a question!!!](#i-dont-want-to-read-this-whole-thing-i-just-have-a-question)
+
+[How Can I Contribute?](#how-can-i-contribute)
+  * [Your First Code Contribution](#your-first-code-contribution)
+  * [Adding New Models](#adding-new-models)
+
+[Styleguides](#styleguides)
+  * [Code Lint](#code-lint)
+  * [Pre-Commit Hooks](#pre-commit-hooks)
+ 
+## I don't want to read this whole thing I just have a question!!!
+
+> **Note:** Please don't file an issue to ask a question. You'll get faster results by using the resources below:
+
+* [FastEmbed Docs](https://qdrant.github.io/fastembed/)
+* [Qdrant Discord](https://discord.gg/Qy6HCJK9Dc)
+
+## How Can I Contribute?
+
+## How Do I Submit A (Good) Bug Report?
+
+Bugs are tracked as [GitHub issues](https://guides.github.com/features/issues/). 
+
+Explain the problem and include additional details to help maintainers reproduce the problem:
+
+* **Use a clear and descriptive title** for the issue to identify the problem.
+* **Describe the exact steps which reproduce the problem** in as many details as possible. For example, start by explaining how you are using FastEmbed, e.g. with Langchain, Qdrant Client, Llama Index and which command exactly you used. When listing steps, **don't just say what you did, but explain how you did it**.
+* **Provide specific examples to demonstrate the steps**. Include links to files or GitHub projects, or copy/pasteable snippets, which you use in those examples. If you're providing snippets in the issue, use [Markdown code blocks](https://help.github.com/articles/markdown-basics/#multiple-lines).
+* **Describe the behavior you observed after following the steps** and point out what exactly is the problem with that behavior.
+* **Explain which behavior you expected to see instead and why.**
+* **If the problem is related to performance or memory**, include a [call stack profile capture](https://github.com/joerick/pyinstrument) and your observations. 
+
+Include details about your configuration and environment:
+
+* **Which version of FastEmbed are you using?** You can get the exact version by running `python -c "import fastembed; print(fastembed.__version__)"`.
+* **What's the name and version of the OS you're using**?
+* **Which packages do you have installed?** You can get that list by running `pip freeze`
+
+### Your First Code Contribution
+
+Unsure where to begin contributing to FastEmbed? You can start by looking through these `good-first-issue`issues:
+
+* [Good First Issue](https://github.com/qdrant/fastembed/labels/good%20first%20issue) - issues which should only require a few lines of code, and a test or two. These are a great way to get started with FastEmbed. This includes adding new models which are already tested and ready on Huggingface Hub. 
+
+## Pull Requests
+
+The best way to learn about the mechanics of FastEmbed is to start working on it. 
+
+### Your First Code Contribution
+Your first code contribution can be small bug fixes:
+1. This PR adds a small bug fix for a single input: https://github.com/qdrant/fastembed/pull/148
+2. This PR adds a check for the right file location and extension, specific to an OS: https://github.com/qdrant/fastembed/pull/128
+
+Even documentation improvements and tests are most welcome:
+1. This PR fixes a README link: https://github.com/qdrant/fastembed/pull/143
+
+### Adding New Models
+1. Open Requests for New Models are [here](https://github.com/qdrant/fastembed/labels/model%20request). 
+2. There are quite a few pull requests that were merged for this purpose and you can use them as a reference. Here is an example: https://github.com/qdrant/fastembed/pull/129
+3. Make sure to add tests for the new model
+    - The CANONICAL_VECTOR values must come from a reference implementation usually from Huggingface Transformers or Sentence Transformers
+    - Here is a reference [Colab Notebook](https://colab.research.google.com/drive/1tNdV3DsiwsJzu2AXnUnoeF5av1Hp8HF1?usp=sharing) for how we will evaluate whether your VECTOR values in the test are correct or not.
+
+## Styleguides
+
+### Code Lint
+We use ruff for code linting. It should be installed with poetry since it's a dev dependency.
+
+### Pre-Commit Hooks
+We use pre-commit hooks to ensure that the code is linted before it's committed. You can install pre-commit hooks by running `pre-commit install` in the root directory of the project.

README.md

@@ -2,18 +2,15 @@
 
 FastEmbed is a lightweight, fast, Python library built for embedding generation. We [support popular text models](https://qdrant.github.io/fastembed/examples/Supported_Models/). Please [open a GitHub issue](https://github.com/qdrant/fastembed/issues/new) if you want us to add a new model.
 
-The default text embedding (`TextEmbedding`) model is Flag Embedding, the top model in the [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard. It supports "query" and "passage" prefixes for the input text. Here is an example for [Retrieval Embedding Generation](https://qdrant.github.io/fastembed/examples/Retrieval_with_FastEmbed/) and how to use [FastEmbed with Qdrant](https://qdrant.github.io/fastembed/examples/Usage_With_Qdrant/).
+The default text embedding (`TextEmbedding`) model is Flag Embedding, presented in the [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard. It supports "query" and "passage" prefixes for the input text. Here is an example for [Retrieval Embedding Generation](https://qdrant.github.io/fastembed/examples/Retrieval_with_FastEmbed/) and how to use [FastEmbed with Qdrant](https://qdrant.github.io/fastembed/examples/Usage_With_Qdrant/).
 
-1. Light & Fast
-    - Quantized model weights
-    - ONNX Runtime, no PyTorch dependency
-    - CPU-first design
-    - Data-parallelism for encoding of large datasets
+## 📈 Why FastEmbed?
 
-2. Accuracy/Recall
-    - Better than OpenAI Ada-002
-    - Default is Flag Embedding, which is top of the [MTEB](https://huggingface.co/spaces/mteb/leaderboard) leaderboard
-    - List of [supported models](https://qdrant.github.io/fastembed/examples/Supported_Models/) - including multilingual models
+1. Light: FastEmbed is a lightweight library with few external dependencies. We don't require a GPU and don't download GBs of PyTorch dependencies, and instead use the ONNX Runtime. This makes it a great candidate for serverless runtimes like AWS Lambda. 
+
+2. Fast: FastEmbed is designed for speed. We use the ONNX Runtime, which is faster than PyTorch. We also use data-parallelism for encoding large datasets.
+
+3. Accurate: FastEmbed is better than OpenAI Ada-002. We also [supported](https://qdrant.github.io/fastembed/examples/Supported_Models/) an ever expanding set of models, including a few multilingual models.
 
 ## 🚀 Installation
 
@@ -28,16 +25,21 @@ pip install fastembed
 ```python
 from fastembed import TextEmbedding
 from typing import List
-import numpy as np
 
+# Example list of documents
 documents: List[str] = [
-    "passage: Hello, World!",
-    "query: Hello, World!", # these are two different embedding
-    "passage: This is an example passage.",
-    "fastembed is supported by and maintained by Qdrant." # You can leave out the prefix but it's recommended
+    "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.",
 ]
-embedding_model = TextEmbedding(model_name="BAAI/bge-base-en")
-embeddings: List[np.ndarray] = list(embedding_model.embed(documents)) # Note the list() call - this is a generator
+
+# This will trigger the model download and initialization
+embedding_model = TextEmbedding()
+print("The model BAAI/bge-small-en-v1.5 is ready to use.")
+
+embeddings_generator = embedding_model.embed(documents)  # reminder this is a generator
+embeddings_list = list(embedding_model.embed(documents))
+  # you can also convert the generator to a list, and that to a numpy array
+len(embeddings_list[0]) # Vector of 384 dimensions
 ```
 
 ## Usage with Qdrant

docs/Getting Started.ipynb

@@ -11,7 +11,9 @@
     "\n",
     "## Quick Start\n",
     "\n",
-    "The fastembed package is designed to be easy to use. The main class is the `Embedding` class. It takes a list of strings as input and returns a list of vectors as output. The `Embedding` class is initialized with a model file."
+    "The fastembed package is designed to be easy to use. We'll be using `TextEmbedding` class. It takes a list of strings as input and returns an generator of vectors. If you're seeing generators for the first time, don't worry, you can convert it to a list using `list()`.\n",
+    "\n",
+    "> 💡 You can learn more about generators from [Python Wiki](https://wiki.python.org/moin/Generators)"
    ]
   },
   {
@@ -21,15 +23,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!pip install fastembed --upgrade --quiet # Install fastembed "
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "ed81d725",
-   "metadata": {},
-   "source": [
-    "Make the necessary imports, initialize the `Embedding` class, and embed your data into vectors:"
+    "!pip install -Uqq fastembed # Install fastembed"
    ]
   },
   {
@@ -39,43 +33,115 @@
    "metadata": {},
    "outputs": [
     {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 76.7M/76.7M [00:05<00:00, 15.0MiB/s]\n",
-      "100%|██████████| 3/3 [00:00<00:00, 455.37it/s]"
-     ]
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "890cc3b969354eec8d149d143e301a7a",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Fetching 9 files:   0%|          | 0/9 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
     },
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "(384,)\n"
+      "The model BAAI/bge-small-en-v1.5 is ready to use.\n"
      ]
     },
     {
-     "name": "stderr",
+     "data": {
+      "text/plain": [
+       "384"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import numpy as np\n",
+    "from fastembed import TextEmbedding\n",
+    "from typing import List\n",
+    "\n",
+    "# Example list of documents\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",
+    "\n",
+    "# This will trigger the model download and initialization\n",
+    "embedding_model = TextEmbedding()\n",
+    "print(\"The model BAAI/bge-small-en-v1.5 is ready to use.\")\n",
+    "\n",
+    "embeddings_generator = embedding_model.embed(documents)  # reminder this is a generator\n",
+    "embeddings_list = list(embeddings_generator)\n",
+    "# you can also convert the generator to a list, and that to a numpy array\n",
+    "len(embeddings_list[0])  # Vector of 384 dimensions"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d772190b",
+   "metadata": {},
+   "source": [
+    "> 💡 **Why do we use generators?**\n",
+    "> \n",
+    "> We use them to save memory mostly. Instead of loading all the vectors into memory, we can load them one by one. This is useful when you have a large dataset and you don't want to load all the vectors at once."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "8a225cb8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
      "output_type": "stream",
      "text": [
-      "\n"
+      "Document: This is built to be faster and lighter than other embedding libraries e.g. Transformers, Sentence-Transformers, etc.\n",
+      "Vector of type: <class 'numpy.ndarray'> with shape: (384,)\n",
+      "Document: fastembed is supported by and maintained by Qdrant.\n",
+      "Vector of type: <class 'numpy.ndarray'> with shape: (384,)\n"
      ]
     }
    ],
    "source": [
-    "from typing import List\n",
-    "import numpy as np\n",
-    "from fastembed.embedding import DefaultEmbedding\n",
+    "embeddings_generator = embedding_model.embed(documents)  # reminder this is a generator\n",
     "\n",
-    "# Example list of documents\n",
-    "documents: List[str] = [\n",
-    "    \"Hello, World!\",\n",
-    "    \"This is an example document.\",\n",
-    "    \"fastembed is supported by and maintained by Qdrant.\",\n",
-    "]\n",
-    "# Initialize the DefaultEmbedding class\n",
-    "embedding_model = DefaultEmbedding()\n",
-    "embeddings: List[np.ndarray] = list(embedding_model.embed(documents))\n",
-    "print(embeddings[0].shape)"
+    "for doc, vector in zip(documents, embeddings_generator):\n",
+    "    print(\"Document:\", doc)\n",
+    "    print(f\"Vector of type: {type(vector)} with shape: {vector.shape}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "769a1be9",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "(2, 384)"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "embeddings_list = np.array(\n",
+    "    list(embedding_model.embed(documents))\n",
+    ")  # you can also convert the generator to a list, and that to a numpy array\n",
+    "embeddings_list.shape"
    ]
   },
   {
@@ -83,142 +149,74 @@
    "id": "8c49ae50",
    "metadata": {},
    "source": [
-    "## Let's think step by step"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "92cf4b76",
-   "metadata": {},
-   "source": [
-    "### Setup\n",
-    "\n",
-    "Importing the required classes and modules:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "id": "c0a6f634",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from typing import List\n",
-    "import numpy as np\n",
-    "from fastembed.embedding import DefaultEmbedding"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "3fd03a71",
-   "metadata": {},
-   "source": [
-    "Notice that we are using the DefaultEmbedding -- which is a quantized, state of the Art Flag Embedding model which beats OpenAI's Embedding by a large margin. \n",
-    "\n",
-    "### Prepare your Documents\n",
-    "You can define a list of documents that you'd like to embed. These can be sentences, paragraphs, or even entire documents. \n",
+    "We're using [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) a state of the art Flag Embedding model. The model does better than OpenAI text-embedding-ada-002. We've made it even faster by converting it to ONNX format and quantizing the model for you.\n",
     "\n",
     "#### Format of the Document List\n",
+    "\n",
     "1. List of Strings: Your documents must be in a list, and each document must be a string\n",
-    "2. For Retrieval Tasks: If you're working with queries and passages, you can add special labels to them:\n",
+    "2. For Retrieval Tasks with our default: If you're working with queries and passages, you can add special labels to them:\n",
     "- **Queries**: Add \"query:\" at the beginning of each query string\n",
-    "- **Passages**: Add \"passage:\" at the beginning of each passage string"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "id": "145a56ce",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Example list of documents\n",
-    "documents: List[str] = [\n",
-    "    \"passage: Hello, World!\",\n",
-    "    \"query: Hello, World!\",  # these are two different embedding\n",
-    "    \"passage: This is an example passage.\",\n",
-    "    # You can leave out the prefix but it's recommended\n",
-    "    \"fastembed is supported by and maintained by Qdrant.\",\n",
-    "]"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "1cb3cc87",
-   "metadata": {},
-   "source": [
-    "### Load the Embedding Model Weights\n",
-    "Next, initialize the Embedding class with the desired parameters. Here, \"BAAI/bge-small-en\" is the pre-trained model name, and max_length=512 is the maximum token length for each document.\n",
+    "- **Passages**: Add \"passage:\" at the beginning of each passage string\n",
     "\n",
-    "This will download the model weights, decompress to directory `local_cache` and load them into the Embedding class.\n",
+    "## Beyond the default model\n",
     "\n",
-    "#### Initialize DefaultEmbedding\n",
-    "\n",
-    "We will initialize Flag Embeddings with the model name and the maximum token length. That is the DefaultEmbedding class with the model name \"BAAI/bge-small-en\" and max_length=512."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "id": "272c8915",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "embedding_model = DefaultEmbedding()"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "5549d501",
-   "metadata": {},
-   "source": [
-    "### Embed your Documents\n",
-    "\n",
-    "Use the embed method of the embedding model to transform the documents into a List of np.array. The method returns a generator, so we cast it to a list to get the embeddings."
+    "The default model is built for speed and efficiency. If you need a more accurate model, you can use the `TextEmbedding` class to load any model from our list of available models. You can find the list of available models using `TextEmbedding.list_supported_models()`."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": 5,
-   "id": "8013eee9",
+   "id": "2e9c8766",
    "metadata": {},
    "outputs": [
     {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 4/4 [00:00<00:00, 361.82it/s]\n"
-     ]
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "9470ec542f3c4400a42452c2489a1abc",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Fetching 8 files:   0%|          | 0/8 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
     }
    ],
    "source": [
-    "embeddings: List[np.ndarray] = list(embedding_model.embed(documents))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "e5b5a6ad",
-   "metadata": {},
-   "source": [
-    "You can print the shape of the embeddings to understand their dimensions. Typically, the shape will indicate the number of dimensions in the vector."
+    "multilingual_large_model = TextEmbedding(\"intfloat/multilingual-e5-large\")  # This can take a few minutes to download"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": 6,
-   "id": "0d8c8e08",
+   "id": "a9e70f0e",
    "metadata": {},
    "outputs": [
     {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "(384,)\n"
-     ]
+     "data": {
+      "text/plain": [
+       "(4, 1024)"
+      ]
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
     }
    ],
    "source": [
-    "print(embeddings[0].shape)  # (384,) or similar output"
+    "np.array(\n",
+    "    list(multilingual_large_model.embed([\"Hello, world!\", \"你好世界\", \"¡Hola Mundo!\", \"नमस्ते!\"]))\n",
+    ").shape  # Vector of 1024 dimensions"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "64fe20ed",
+   "metadata": {},
+   "source": [
+    "Next: Checkout how to use FastEmbed with Qdrant for similarity search: [FastEmbed with Qdrant](https://qdrant.github.io/fastembed/examples/Usage_With_Qdrant/)"
    ]
   }
  ],
@@ -238,7 +236,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.17"
+   "version": "3.10.13"
   }
  },
  "nbformat": 4,

docs/examples/SPLADE_with_FastEmbed.ipynb

@@ -0,0 +1,384 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Introduction to SPLADE with FastEmbed\n",
+    "\n",
+    "In this notebook, we will explore how to generate Sparse Vectors -- in particular a variant of the [SPLADE](https://arxiv.org/abs/2107.05720).\n",
+    "\n",
+    "> 💡 The original [naver/SPLADE](https://github.com/naver/splade) models were licensed CC BY-NC-SA 4.0 -- Not for Commercial Use. This [SPLADE++](https://huggingface.co/prithivida/Splade_PP_en_v1) model is Apache License and hence, licensed for commercial use. \n",
+    "\n",
+    "## Outline:\n",
+    "1. [What is SPLADE?](#What-is-SPLADE?)\n",
+    "2. [Setting up the environment](#Setting-up-the-environment)\n",
+    "3. [Generating SPLADE vectors with FastEmbed](#Generating-SPLADE-vectors-with-FastEmbed)\n",
+    "4. [Understanding SPLADE vectors](#Understanding-SPLADE-vectors)\n",
+    "5. [Observations and Design Choices](#Observations-and-Model-Design-Choices)\n",
+    "\n",
+    "\n",
+    "## What is SPLADE?\n",
+    "\n",
+    "SPLADE was a novel method for _learning_ sparse vectors for text representation. This model beats BM25 -- the underlying approach for the Elastic/Lucene family of implementations. Thus making it highly effective for tasks such as information retrieval, document classification, and more. \n",
+    "\n",
+    "The key advantage of SPLADE is its ability to generate sparse vectors, which are more efficient and interpretable than dense vectors. This makes SPLADE a powerful tool for handling large-scale text data.\n",
+    "\n",
+    "## Setting up the environment\n",
+    "\n",
+    "This notebook uses few dependencies, which are installed below: "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# !pip install -q fastembed"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Let's get started! 🚀"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:20.516644Z",
+     "start_time": "2024-03-30T00:49:20.188543Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "from fastembed import SparseTextEmbedding, SparseEmbedding\n",
+    "from typing import List"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "> You can find the list of all supported Sparse Embedding models by calling this API: `SparseTextEmbedding.list_supported_models()`"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:22.366294Z",
+     "start_time": "2024-03-30T00:49:22.362384Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'model': 'prithvida/Splade_PP_en_v1',\n",
+       "  'vocab_size': 30522,\n",
+       "  'description': 'Misspelled version of the model. Retained for backward compatibility. Independent Implementation of SPLADE++ Model for English',\n",
+       "  'size_in_GB': 0.532,\n",
+       "  'sources': {'hf': 'Qdrant/SPLADE_PP_en_v1'}},\n",
+       " {'model': 'prithivida/Splade_PP_en_v1',\n",
+       "  'vocab_size': 30522,\n",
+       "  'description': 'Independent Implementation of SPLADE++ Model for English',\n",
+       "  'size_in_GB': 0.532,\n",
+       "  'sources': {'hf': 'Qdrant/SPLADE_PP_en_v1'}}]"
+      ]
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "SparseTextEmbedding.list_supported_models()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:27.193530Z",
+     "start_time": "2024-03-30T00:49:26.139248Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "2aa47b26ab01475e8d3577433037f685",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Fetching 9 files:   0%|          | 0/9 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "model_name = \"prithvida/Splade_PP_en_v1\"\n",
+    "# This triggers the model download\n",
+    "model = SparseTextEmbedding(model_name=model_name)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:28.624109Z",
+     "start_time": "2024-03-30T00:49:28.399960Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "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",
+    "    \"Chandrayaan-3 will be launched by the Indian Space Research Organisation (ISRO)\",\n",
+    "    \"The estimated cost of the mission is around $35 million\",\n",
+    "    \"It will carry instruments to study the lunar surface and atmosphere\",\n",
+    "    \"Chandrayaan-3 landed on the Moon's surface on 23rd August 2023\",\n",
+    "    \"It consists of a lander named Vikram and a rover named Pragyan similar to Chandrayaan-2. Its propulsion module would act like an orbiter.\",\n",
+    "    \"The propulsion module carries the lander and rover configuration until the spacecraft is in a 100-kilometre (62 mi) lunar orbit\",\n",
+    "    \"The mission used GSLV Mk III rocket for its launch\",\n",
+    "    \"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",
+    "    model.embed(documents, batch_size=6)\n",
+    ")  # batch_size is optional, notice the generator"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:29.646340Z",
+     "start_time": "2024-03-30T00:49:29.643411Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "SparseEmbedding(values=array([0.05297208, 0.01963477, 0.36459631, 1.38508618, 0.71776593,\n",
+       "       0.12667948, 0.46230844, 0.446771  , 0.26897505, 1.01519883,\n",
+       "       1.5655334 , 0.29412213, 1.53102326, 0.59785569, 1.1001817 ,\n",
+       "       0.02079751, 0.09955651, 0.44249091, 0.09747757, 1.53519952,\n",
+       "       1.36765671, 0.15740395, 0.49882549, 0.38629025, 0.76612782,\n",
+       "       1.25805044, 0.39058095, 0.27236196, 0.45152301, 0.48262018,\n",
+       "       0.26085234, 1.35912788, 0.70710695, 1.71639752]), indices=array([ 1010,  1011,  1016,  1017,  2001,  2018,  2034,  2093,  2117,\n",
+       "        2319,  2353,  2509,  2634,  2686,  2796,  2817,  2922,  2959,\n",
+       "        3003,  3148,  3260,  3390,  3462,  3523,  3822,  4231,  4316,\n",
+       "        4774,  5590,  5871,  6416, 11926, 12076, 16469]))"
+      ]
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "index = 0\n",
+    "sparse_embeddings_list[index]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "The previous output is a SparseEmbedding object for the first document in our list.\n",
+    "\n",
+    "It contains two arrays: values and indices. \n",
+    "- The 'values' array represents the weights of the features (tokens) in the document.\n",
+    "- The 'indices' array represents the indices of these features in the model's vocabulary.\n",
+    "\n",
+    "Each pair of corresponding values and indices represents a token and its weight in the document."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:31.549533Z",
+     "start_time": "2024-03-30T00:49:31.546398Z"
+    }
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Token at index 1010 has weight 0.05297207832336426\n",
+      "Token at index 1011 has weight 0.01963476650416851\n",
+      "Token at index 1016 has weight 0.36459630727767944\n",
+      "Token at index 1017 has weight 1.385086178779602\n",
+      "Token at index 2001 has weight 0.7177659273147583\n"
+     ]
+    }
+   ],
+   "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]}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Understanding SPLADE vectors\n",
+    "\n",
+    "This is still a little abstract, so let's use the tokenizer vocab to make sense of these indices."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:36.203640Z",
+     "start_time": "2024-03-30T00:49:34.889654Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "import json\n",
+    "from transformers import AutoTokenizer\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(SparseTextEmbedding.list_supported_models()[0][\"sources\"][\"hf\"])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T00:49:36.210049Z",
+     "start_time": "2024-03-30T00:49:36.206825Z"
+    }
+   },
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "{\n",
+      "    \"chandra\": 1.7163975238800049,\n",
+      "    \"third\": 1.5655333995819092,\n",
+      "    \"##ya\": 1.535199522972107,\n",
+      "    \"india\": 1.5310232639312744,\n",
+      "    \"3\": 1.385086178779602,\n",
+      "    \"mission\": 1.3676567077636719,\n",
+      "    \"lunar\": 1.3591278791427612,\n",
+      "    \"moon\": 1.2580504417419434,\n",
+      "    \"indian\": 1.1001816987991333,\n",
+      "    \"##an\": 1.015198826789856,\n",
+      "    \"3rd\": 0.7661278247833252,\n",
+      "    \"was\": 0.7177659273147583,\n",
+      "    \"spacecraft\": 0.7071069478988647,\n",
+      "    \"space\": 0.5978556871414185,\n",
+      "    \"flight\": 0.4988254904747009,\n",
+      "    \"satellite\": 0.4826201796531677,\n",
+      "    \"first\": 0.46230843663215637,\n",
+      "    \"expedition\": 0.4515230059623718,\n",
+      "    \"three\": 0.4467709958553314,\n",
+      "    \"fourth\": 0.44249090552330017,\n",
+      "    \"vehicle\": 0.390580952167511,\n",
+      "    \"iii\": 0.3862902522087097,\n",
+      "    \"2\": 0.36459630727767944,\n",
+      "    \"##3\": 0.2941221296787262,\n",
+      "    \"planet\": 0.27236196398735046,\n",
+      "    \"second\": 0.26897504925727844,\n",
+      "    \"missions\": 0.2608523368835449,\n",
+      "    \"launched\": 0.15740394592285156,\n",
+      "    \"had\": 0.12667948007583618,\n",
+      "    \"largest\": 0.09955651313066483,\n",
+      "    \"leader\": 0.09747757017612457,\n",
+      "    \",\": 0.05297207832336426,\n",
+      "    \"study\": 0.02079751156270504,\n",
+      "    \"-\": 0.01963476650416851\n",
+      "}\n"
+     ]
+    }
+   ],
+   "source": [
+    "def get_tokens_and_weights(sparse_embedding, tokenizer):\n",
+    "    token_weight_dict = {}\n",
+    "    for i in range(len(sparse_embedding.indices)):\n",
+    "        token = tokenizer.decode([sparse_embedding.indices[i]])\n",
+    "        weight = sparse_embedding.values[i]\n",
+    "        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",
+    "    return token_weight_dict\n",
+    "\n",
+    "\n",
+    "# Test the function with the first SparseEmbedding\n",
+    "print(json.dumps(get_tokens_and_weights(sparse_embeddings_list[index], tokenizer), indent=4))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Observations and Model Design Choices\n",
+    "\n",
+    "1. The relative order of importance is quite useful. The most important tokens in the sentence have the highest weights.\n",
+    "1. **Term Expansion**: The model can expand the terms in the document. This means that the model can generate weights for tokens that are not present in the document but are related to the tokens in the document. This is a powerful feature that allows the model to capture the context of the document. Here, you'll see that the model has added the tokens '3' from 'third' and 'moon' from 'lunar' to the sparse vector.\n",
+    "\n",
+    "### Design Choices\n",
+    "\n",
+    "1. The weights are not normalized. This means that the sum of the weights is not 1 or 100. This is a common practice in sparse embeddings, as it allows the model to capture the importance of each token in the document.\n",
+    "1. Tokens are included in the sparse vector only if they are present in the model's vocabulary. This means that the model will not generate a weight for tokens that it has not seen during training.\n",
+    "1. Tokens do not map to words directly -- allowing you to gracefully handle typo errors and out-of-vocabulary tokens."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {
+    "collapsed": false
+   },
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "fst",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

docs/examples/Supported_Models.ipynb

@@ -3,7 +3,12 @@
   {
    "cell_type": "code",
    "execution_count": 1,
-   "metadata": {},
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T11:18:52.052764Z",
+     "start_time": "2024-03-30T11:18:52.039616Z"
+    }
+   },
    "outputs": [],
    "source": [
     "%load_ext autoreload\n",
@@ -12,7 +17,25 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 6,
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pandas as pd\n",
+    "\n",
+    "from fastembed import SparseTextEmbedding, TextEmbedding"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Supported Text Embedding Models"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
    "metadata": {},
    "outputs": [
     {
@@ -40,233 +63,316 @@
        "      <th>dim</th>\n",
        "      <th>description</th>\n",
        "      <th>size_in_GB</th>\n",
-       "      <th>sources</th>\n",
        "    </tr>\n",
        "  </thead>\n",
        "  <tbody>\n",
        "    <tr>\n",
        "      <th>0</th>\n",
-       "      <td>BAAI/bge-base-en</td>\n",
-       "      <td>768</td>\n",
-       "      <td>Base English model</td>\n",
-       "      <td>0.50</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz'}</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>1</th>\n",
-       "      <td>BAAI/bge-base-en-v1.5</td>\n",
-       "      <td>768</td>\n",
-       "      <td>Base English model, v1.5</td>\n",
-       "      <td>0.44</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en-v1.5.tar.gz', 'hf': 'qdrant/bge-base-en-v1.5-onnx-q'}</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>2</th>\n",
-       "      <td>BAAI/bge-large-en-v1.5-quantized</td>\n",
-       "      <td>1024</td>\n",
-       "      <td>Large English model, v1.5</td>\n",
-       "      <td>1.34</td>\n",
-       "      <td>{'hf': 'qdrant/bge-large-en-v1.5-onnx-q'}</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>3</th>\n",
-       "      <td>BAAI/bge-large-en-v1.5</td>\n",
-       "      <td>1024</td>\n",
-       "      <td>Large English model, v1.5</td>\n",
-       "      <td>1.34</td>\n",
-       "      <td>{'hf': 'qdrant/bge-large-en-v1.5-onnx'}</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>4</th>\n",
-       "      <td>BAAI/bge-small-en</td>\n",
-       "      <td>384</td>\n",
-       "      <td>Fast English model</td>\n",
-       "      <td>0.20</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz'}</td>\n",
-       "    </tr>\n",
-       "    <tr>\n",
-       "      <th>5</th>\n",
        "      <td>BAAI/bge-small-en-v1.5</td>\n",
        "      <td>384</td>\n",
        "      <td>Fast and Default English model</td>\n",
-       "      <td>0.13</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-en-v1.5.tar.gz', 'hf': 'qdrant/bge-small-en-v1.5-onnx-q'}</td>\n",
+       "      <td>0.067</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>6</th>\n",
+       "      <th>1</th>\n",
        "      <td>BAAI/bge-small-zh-v1.5</td>\n",
        "      <td>512</td>\n",
        "      <td>Fast and recommended Chinese model</td>\n",
-       "      <td>0.10</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-zh-v1.5.tar.gz'}</td>\n",
+       "      <td>0.090</td>\n",
        "    </tr>\n",
        "    <tr>\n",
-       "      <th>7</th>\n",
+       "      <th>2</th>\n",
        "      <td>sentence-transformers/all-MiniLM-L6-v2</td>\n",
        "      <td>384</td>\n",
        "      <td>Sentence Transformer model, MiniLM-L6-v2</td>\n",
-       "      <td>0.09</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/sentence-transformers-all-MiniLM-L6-v2.tar.gz', 'hf': 'qdrant/all-MiniLM-L6-v2-onnx'}</td>\n",
+       "      <td>0.090</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>3</th>\n",
+       "      <td>snowflake/snowflake-arctic-embed-xs</td>\n",
+       "      <td>384</td>\n",
+       "      <td>Based on all-MiniLM-L6-v2 model with only 22m ...</td>\n",
+       "      <td>0.090</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>4</th>\n",
+       "      <td>jinaai/jina-embeddings-v2-small-en</td>\n",
+       "      <td>512</td>\n",
+       "      <td>English embedding model supporting 8192 sequen...</td>\n",
+       "      <td>0.120</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5</th>\n",
+       "      <td>snowflake/snowflake-arctic-embed-s</td>\n",
+       "      <td>384</td>\n",
+       "      <td>Based on infloat/e5-small-unsupervised, does n...</td>\n",
+       "      <td>0.130</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>6</th>\n",
+       "      <td>BAAI/bge-small-en</td>\n",
+       "      <td>384</td>\n",
+       "      <td>Fast English model</td>\n",
+       "      <td>0.130</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>7</th>\n",
+       "      <td>BAAI/bge-base-en-v1.5</td>\n",
+       "      <td>768</td>\n",
+       "      <td>Base English model, v1.5</td>\n",
+       "      <td>0.210</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>8</th>\n",
-       "      <td>nomic-ai/nomic-embed-text-v1</td>\n",
-       "      <td>768</td>\n",
-       "      <td>8192 context length english model</td>\n",
-       "      <td>0.54</td>\n",
-       "      <td>{'hf': 'nomic-ai/nomic-embed-text-v1'}</td>\n",
+       "      <td>sentence-transformers/paraphrase-multilingual-...</td>\n",
+       "      <td>384</td>\n",
+       "      <td>Sentence Transformer model, paraphrase-multili...</td>\n",
+       "      <td>0.220</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>9</th>\n",
-       "      <td>nomic-ai/nomic-embed-text-v1.5</td>\n",
+       "      <td>BAAI/bge-base-en</td>\n",
        "      <td>768</td>\n",
-       "      <td>8192 context length english model</td>\n",
-       "      <td>0.54</td>\n",
-       "      <td>{'hf': 'nomic-ai/nomic-embed-text-v1.5'}</td>\n",
+       "      <td>Base English model</td>\n",
+       "      <td>0.420</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>10</th>\n",
-       "      <td>thenlper/gte-large</td>\n",
-       "      <td>1024</td>\n",
-       "      <td>Large general text embeddings model</td>\n",
-       "      <td>1.34</td>\n",
-       "      <td>{'hf': 'qdrant/gte-large-onnx'}</td>\n",
+       "      <td>snowflake/snowflake-arctic-embed-m</td>\n",
+       "      <td>768</td>\n",
+       "      <td>Based on intfloat/e5-base-unsupervised model, ...</td>\n",
+       "      <td>0.430</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>11</th>\n",
-       "      <td>intfloat/multilingual-e5-large</td>\n",
-       "      <td>1024</td>\n",
-       "      <td>Multilingual model, e5-large. Recommend using this model for non-English languages</td>\n",
-       "      <td>2.24</td>\n",
-       "      <td>{'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-multilingual-e5-large.tar.gz', 'hf': 'qdrant/multilingual-e5-large-onnx'}</td>\n",
+       "      <td>jinaai/jina-embeddings-v2-base-en</td>\n",
+       "      <td>768</td>\n",
+       "      <td>English embedding model supporting 8192 sequen...</td>\n",
+       "      <td>0.520</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>12</th>\n",
-       "      <td>sentence-transformers/paraphrase-multilingual-mpnet-base-v2</td>\n",
+       "      <td>nomic-ai/nomic-embed-text-v1</td>\n",
        "      <td>768</td>\n",
-       "      <td>Sentence-transformers model for tasks like clustering or semantic search</td>\n",
-       "      <td>1.11</td>\n",
-       "      <td>{'hf': 'xenova/paraphrase-multilingual-mpnet-base-v2'}</td>\n",
+       "      <td>8192 context length english model</td>\n",
+       "      <td>0.520</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>13</th>\n",
-       "      <td>sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2</td>\n",
-       "      <td>384</td>\n",
-       "      <td>Sentence Transformer model, paraphrase-multilingual-MiniLM-L12-v2</td>\n",
-       "      <td>0.46</td>\n",
-       "      <td>{'hf': 'qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q'}</td>\n",
+       "      <td>nomic-ai/nomic-embed-text-v1.5</td>\n",
+       "      <td>768</td>\n",
+       "      <td>8192 context length english model</td>\n",
+       "      <td>0.520</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>14</th>\n",
-       "      <td>jinaai/jina-embeddings-v2-base-en</td>\n",
+       "      <td>snowflake/snowflake-arctic-embed-m-long</td>\n",
        "      <td>768</td>\n",
-       "      <td>English embedding model supporting 8192 sequence length</td>\n",
-       "      <td>0.55</td>\n",
-       "      <td>{'hf': 'xenova/jina-embeddings-v2-base-en'}</td>\n",
+       "      <td>Based on nomic-ai/nomic-embed-text-v1-unsuperv...</td>\n",
+       "      <td>0.540</td>\n",
        "    </tr>\n",
        "    <tr>\n",
        "      <th>15</th>\n",
-       "      <td>jinaai/jina-embeddings-v2-small-en</td>\n",
-       "      <td>512</td>\n",
-       "      <td>English embedding model supporting 8192 sequence length</td>\n",
-       "      <td>0.13</td>\n",
-       "      <td>{'hf': 'xenova/jina-embeddings-v2-small-en'}</td>\n",
+       "      <td>mixedbread-ai/mxbai-embed-large-v1</td>\n",
+       "      <td>1024</td>\n",
+       "      <td>MixedBread Base sentence embedding model, does...</td>\n",
+       "      <td>0.640</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>16</th>\n",
+       "      <td>sentence-transformers/paraphrase-multilingual-...</td>\n",
+       "      <td>768</td>\n",
+       "      <td>Sentence-transformers model for tasks like clu...</td>\n",
+       "      <td>1.000</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>17</th>\n",
+       "      <td>snowflake/snowflake-arctic-embed-l</td>\n",
+       "      <td>1024</td>\n",
+       "      <td>Based on intfloat/e5-large-unsupervised, large...</td>\n",
+       "      <td>1.020</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>18</th>\n",
+       "      <td>BAAI/bge-large-en-v1.5</td>\n",
+       "      <td>1024</td>\n",
+       "      <td>Large English model, v1.5</td>\n",
+       "      <td>1.200</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>19</th>\n",
+       "      <td>thenlper/gte-large</td>\n",
+       "      <td>1024</td>\n",
+       "      <td>Large general text embeddings model</td>\n",
+       "      <td>1.200</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>20</th>\n",
+       "      <td>intfloat/multilingual-e5-large</td>\n",
+       "      <td>1024</td>\n",
+       "      <td>Multilingual model, e5-large. Recommend using ...</td>\n",
+       "      <td>2.240</td>\n",
        "    </tr>\n",
        "  </tbody>\n",
        "</table>\n",
        "</div>"
       ],
       "text/plain": [
-       "                                                          model   dim  \\\n",
-       "0                                              BAAI/bge-base-en   768   \n",
-       "1                                         BAAI/bge-base-en-v1.5   768   \n",
-       "2                              BAAI/bge-large-en-v1.5-quantized  1024   \n",
-       "3                                        BAAI/bge-large-en-v1.5  1024   \n",
-       "4                                             BAAI/bge-small-en   384   \n",
-       "5                                        BAAI/bge-small-en-v1.5   384   \n",
-       "6                                        BAAI/bge-small-zh-v1.5   512   \n",
-       "7                        sentence-transformers/all-MiniLM-L6-v2   384   \n",
-       "8                                  nomic-ai/nomic-embed-text-v1   768   \n",
-       "9                                nomic-ai/nomic-embed-text-v1.5   768   \n",
-       "10                                           thenlper/gte-large  1024   \n",
-       "11                               intfloat/multilingual-e5-large  1024   \n",
-       "12  sentence-transformers/paraphrase-multilingual-mpnet-base-v2   768   \n",
-       "13  sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2   384   \n",
-       "14                            jinaai/jina-embeddings-v2-base-en   768   \n",
-       "15                           jinaai/jina-embeddings-v2-small-en   512   \n",
+       "                                                model   dim  \\\n",
+       "0                              BAAI/bge-small-en-v1.5   384   \n",
+       "1                              BAAI/bge-small-zh-v1.5   512   \n",
+       "2              sentence-transformers/all-MiniLM-L6-v2   384   \n",
+       "3                 snowflake/snowflake-arctic-embed-xs   384   \n",
+       "4                  jinaai/jina-embeddings-v2-small-en   512   \n",
+       "5                  snowflake/snowflake-arctic-embed-s   384   \n",
+       "6                                   BAAI/bge-small-en   384   \n",
+       "7                               BAAI/bge-base-en-v1.5   768   \n",
+       "8   sentence-transformers/paraphrase-multilingual-...   384   \n",
+       "9                                    BAAI/bge-base-en   768   \n",
+       "10                 snowflake/snowflake-arctic-embed-m   768   \n",
+       "11                  jinaai/jina-embeddings-v2-base-en   768   \n",
+       "12                       nomic-ai/nomic-embed-text-v1   768   \n",
+       "13                     nomic-ai/nomic-embed-text-v1.5   768   \n",
+       "14            snowflake/snowflake-arctic-embed-m-long   768   \n",
+       "15                 mixedbread-ai/mxbai-embed-large-v1  1024   \n",
+       "16  sentence-transformers/paraphrase-multilingual-...   768   \n",
+       "17                 snowflake/snowflake-arctic-embed-l  1024   \n",
+       "18                             BAAI/bge-large-en-v1.5  1024   \n",
+       "19                                 thenlper/gte-large  1024   \n",
+       "20                     intfloat/multilingual-e5-large  1024   \n",
        "\n",
-       "                                                                           description  \\\n",
-       "0                                                                   Base English model   \n",
-       "1                                                             Base English model, v1.5   \n",
-       "2                                                            Large English model, v1.5   \n",
-       "3                                                            Large English model, v1.5   \n",
-       "4                                                                   Fast English model   \n",
-       "5                                                       Fast and Default English model   \n",
-       "6                                                   Fast and recommended Chinese model   \n",
-       "7                                             Sentence Transformer model, MiniLM-L6-v2   \n",
-       "8                                                    8192 context length english model   \n",
-       "9                                                    8192 context length english model   \n",
-       "10                                                 Large general text embeddings model   \n",
-       "11  Multilingual model, e5-large. Recommend using this model for non-English languages   \n",
-       "12            Sentence-transformers model for tasks like clustering or semantic search   \n",
-       "13                   Sentence Transformer model, paraphrase-multilingual-MiniLM-L12-v2   \n",
-       "14                             English embedding model supporting 8192 sequence length   \n",
-       "15                             English embedding model supporting 8192 sequence length   \n",
-       "\n",
-       "    size_in_GB  \\\n",
-       "0         0.50   \n",
-       "1         0.44   \n",
-       "2         1.34   \n",
-       "3         1.34   \n",
-       "4         0.20   \n",
-       "5         0.13   \n",
-       "6         0.10   \n",
-       "7         0.09   \n",
-       "8         0.54   \n",
-       "9         0.54   \n",
-       "10        1.34   \n",
-       "11        2.24   \n",
-       "12        1.11   \n",
-       "13        0.46   \n",
-       "14        0.55   \n",
-       "15        0.13   \n",
-       "\n",
-       "                                                                                                                                           sources  \n",
-       "0                                                               {'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz'}  \n",
-       "1                  {'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en-v1.5.tar.gz', 'hf': 'qdrant/bge-base-en-v1.5-onnx-q'}  \n",
-       "2                                                                                                        {'hf': 'qdrant/bge-large-en-v1.5-onnx-q'}  \n",
-       "3                                                                                                          {'hf': 'qdrant/bge-large-en-v1.5-onnx'}  \n",
-       "4                                                              {'url': 'https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz'}  \n",
-       "5                {'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-en-v1.5.tar.gz', 'hf': 'qdrant/bge-small-en-v1.5-onnx-q'}  \n",
-       "6                                                         {'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-zh-v1.5.tar.gz'}  \n",
-       "7   {'url': 'https://storage.googleapis.com/qdrant-fastembed/sentence-transformers-all-MiniLM-L6-v2.tar.gz', 'hf': 'qdrant/all-MiniLM-L6-v2-onnx'}  \n",
-       "8                                                                                                           {'hf': 'nomic-ai/nomic-embed-text-v1'}  \n",
-       "9                                                                                                         {'hf': 'nomic-ai/nomic-embed-text-v1.5'}  \n",
-       "10                                                                                                                 {'hf': 'qdrant/gte-large-onnx'}  \n",
-       "11         {'url': 'https://storage.googleapis.com/qdrant-fastembed/fast-multilingual-e5-large.tar.gz', 'hf': 'qdrant/multilingual-e5-large-onnx'}  \n",
-       "12                                                                                          {'hf': 'xenova/paraphrase-multilingual-mpnet-base-v2'}  \n",
-       "13                                                                                   {'hf': 'qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q'}  \n",
-       "14                                                                                                     {'hf': 'xenova/jina-embeddings-v2-base-en'}  \n",
-       "15                                                                                                    {'hf': 'xenova/jina-embeddings-v2-small-en'}  "
+       "                                          description  size_in_GB  \n",
+       "0                      Fast and Default English model       0.067  \n",
+       "1                  Fast and recommended Chinese model       0.090  \n",
+       "2            Sentence Transformer model, MiniLM-L6-v2       0.090  \n",
+       "3   Based on all-MiniLM-L6-v2 model with only 22m ...       0.090  \n",
+       "4   English embedding model supporting 8192 sequen...       0.120  \n",
+       "5   Based on infloat/e5-small-unsupervised, does n...       0.130  \n",
+       "6                                  Fast English model       0.130  \n",
+       "7                            Base English model, v1.5       0.210  \n",
+       "8   Sentence Transformer model, paraphrase-multili...       0.220  \n",
+       "9                                  Base English model       0.420  \n",
+       "10  Based on intfloat/e5-base-unsupervised model, ...       0.430  \n",
+       "11  English embedding model supporting 8192 sequen...       0.520  \n",
+       "12                  8192 context length english model       0.520  \n",
+       "13                  8192 context length english model       0.520  \n",
+       "14  Based on nomic-ai/nomic-embed-text-v1-unsuperv...       0.540  \n",
+       "15  MixedBread Base sentence embedding model, does...       0.640  \n",
+       "16  Sentence-transformers model for tasks like clu...       1.000  \n",
+       "17  Based on intfloat/e5-large-unsupervised, large...       1.020  \n",
+       "18                          Large English model, v1.5       1.200  \n",
+       "19                Large general text embeddings model       1.200  \n",
+       "20  Multilingual model, e5-large. Recommend using ...       2.240  "
       ]
      },
-     "execution_count": 6,
+     "execution_count": 2,
      "metadata": {},
      "output_type": "execute_result"
     }
    ],
    "source": [
-    "from fastembed import TextEmbedding\n",
-    "import pandas as pd\n",
-    "\n",
-    "pd.set_option(\"display.max_colwidth\", None)\n",
-    "pd.DataFrame(TextEmbedding.list_supported_models())"
+    "supported_models = (\n",
+    "    pd.DataFrame(TextEmbedding.list_supported_models())\n",
+    "    .sort_values(\"size_in_GB\")\n",
+    "    .drop(columns=\"sources\")\n",
+    "    .reset_index(drop=True)\n",
+    ")\n",
+    "supported_models"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Supported Sparse Text Embedding Models"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-03-30T11:19:01.564291Z",
+     "start_time": "2024-03-30T11:19:01.538768Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>model</th>\n",
+       "      <th>vocab_size</th>\n",
+       "      <th>description</th>\n",
+       "      <th>size_in_GB</th>\n",
+       "      <th>sources</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>0</th>\n",
+       "      <td>prithvida/Splade_PP_en_v1</td>\n",
+       "      <td>30522</td>\n",
+       "      <td>Misspelled version of the model. Retained for ...</td>\n",
+       "      <td>0.532</td>\n",
+       "      <td>{'hf': 'Qdrant/SPLADE_PP_en_v1'}</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>1</th>\n",
+       "      <td>prithivida/Splade_PP_en_v1</td>\n",
+       "      <td>30522</td>\n",
+       "      <td>Independent Implementation of SPLADE++ Model f...</td>\n",
+       "      <td>0.532</td>\n",
+       "      <td>{'hf': 'Qdrant/SPLADE_PP_en_v1'}</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "                        model  vocab_size  \\\n",
+       "0   prithvida/Splade_PP_en_v1       30522   \n",
+       "1  prithivida/Splade_PP_en_v1       30522   \n",
+       "\n",
+       "                                         description  size_in_GB  \\\n",
+       "0  Misspelled version of the model. Retained for ...       0.532   \n",
+       "1  Independent Implementation of SPLADE++ Model f...       0.532   \n",
+       "\n",
+       "                            sources  \n",
+       "0  {'hf': 'Qdrant/SPLADE_PP_en_v1'}  \n",
+       "1  {'hf': 'Qdrant/SPLADE_PP_en_v1'}  "
+      ]
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "pd.DataFrame(SparseTextEmbedding.list_supported_models())"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "fst",
+   "display_name": "Python 3.8.18 ('base')",
    "language": "python",
    "name": "python3"
   },
@@ -282,7 +388,12 @@
    "pygments_lexer": "ipython3",
    "version": "3.10.13"
   },
-  "orig_nbformat": 4
+  "orig_nbformat": 4,
+  "vscode": {
+   "interpreter": {
+    "hash": "c4a27af61e455bc18dcf16f5867a2ff0402fa12b01dd0f6ce3a79ae73ad15e91"
+   }
+  }
  },
  "nbformat": 4,
  "nbformat_minor": 2

docs/experimental/Binary Quantization from Scratch.ipynb

@@ -3,22 +3,7 @@
   {
    "cell_type": "markdown",
    "metadata": {},
-   "source": [
-    "# Binary Quantization of OpenAI Embedding\n",
-    "---\n",
-    "\n",
-    "In the world of large-scale data retrieval and processing, efficiency is crucial. With the exponential growth of data, the ability to retrieve information quickly and accurately can significantly affect system performance. This blog post explores a technique known as binary quantization applied to OpenAI embeddings, demonstrating how it can enhance **retrieval latency by 20x** or more.\n",
-    "\n",
-    "## What Are OpenAI Embeddings?\n",
-    "OpenAI embeddings are numerical representations of textual information. They transform text into a vector space where semantically similar texts are mapped close together. This mathematical representation enables computers to understand and process human language more effectively.\n",
-    "\n",
-    "## Binary Quantization\n",
-    "Binary quantization is a method which converts continuous numerical values into binary values (0 or 1). It simplifies the data structure, allowing faster computations. Here's a brief overview of the binary quantization process applied to OpenAI embeddings:\n",
-    "\n",
-    "1. **Load Embeddings**: OpenAI embeddings are loaded from parquet files.\n",
-    "2. **Binary Transformation**: The continuous valued vectors are converted into binary form. Here, values greater than 0 are set to 1, and others remain 0.\n",
-    "3. **Comparison & Retrieval**: Binary vectors are used for comparison using logical XOR operations and other efficient algorithms."
-   ]
+   "source": []
   },
   {
    "cell_type": "markdown",

docs/index.md

@@ -6,7 +6,7 @@ The default embedding supports "query" and "passage" prefixes for the input text
 
 1. Light & Fast
     - Quantized model weights
-    - ONNX Runtime for inference via [Optimum](github.com/huggingface/optimum)
+    - ONNX Runtime for inference via [Optimum](https://github.com/huggingface/optimum)
 
 2. Accuracy/Recall
     - Better than OpenAI Ada-002

docs/qdrant/Binary_Quantization_with_Qdrant.ipynb

@@ -0,0 +1,377 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Binary Quantization with Qdrant & OpenAI Embedding\n",
+    "\n",
+    "---\n",
+    "In the world of large-scale data retrieval and processing, efficiency is crucial. With the exponential growth of data, the ability to retrieve information quickly and accurately can significantly affect system performance. This blog post explores a technique known as binary quantization applied to OpenAI embeddings, demonstrating how it can enhance **retrieval latency by 20x** or more.\n",
+    "\n",
+    "## What Are OpenAI Embeddings?\n",
+    "OpenAI embeddings are numerical representations of textual information. They transform text into a vector space where semantically similar texts are mapped close together. This mathematical representation enables computers to understand and process human language more effectively.\n",
+    "\n",
+    "## Binary Quantization\n",
+    "Binary quantization is a method which converts continuous numerical values into binary values (0 or 1). It simplifies the data structure, allowing faster computations. Here's a brief overview of the binary quantization process applied to OpenAI embeddings:\n",
+    "\n",
+    "1. **Load Embeddings**: OpenAI embeddings are loaded from parquet files.\n",
+    "2. **Binary Transformation**: The continuous valued vectors are converted into binary form. Here, values greater than 0 are set to 1, and others remain 0.\n",
+    "3. **Comparison & Retrieval**: Binary vectors are used for comparison using logical XOR operations and other efficient algorithms.\n",
+    "\n",
+    "Binary Quantization is a promising approach to improve retrieval speeds and reduce memory footprint of vector search engines. In this notebook we will show how to use Qdrant to perform binary quantization of vectors and perform fast similarity search on the resulting index.\n",
+    "\n",
+    "## Table of Contents\n",
+    "1. Imports\n",
+    "2. Download and Slice Dataset\n",
+    "3. Create Qdrant Collection\n",
+    "4. Indexing\n",
+    "5. Search\n",
+    "\n",
+    "## 1. Imports"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:33:03.013948Z",
+     "start_time": "2024-04-01T16:33:01.019043Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "!pip install qdrant-client pandas dataset --quiet --upgrade"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:33:03.914729Z",
+     "start_time": "2024-04-01T16:33:03.015394Z"
+    }
+   },
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/Users/joein/work/qdrant/fastembed/venv/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
+      "  from .autonotebook import tqdm as notebook_tqdm\n"
+     ]
+    }
+   ],
+   "source": [
+    "import os\n",
+    "import random\n",
+    "import time\n",
+    "\n",
+    "import numpy as np\n",
+    "import pandas as pd\n",
+    "from qdrant_client import QdrantClient, models\n",
+    "\n",
+    "random.seed(37)\n",
+    "np.random.seed(37)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## 2. Download and Slice Dataset\n",
+    "\n",
+    "We will be using the [dbpedia-entities](https://huggingface.co/datasets/Qdrant/dbpedia-entities-openai3-text-embedding-3-small-1536-100K) dataset from the [HuggingFace Datasets](https://huggingface.co/datasets) library. This contains 100K vectors of 1536 dimensions each"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:33:09.085853Z",
+     "start_time": "2024-04-01T16:33:03.912688Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "100000"
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import datasets\n",
+    "\n",
+    "dataset = datasets.load_dataset(\n",
+    "    \"Qdrant/dbpedia-entities-openai3-text-embedding-3-small-1536-100K\", split=\"train\"\n",
+    ")\n",
+    "len(dataset)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:33:09.176212Z",
+     "start_time": "2024-04-01T16:33:09.084550Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "True"
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "client = QdrantClient(\n",
+    "    prefer_grpc=True,\n",
+    ")\n",
+    "\n",
+    "collection_name = \"binary-quantization\"\n",
+    "client.recreate_collection(\n",
+    "    collection_name=collection_name,\n",
+    "    vectors_config=models.VectorParams(\n",
+    "        size=1536,\n",
+    "        distance=models.Distance.DOT,\n",
+    "        on_disk=True,\n",
+    "    ),\n",
+    "    quantization_config=models.BinaryQuantization(\n",
+    "        binary=models.BinaryQuantizationConfig(always_ram=True),\n",
+    "    ),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:34:13.383986Z",
+     "start_time": "2024-04-01T16:33:09.175725Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "def iter_dataset(dataset):\n",
+    "    for point in dataset:\n",
+    "        yield point[\"openai\"], {\"text\": point[\"text\"]}\n",
+    "\n",
+    "\n",
+    "vectors, payload = zip(*iter_dataset(dataset))\n",
+    "client.upload_collection(\n",
+    "    collection_name=collection_name,\n",
+    "    vectors=vectors,\n",
+    "    payload=payload,\n",
+    "    parallel=max(1, (os.cpu_count() // 2)),\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:34:13.390886Z",
+     "start_time": "2024-04-01T16:34:13.385961Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "{'status': <CollectionStatus.YELLOW: 'yellow'>,\n 'optimizer_status': <OptimizersStatusOneOf.OK: 'ok'>,\n 'vectors_count': 116640,\n 'indexed_vectors_count': 43520,\n 'points_count': 116640,\n 'segments_count': 6,\n 'config': {'params': {'vectors': {'size': 1536,\n    'distance': <Distance.DOT: 'Dot'>,\n    'hnsw_config': None,\n    'quantization_config': None,\n    'on_disk': True},\n   'shard_number': 1,\n   'sharding_method': None,\n   'replication_factor': 1,\n   'write_consistency_factor': 1,\n   'read_fan_out_factor': None,\n   'on_disk_payload': True,\n   'sparse_vectors': None},\n  'hnsw_config': {'m': 16,\n   'ef_construct': 100,\n   'full_scan_threshold': 10000,\n   'max_indexing_threads': 0,\n   'on_disk': False,\n   'payload_m': None},\n  'optimizer_config': {'deleted_threshold': 0.2,\n   'vacuum_min_vector_number': 1000,\n   'default_segment_number': 0,\n   'max_segment_size': None,\n   'memmap_threshold': None,\n   'indexing_threshold': 20000,\n   'flush_interval_sec': 5,\n   'max_optimization_threads': None},\n  'wal_config': {'wal_capacity_mb': 32, 'wal_segments_ahead': 0},\n  'quantization_config': {'binary': {'always_ram': True}}},\n 'payload_schema': {}}"
+     },
+     "execution_count": 6,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "collection_info = client.get_collection(collection_name=f\"{collection_name}\")\n",
+    "collection_info.dict()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Oversampling vs Recall\n",
+    "\n",
+    "### Preparing a query dataset\n",
+    "\n",
+    "For the purpose of this illustration, we'll take a few vectors which we know are already in the index and query them. We should get the same vectors back as results from the Qdrant index. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:34:13.453626Z",
+     "start_time": "2024-04-01T16:34:13.391567Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": "[89391,\n 79659,\n 12006,\n 80978,\n 87219,\n 97885,\n 83155,\n 67504,\n 4645,\n 82711,\n 48395,\n 57375,\n 69208,\n 14136,\n 89515,\n 59880,\n 78730,\n 36952,\n 49620,\n 96486,\n 55473,\n 58179,\n 18926,\n 6489,\n 11931,\n 54146,\n 9850,\n 71259,\n 37825,\n 47331,\n 84964,\n 92399,\n 56669,\n 77042,\n 73744,\n 47993,\n 83780,\n 92429,\n 75114,\n 4463,\n 69030,\n 81185,\n 27950,\n 66217,\n 54652,\n 8260,\n 1151,\n 993,\n 85954,\n 66863,\n 47303,\n 8992,\n 92688,\n 76030,\n 29472,\n 3077,\n 42454,\n 46120,\n 69140,\n 20877,\n 2844,\n 95423,\n 1770,\n 28568,\n 96448,\n 94227,\n 40837,\n 91684,\n 29785,\n 66936,\n 85121,\n 39546,\n 81910,\n 5514,\n 37068,\n 35731,\n 93990,\n 26685,\n 63076,\n 18762,\n 27922,\n 34916,\n 80976,\n 83189,\n 6328,\n 57508,\n 58860,\n 13758,\n 72976,\n 85030,\n 332,\n 34963,\n 85009,\n 31344,\n 11560,\n 58108,\n 85163,\n 17064,\n 44712,\n 45962]"
+     },
+     "execution_count": 7,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "query_indices = random.sample(range(len(dataset)), 100)\n",
+    "query_dataset = dataset[query_indices]\n",
+    "query_indices"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:34:13.453928Z",
+     "start_time": "2024-04-01T16:34:13.452405Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "## Add Gaussian noise to any vector\n",
+    "\n",
+    "\n",
+    "def add_noise(vector, noise=0.05):\n",
+    "    return vector + noise * np.random.randn(*vector.shape)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:34:13.457839Z",
+     "start_time": "2024-04-01T16:34:13.455431Z"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "def correct(results, text):\n",
+    "    return text in [x.payload[\"text\"] for x in results]\n",
+    "\n",
+    "\n",
+    "def count_correct(query_dataset, limit=1, oversampling=1, rescore=False):\n",
+    "    correct_results = 0\n",
+    "    for query_vector, text in zip(query_dataset[\"openai\"], query_dataset[\"text\"]):\n",
+    "        results = client.search(\n",
+    "            collection_name=collection_name,\n",
+    "            query_vector=add_noise(np.array(query_vector)),\n",
+    "            limit=limit,\n",
+    "            search_params=models.SearchParams(\n",
+    "                quantization=models.QuantizationSearchParams(\n",
+    "                    rescore=rescore,\n",
+    "                    oversampling=oversampling,\n",
+    "                )\n",
+    "            ),\n",
+    "        )\n",
+    "        correct_results += correct(results, text)\n",
+    "    return correct_results"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:40:48.347002Z",
+     "start_time": "2024-04-01T16:40:42.228551Z"
+    },
+    "collapsed": false
+   },
+   "outputs": [],
+   "source": [
+    "limit_grid = [1, 3, 10, 20, 50]\n",
+    "oversampling_grid = [1.0, 3.0, 5.0]\n",
+    "rescore_grid = [False, True]\n",
+    "results = []\n",
+    "\n",
+    "for limit in limit_grid:\n",
+    "    for oversampling in oversampling_grid:\n",
+    "        for rescore in rescore_grid:\n",
+    "            start = time.perf_counter()\n",
+    "            correct_results = count_correct(\n",
+    "                query_dataset, limit=limit, oversampling=oversampling, rescore=rescore\n",
+    "            )\n",
+    "            end = time.perf_counter()\n",
+    "            results.append(\n",
+    "                {\n",
+    "                    \"limit\": limit,\n",
+    "                    \"oversampling\": oversampling,\n",
+    "                    \"candidates\": int(oversampling * limit),\n",
+    "                    \"rescore\": rescore,\n",
+    "                    \"accuracy\": correct_results / 100,\n",
+    "                    \"total queries\": len(query_dataset[\"text\"]),\n",
+    "                    \"time\": end - start,\n",
+    "                }\n",
+    "            )"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 22,
+   "metadata": {
+    "ExecuteTime": {
+     "end_time": "2024-04-01T16:41:55.445405Z",
+     "start_time": "2024-04-01T16:41:55.442687Z"
+    }
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/html": "<div>\n<style scoped>\n    .dataframe tbody tr th:only-of-type {\n        vertical-align: middle;\n    }\n\n    .dataframe tbody tr th {\n        vertical-align: top;\n    }\n\n    .dataframe thead th {\n        text-align: right;\n    }\n</style>\n<table border=\"1\" class=\"dataframe\">\n  <thead>\n    <tr style=\"text-align: right;\">\n      <th></th>\n      <th>candidates</th>\n      <th>rescore</th>\n      <th>accuracy</th>\n      <th>time</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>1</td>\n      <td>False</td>\n      <td>0.90</td>\n      <td>0.221826</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>1</td>\n      <td>True</td>\n      <td>0.91</td>\n      <td>0.134167</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>3</td>\n      <td>False</td>\n      <td>0.88</td>\n      <td>0.115299</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>3</td>\n      <td>True</td>\n      <td>0.97</td>\n      <td>0.209320</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>5</td>\n      <td>False</td>\n      <td>0.84</td>\n      <td>0.154485</td>\n    </tr>\n    <tr>\n      <th>5</th>\n      <td>5</td>\n      <td>True</td>\n      <td>0.91</td>\n      <td>0.124424</td>\n    </tr>\n    <tr>\n      <th>6</th>\n      <td>3</td>\n      <td>False</td>\n      <td>0.99</td>\n      <td>0.121695</td>\n    </tr>\n    <tr>\n      <th>7</th>\n      <td>3</td>\n      <td>True</td>\n      <td>0.96</td>\n      <td>0.123257</td>\n    </tr>\n    <tr>\n      <th>8</th>\n      <td>9</td>\n      <td>False</td>\n      <td>0.94</td>\n      <td>0.119629</td>\n    </tr>\n    <tr>\n      <th>9</th>\n      <td>9</td>\n      <td>True</td>\n      <td>0.98</td>\n      <td>0.119372</td>\n    </tr>\n    <tr>\n      <th>10</th>\n      <td>15</td>\n      <td>False</td>\n      <td>0.90</td>\n      <td>0.121621</td>\n    </tr>\n    <tr>\n      <th>11</th>\n      <td>15</td>\n      <td>True</td>\n      <td>0.97</td>\n      <td>0.125466</td>\n    </tr>\n    <tr>\n      <th>12</th>\n      <td>10</td>\n      <td>False</td>\n      <td>0.93</td>\n      <td>0.135910</td>\n    </tr>\n    <tr>\n      <th>13</th>\n      <td>10</td>\n      <td>True</td>\n      <td>0.95</td>\n      <td>0.138135</td>\n    </tr>\n    <tr>\n      <th>14</th>\n      <td>30</td>\n      <td>False</td>\n      <td>0.94</td>\n      <td>0.177928</td>\n    </tr>\n    <tr>\n      <th>15</th>\n      <td>30</td>\n      <td>True</td>\n      <td>0.98</td>\n      <td>0.254588</td>\n    </tr>\n    <tr>\n      <th>16</th>\n      <td>50</td>\n      <td>False</td>\n      <td>0.94</td>\n      <td>0.268659</td>\n    </tr>\n    <tr>\n      <th>17</th>\n      <td>50</td>\n      <td>True</td>\n      <td>0.96</td>\n      <td>0.269792</td>\n    </tr>\n    <tr>\n      <th>18</th>\n      <td>20</td>\n      <td>False</td>\n      <td>0.96</td>\n      <td>0.249941</td>\n    </tr>\n    <tr>\n      <th>19</th>\n      <td>20</td>\n      <td>True</td>\n      <td>0.96</td>\n      <td>0.247138</td>\n    </tr>\n    <tr>\n      <th>20</th>\n      <td>60</td>\n      <td>False</td>\n      <td>0.97</td>\n      <td>0.251301</td>\n    </tr>\n    <tr>\n      <th>21</th>\n      <td>60</td>\n      <td>True</td>\n      <td>0.98</td>\n      <td>0.256504</td>\n    </tr>\n    <tr>\n      <th>22</th>\n      <td>100</td>\n      <td>False</td>\n      <td>0.98</td>\n      <td>0.270049</td>\n    </tr>\n    <tr>\n      <th>23</th>\n      <td>100</td>\n      <td>True</td>\n      <td>0.97</td>\n      <td>0.248972</td>\n    </tr>\n    <tr>\n      <th>24</th>\n      <td>50</td>\n      <td>False</td>\n      <td>0.97</td>\n      <td>0.306356</td>\n    </tr>\n    <tr>\n      <th>25</th>\n      <td>50</td>\n      <td>True</td>\n      <td>0.98</td>\n      <td>0.257544</td>\n    </tr>\n    <tr>\n      <th>26</th>\n      <td>150</td>\n      <td>False</td>\n      <td>0.98</td>\n      <td>0.238811</td>\n    </tr>\n    <tr>\n      <th>27</th>\n      <td>150</td>\n      <td>True</td>\n      <td>0.99</td>\n      <td>0.263939</td>\n    </tr>\n    <tr>\n      <th>28</th>\n      <td>250</td>\n      <td>False</td>\n      <td>0.99</td>\n      <td>0.256558</td>\n    </tr>\n    <tr>\n      <th>29</th>\n      <td>250</td>\n      <td>True</td>\n      <td>1.00</td>\n      <td>0.335823</td>\n    </tr>\n  </tbody>\n</table>\n</div>",
+      "text/plain": "    candidates  rescore  accuracy      time\n0            1    False      0.90  0.221826\n1            1     True      0.91  0.134167\n2            3    False      0.88  0.115299\n3            3     True      0.97  0.209320\n4            5    False      0.84  0.154485\n5            5     True      0.91  0.124424\n6            3    False      0.99  0.121695\n7            3     True      0.96  0.123257\n8            9    False      0.94  0.119629\n9            9     True      0.98  0.119372\n10          15    False      0.90  0.121621\n11          15     True      0.97  0.125466\n12          10    False      0.93  0.135910\n13          10     True      0.95  0.138135\n14          30    False      0.94  0.177928\n15          30     True      0.98  0.254588\n16          50    False      0.94  0.268659\n17          50     True      0.96  0.269792\n18          20    False      0.96  0.249941\n19          20     True      0.96  0.247138\n20          60    False      0.97  0.251301\n21          60     True      0.98  0.256504\n22         100    False      0.98  0.270049\n23         100     True      0.97  0.248972\n24          50    False      0.97  0.306356\n25          50     True      0.98  0.257544\n26         150    False      0.98  0.238811\n27         150     True      0.99  0.263939\n28         250    False      0.99  0.256558\n29         250     True      1.00  0.335823"
+     },
+     "execution_count": 22,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "df = pd.DataFrame(results)\n",
+    "df[[\"candidates\", \"rescore\", \"accuracy\", \"time\"]]\n",
+    "# df.to_csv(\"candidates-rescore-time.csv\", index=False)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 4
+}

docs/qdrant/Retrieval_with_FastEmbed.ipynb

@@ -39,15 +39,7 @@
    "cell_type": "code",
    "execution_count": 2,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "\u001b[32m2024-02-07 22:20:57.013\u001b[0m | \u001b[33m\u001b[1mWARNING \u001b[0m | \u001b[36mfastembed.embedding\u001b[0m:\u001b[36m<module>\u001b[0m:\u001b[36m7\u001b[0m - \u001b[33m\u001b[1mDefaultEmbedding, FlagEmbedding, JinaEmbedding are deprecated. Use TextEmbedding instead.\u001b[0m\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "from typing import List\n",
     "import numpy as np\n",
@@ -92,7 +84,7 @@
     "    \"His life has been depicted in various films, TV shows, and books\",\n",
     "]\n",
     "# Initialize the DefaultEmbedding class with the desired parameters\n",
-    "embedding_model = TextEmbedding(model_name=\"BAAI/bge-small-en\", max_length=512)\n",
+    "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",
@@ -140,7 +132,7 @@
     {
      "data": {
       "text/plain": [
-       "(array([-0.04393955,  0.04452892, -0.00760788, -0.03399807,  0.01951348],\n",
+       "(array([-0.06002192,  0.04322132, -0.00545516, -0.04419701, -0.00542277],\n",
        "       dtype=float32),\n",
        " array([-0.06002192,  0.04322132, -0.00545516, -0.04419701, -0.00542277],\n",
        "       dtype=float32))"
@@ -183,7 +175,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.11.5"
+   "version": "3.10.13"
   },
   "orig_nbformat": 4
  },

experiments/02_SPLADE_to_ONNX.ipynb

@@ -0,0 +1,371 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "import torch\n",
+    "from transformers import AutoModelForMaskedLM, AutoTokenizer"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Running the model with Transformers and Torch"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "sentences = [\n",
+    "    \"Hello World\",\n",
+    "    \"Built by Nirant Kasliwal\",\n",
+    "]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## PyTorch Code from the [SPLADERunner](https://github.com/PrithivirajDamodaran/SPLADERunner) library"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "hf_token = \"<your_hf_token_here>\""
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Output Logits shape:  torch.Size([2, 10, 30522])\n",
+      "Output Attention mask shape:  torch.Size([2, 10])\n",
+      "Sparse Vector shape:  torch.Size([2, 30522])\n",
+      "SPLADE BOW rep for sentence:\tBuilt by Nirant Kasliwal\n",
+      "[('##rant', 2.02), ('built', 1.94), ('##wal', 1.79), ('##sl', 1.69), ('build', 1.57), ('ka', 1.4), ('ni', 1.26), ('made', 0.93), ('architect', 0.76), ('was', 0.69), ('who', 0.61), ('his', 0.5), ('wrote', 0.47), ('india', 0.45), ('company', 0.41), ('##i', 0.41), ('he', 0.37), ('manufacturer', 0.36), ('by', 0.35), ('engineer', 0.33), ('architecture', 0.33), ('ko', 0.23), ('him', 0.22), ('invented', 0.19), ('said', 0.14), ('k', 0.11), ('man', 0.11), ('statue', 0.11), ('bomb', 0.1), ('##wa', 0.1), ('builder', 0.09), ('.', 0.07), ('started', 0.06), (',', 0.04), ('ku', 0.03)]\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Download the model and tokenizer\n",
+    "device = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\"prithivida/Splade_PP_en_v1\", token=hf_token)\n",
+    "reverse_voc = {v: k for k, v in tokenizer.vocab.items()}\n",
+    "model = AutoModelForMaskedLM.from_pretrained(\"prithivida/Splade_PP_en_v1\", token=hf_token)\n",
+    "model.to(device)\n",
+    "\n",
+    "# Tokenize the input\n",
+    "inputs = tokenizer(sentences, return_tensors=\"pt\", padding=True, truncation=True, max_length=512)\n",
+    "inputs = {key: val.to(device) for key, val in inputs.items()}\n",
+    "input_ids = inputs[\"input_ids\"]\n",
+    "attention_mask = inputs[\"attention_mask\"]\n",
+    "token_type_ids = inputs[\"token_type_ids\"]\n",
+    "\n",
+    "# Run model and prepare sparse vector\n",
+    "outputs = model(**inputs)\n",
+    "logits = outputs.logits\n",
+    "print(\"Output Logits shape: \", logits.shape)\n",
+    "print(\"Output Attention mask shape: \", attention_mask.shape)\n",
+    "relu_log = torch.log(1 + torch.relu(logits))\n",
+    "weighted_log = relu_log * attention_mask.unsqueeze(-1)\n",
+    "max_val, _ = torch.max(weighted_log, dim=1)\n",
+    "vector = max_val.squeeze()\n",
+    "print(\"Sparse Vector shape: \", vector.shape)\n",
+    "# print(\"Number of Actual Dimensions: \", len(cols))\n",
+    "cols = [vec.nonzero().squeeze().cpu().tolist() for vec in vector]\n",
+    "weights = [vec[col].cpu().tolist() for vec, col in zip(vector, cols)]\n",
+    "\n",
+    "idx = 1\n",
+    "cols, weights = cols[idx], weights[idx]\n",
+    "# Print the BOW representation\n",
+    "d = {k: v for k, v in zip(cols, weights)}\n",
+    "sorted_d = {k: v for k, v in sorted(d.items(), key=lambda item: item[1], reverse=True)}\n",
+    "bow_rep = []\n",
+    "for k, v in sorted_d.items():\n",
+    "    bow_rep.append((reverse_voc[k], round(v, 2)))\n",
+    "print(f\"SPLADE BOW rep for sentence:\\t{sentences[idx]}\\n{bow_rep}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Export with output_attentions and logits"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Exporting model to models/nirantk_SPLADE_PP_en_v1\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "('models/nirantk_SPLADE_PP_en_v1/tokenizer_config.json',\n",
+       " 'models/nirantk_SPLADE_PP_en_v1/special_tokens_map.json',\n",
+       " 'models/nirantk_SPLADE_PP_en_v1/vocab.txt',\n",
+       " 'models/nirantk_SPLADE_PP_en_v1/added_tokens.json',\n",
+       " 'models/nirantk_SPLADE_PP_en_v1/tokenizer.json')"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from transformers import AutoTokenizer\n",
+    "\n",
+    "model_id = \"nirantk/SPLADE_PP_en_v1\"\n",
+    "output_dir = f\"models/{model_id.replace('/', '_')}\"\n",
+    "model_kwargs = {\"output_attentions\": True, \"return_dict\": True}\n",
+    "\n",
+    "print(f\"Exporting model to {output_dir}\")\n",
+    "tokenizer.save_pretrained(output_dir)\n",
+    "# main_export(\n",
+    "#     model_id,\n",
+    "#     output=output_dir,\n",
+    "#     no_post_process=True,\n",
+    "#     model_kwargs=model_kwargs,\n",
+    "#     token=hf_token,\n",
+    "# )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Running the model with ONNX"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from optimum.onnxruntime import ORTModelForMaskedLM\n",
+    "\n",
+    "model = ORTModelForMaskedLM.from_pretrained(\"nirantk/SPLADE_PP_en_v1\")\n",
+    "tokenizer = AutoTokenizer.from_pretrained(\"nirantk/SPLADE_PP_en_v1\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "inputs = tokenizer(sentences, return_tensors=\"pt\", padding=True, truncation=True, max_length=512)\n",
+    "inputs = {key: val.to(device) for key, val in inputs.items()}\n",
+    "input_ids = inputs[\"input_ids\"]\n",
+    "attention_mask = inputs[\"attention_mask\"]\n",
+    "token_type_ids = inputs[\"token_type_ids\"]\n",
+    "\n",
+    "onnx_input = {\n",
+    "    \"input_ids\": input_ids.cpu().numpy(),\n",
+    "    \"attention_mask\": attention_mask.cpu().numpy(),\n",
+    "    \"token_type_ids\": token_type_ids.cpu().numpy(),\n",
+    "}\n",
+    "\n",
+    "logits = model(**onnx_input).logits"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "(2, 10, 30522)"
+      ]
+     },
+     "execution_count": 8,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "logits.shape"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Output Logits shape:  (2, 10, 30522)\n",
+      "Sparse Vector shape:  (2, 30522)\n",
+      "SPLADE BOW rep for sentence:\tBuilt by Nirant Kasliwal\n",
+      "[('##rant', 2.02), ('built', 1.94), ('##wal', 1.79), ('##sl', 1.69), ('build', 1.57), ('ka', 1.4), ('ni', 1.26), ('made', 0.93), ('architect', 0.76), ('was', 0.69), ('who', 0.61), ('his', 0.5), ('wrote', 0.47), ('india', 0.45), ('company', 0.41), ('##i', 0.41), ('he', 0.37), ('manufacturer', 0.36), ('by', 0.35), ('engineer', 0.33), ('architecture', 0.33), ('ko', 0.23), ('him', 0.22), ('invented', 0.19), ('said', 0.14), ('k', 0.11), ('man', 0.11), ('statue', 0.11), ('bomb', 0.1), ('##wa', 0.1), ('builder', 0.09), ('.', 0.07), ('started', 0.06), (',', 0.04), ('ku', 0.03)]\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(\"Output Logits shape: \", logits.shape)\n",
+    "\n",
+    "relu_log = np.log(1 + np.maximum(logits, 0))\n",
+    "\n",
+    "# Equivalent to relu_log * attention_mask.unsqueeze(-1)\n",
+    "# For NumPy, you might need to explicitly expand dimensions if 'attention_mask' is not already 2D\n",
+    "weighted_log = relu_log * np.expand_dims(attention_mask, axis=-1)\n",
+    "\n",
+    "# Equivalent to torch.max(weighted_log, dim=1)\n",
+    "# NumPy's max function returns only the max values, not the indices, so we don't need to unpack two values\n",
+    "max_val = np.max(weighted_log, axis=1)\n",
+    "\n",
+    "# Equivalent to max_val.squeeze()\n",
+    "# This step may be unnecessary in NumPy if max_val doesn't have unnecessary dimensions\n",
+    "vector = np.squeeze(max_val)\n",
+    "print(\"Sparse Vector shape: \", vector.shape)\n",
+    "\n",
+    "# print(vector[0].nonzero())\n",
+    "\n",
+    "cols = [vec.nonzero()[0].squeeze().tolist() for vec in vector]\n",
+    "weights = [vec[col].tolist() for vec, col in zip(vector, cols)]\n",
+    "\n",
+    "idx = 1\n",
+    "cols, weights = cols[idx], weights[idx]\n",
+    "# Print the BOW representation\n",
+    "d = {k: v for k, v in zip(cols, weights)}\n",
+    "sorted_d = {k: v for k, v in sorted(d.items(), key=lambda item: item[1], reverse=True)}\n",
+    "bow_rep = []\n",
+    "for k, v in sorted_d.items():\n",
+    "    bow_rep.append((reverse_voc[k], round(v, 2)))\n",
+    "print(f\"SPLADE BOW rep for sentence:\\t{sentences[idx]}\\n{bow_rep}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "35"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "len(cols)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[1010,\n",
+       " 1012,\n",
+       " 1047,\n",
+       " 2001,\n",
+       " 2002,\n",
+       " 2010,\n",
+       " 2011,\n",
+       " 2032,\n",
+       " 2040,\n",
+       " 2056,\n",
+       " 2072,\n",
+       " 2081,\n",
+       " 2158,\n",
+       " 2194,\n",
+       " 2318,\n",
+       " 2328,\n",
+       " 2626,\n",
+       " 2634,\n",
+       " 3857,\n",
+       " 3992,\n",
+       " 4213,\n",
+       " 4294,\n",
+       " 4944,\n",
+       " 5968,\n",
+       " 6231,\n",
+       " 7751,\n",
+       " 8826,\n",
+       " 9152,\n",
+       " 10556,\n",
+       " 12508,\n",
+       " 12849,\n",
+       " 13476,\n",
+       " 13970,\n",
+       " 14540,\n",
+       " 17884]"
+      ]
+     },
+     "execution_count": 11,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "cols"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "fst",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

experiments/attention_export.py

@@ -0,0 +1,15 @@
+from optimum.exporters.onnx import main_export
+from transformers import AutoTokenizer
+
+model_id = "sentence-transformers/paraphrase-MiniLM-L6-v2"
+output_dir = f"models/{model_id.replace('/', '_')}"
+model_kwargs = {"output_attentions": True, "return_dict": True}
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+# export if the output model does not exist
+# try:
+#     sess = onnxruntime.InferenceSession(f"{output_dir}/model.onnx")
+#     print("Model already exported")
+# except FileNotFoundError:
+print(f"Exporting model to {output_dir}")
+main_export(model_id, output=output_dir, no_post_process=True, model_kwargs=model_kwargs)

experiments/dynamic_quantization_export.py

@@ -1,473 +0,0 @@
-"""
-This script is used to convert HuggingFace models to ONNX format and optionally quantize the model using dynamic quantization. 
-This is courtesy of Joshua aka @Xenova
-"""
-
-import json
-import os
-import shutil
-from dataclasses import dataclass, field
-from typing import Optional, Set
-
-import onnx
-from onnxruntime.quantization import QuantType, quantize_dynamic
-from optimum.exporters.onnx import export_models, main_export
-from optimum.exporters.tasks import TasksManager
-from tqdm import tqdm
-from transformers import AutoConfig, AutoTokenizer, HfArgumentParser
-
-DEFAULT_QUANTIZE_PARAMS = {
-    "per_channel": True,
-    "reduce_range": True,
-}
-
-MODEL_SPECIFIC_QUANTIZE_PARAMS = {
-    # Decoder-only models
-    "codegen": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "gpt2": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "gpt_bigcode": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "gptj": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "gpt-neo": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "gpt-neox": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "mpt": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "bloom": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "llama": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "opt": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "mistral": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "falcon": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "phi": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "qwen2": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    # Encoder-decoder models
-    "whisper": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    "vision-encoder-decoder": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-    # Encoder-only models
-    "owlv2": {
-        "per_channel": False,
-        "reduce_range": False,
-    },
-}
-
-MODELS_WITHOUT_TOKENIZERS = [
-    "wav2vec2",
-    "wav2vec2-bert",
-    "wavlm",
-    "hubert",
-]
-
-
-@dataclass
-class ConversionArguments:
-    """
-    Arguments used for converting HuggingFace models to onnx.
-    """
-
-    model_id: str = field(metadata={"help": "Model identifier"})
-    tokenizer_id: str = field(default=None, metadata={"help": "Tokenizer identifier (if different to `model_id`)"})
-    quantize: bool = field(default=False, metadata={"help": "Whether to quantize the model."})
-    output_parent_dir: str = field(
-        default="./models/", metadata={"help": "Path where the converted model will be saved to."}
-    )
-
-    task: Optional[str] = field(
-        default="auto",
-        metadata={
-            "help": (
-                "The task to export the model for. If not specified, the task will be auto-inferred based on the model. Available tasks depend on the model, but are among:"
-                f" {str(TasksManager.get_all_tasks())}. For decoder models, use `xxx-with-past` to export the model using past key values in the decoder."
-            )
-        },
-    )
-
-    opset: int = field(
-        default=None,
-        metadata={
-            "help": (
-                "If specified, ONNX opset version to export the model with. Otherwise, the default opset will be used."
-            )
-        },
-    )
-
-    device: str = field(default="cpu", metadata={"help": "The device to use to do the export."})
-    skip_validation: bool = field(default=False, metadata={"help": "Whether to skip validation of the converted model"})
-
-    per_channel: bool = field(default=None, metadata={"help": "Whether to quantize weights per channel"})
-    reduce_range: bool = field(
-        default=None,
-        metadata={
-            "help": "Whether to quantize weights with 7-bits. It may improve the accuracy for some models running on non-VNNI machine, especially for per-channel mode"
-        },
-    )
-
-    output_attentions: bool = field(
-        default=False,
-        metadata={
-            "help": "Whether to output attentions from the model. NOTE: This is only supported for whisper models right now."
-        },
-    )
-
-    split_modalities: bool = field(
-        default=False,
-        metadata={
-            "help": "Whether to split multimodal models. NOTE: This is only supported for CLIP models right now."
-        },
-    )
-
-    trust_remote_code: bool = field(
-        default=False,
-        metadata={
-            "help": "Allows to use custom code for the modeling hosted in the model repository. This option should only be set for repositories"
-            "you trust and in which you have read the code, as it will execute on your local machine arbitrary code present in the model repository."
-        },
-    )
-
-    custom_onnx_configs: str = field(
-        default=None,
-        metadata={
-            "help": "Experimental usage: override the default ONNX config used for the given model. This argument may be useful for advanced users "
-            "that desire a finer-grained control on the export."
-        },
-    )
-
-
-def get_operators(model: onnx.ModelProto) -> Set[str]:
-    operators = set()
-
-    def traverse_graph(graph):
-        for node in graph.node:
-            operators.add(node.op_type)
-            for attr in node.attribute:
-                if attr.type == onnx.AttributeProto.GRAPH:
-                    subgraph = attr.g
-                    traverse_graph(subgraph)
-
-    traverse_graph(model.graph)
-    return operators
-
-
-def quantize(model_names_or_paths, **quantize_kwargs):
-    """
-    Quantize the weights of the model from float32 to int8 to allow very efficient inference on modern CPU
-
-    Uses unsigned ints for activation values, signed ints for weights, per
-    https://onnxruntime.ai/docs/performance/quantization.html#data-type-selection
-    it is faster on most CPU architectures
-    Args:
-        onnx_model_path: Path to location the exported ONNX model is stored
-    Returns: The Path generated for the quantized
-    """
-
-    quantize_config = dict(**quantize_kwargs, per_model_config={})
-
-    for model in tqdm(model_names_or_paths, desc="Quantizing"):
-        directory_path = os.path.dirname(model)
-        file_name_without_extension = os.path.splitext(os.path.basename(model))[0]
-
-        # NOTE:
-        # As of 2023/04/20, the current latest version of onnxruntime-web is 1.14.0, and does not support INT8 weights for Conv layers.
-        # For this reason, we choose model weight types to ensure compatibility with onnxruntime-web.
-        #
-        # As per docs, signed weight type (QInt8) is faster on most CPUs, so, we use that unless the model contains a Conv layer.
-        # For more information, see:
-        #  - https://github.com/microsoft/onnxruntime/issues/3130#issuecomment-1105200621
-        #  - https://github.com/microsoft/onnxruntime/issues/2339
-
-        loaded_model = onnx.load_model(model)
-        op_types = get_operators(loaded_model)
-        weight_type = QuantType.QUInt8 if "Conv" in op_types else QuantType.QInt8
-
-        quantize_dynamic(
-            model_input=model,
-            model_output=os.path.join(directory_path, f"{file_name_without_extension}_quantized.onnx"),
-            weight_type=weight_type,
-            # TODO allow user to specify these
-            # op_types_to_quantize=['MatMul', 'Add', 'Conv'],
-            extra_options=dict(EnableSubgraph=True),
-            **quantize_kwargs,
-        )
-
-        quantize_config["per_model_config"][file_name_without_extension] = dict(
-            op_types=list(op_types),
-            weight_type=str(weight_type),
-        )
-
-    # Save quantization config
-    with open(os.path.join(directory_path, "quantize_config.json"), "w") as fp:
-        json.dump(quantize_config, fp, indent=4)
-
-
-def main():
-    parser = HfArgumentParser((ConversionArguments,))
-    (conv_args,) = parser.parse_args_into_dataclasses()
-
-    model_id = conv_args.model_id
-    tokenizer_id = conv_args.tokenizer_id or model_id
-
-    output_model_folder = os.path.join(conv_args.output_parent_dir, model_id)
-
-    # Create output folder
-    os.makedirs(output_model_folder, exist_ok=True)
-
-    from_pretrained_kwargs = dict(
-        trust_remote_code=conv_args.trust_remote_code,
-    )
-
-    # Saving the model config
-    config = AutoConfig.from_pretrained(model_id, **from_pretrained_kwargs)
-
-    custom_kwargs = {}
-    if conv_args.custom_onnx_configs is not None:
-        if conv_args.task == "auto":
-            raise Exception("`--task` must be set when exporting with `--custom_onnx_configs`")
-        custom_onnx_configs = json.loads(conv_args.custom_onnx_configs)
-
-        for key in custom_onnx_configs:
-            onnx_configs = TasksManager._SUPPORTED_MODEL_TYPE[custom_onnx_configs[key]]["onnx"]
-            mapping = onnx_configs[conv_args.task]
-            custom_onnx_configs[key] = mapping.func(config, **mapping.keywords)
-
-        custom_kwargs["custom_onnx_configs"] = custom_onnx_configs
-
-    tokenizer = None
-    try:
-        # Load tokenizer
-        tokenizer = AutoTokenizer.from_pretrained(tokenizer_id, **from_pretrained_kwargs)
-
-        # To avoid inserting all chat templates into tokenizers.js, we save the chat template
-        # to the tokenizer_config.json file, and load it when the tokenizer is loaded.
-        if getattr(tokenizer, "chat_template", None) is None and getattr(tokenizer, "use_default_system_prompt", False):
-            # No chat template specified, and we use the default
-            setattr(tokenizer, "chat_template", tokenizer.default_chat_template)
-
-    except KeyError:
-        pass  # No Tokenizer
-
-    except Exception as e:
-        if config.model_type not in MODELS_WITHOUT_TOKENIZERS:
-            raise e
-
-    core_export_kwargs = dict(
-        opset=conv_args.opset,
-        device=conv_args.device,
-        trust_remote_code=conv_args.trust_remote_code,
-        **custom_kwargs,
-    )
-
-    export_kwargs = dict(
-        model_name_or_path=model_id,
-        output=output_model_folder,
-        task=conv_args.task,
-        do_validation=not conv_args.skip_validation,
-        library_name="transformers",
-        **core_export_kwargs,
-    )
-
-    # Handle special cases
-    if config.model_type == "marian":
-        from .extra.marian import generate_tokenizer_json
-
-        tokenizer_json = generate_tokenizer_json(model_id, tokenizer)
-
-        with open(os.path.join(output_model_folder, "tokenizer.json"), "w", encoding="utf-8") as fp:
-            json.dump(tokenizer_json, fp, indent=4)
-
-    elif config.model_type == "esm":
-        from .extra.esm import generate_fast_tokenizer
-
-        fast_tokenizer = generate_fast_tokenizer(tokenizer)
-        fast_tokenizer.save(os.path.join(output_model_folder, "tokenizer.json"))
-
-    elif config.model_type == "whisper":
-        if conv_args.output_attentions:
-            from .extra.whisper import get_main_export_kwargs
-
-            export_kwargs.update(**get_main_export_kwargs(config, "automatic-speech-recognition"))
-
-    elif config.model_type in ("wav2vec2", "wav2vec2-bert", "hubert"):
-        if tokenizer is not None:
-            from .extra.wav2vec2 import generate_tokenizer_json
-
-            tokenizer_json = generate_tokenizer_json(tokenizer)
-
-            with open(os.path.join(output_model_folder, "tokenizer.json"), "w", encoding="utf-8") as fp:
-                json.dump(tokenizer_json, fp, indent=4)
-
-    elif config.model_type == "vits":
-        if tokenizer is not None:
-            from .extra.vits import generate_tokenizer_json
-
-            tokenizer_json = generate_tokenizer_json(tokenizer)
-
-            with open(os.path.join(output_model_folder, "tokenizer.json"), "w", encoding="utf-8") as fp:
-                json.dump(tokenizer_json, fp, indent=4)
-
-    elif config.model_type == "speecht5":
-        # TODO allow user to specify vocoder path
-        export_kwargs["model_kwargs"] = {"vocoder": "microsoft/speecht5_hifigan"}
-
-        if tokenizer is not None:
-            from .extra.speecht5 import generate_tokenizer_json
-
-            tokenizer_json = generate_tokenizer_json(tokenizer)
-
-            with open(os.path.join(output_model_folder, "tokenizer.json"), "w", encoding="utf-8") as fp:
-                json.dump(tokenizer_json, fp, indent=4)
-
-    elif config.model_type in ("owlvit", "owlv2"):
-        # Override default batch size to 1, needed because non-maximum suppression is performed for exporting.
-        # For more information, see https://github.com/huggingface/optimum/blob/e3b7efb1257c011db907ef40ab340e795cc5684c/optimum/exporters/onnx/model_configs.py#L1028-L1032
-        export_kwargs["batch_size"] = 1
-
-    else:
-        pass  # TODO
-
-    # Step 1. convert huggingface model to onnx
-    if not conv_args.split_modalities:
-        main_export(**export_kwargs)
-    else:
-        custom_export_kwargs = dict(
-            output_dir=output_model_folder,
-            **core_export_kwargs,
-        )
-
-        if config.model_type == "clip":
-            # Handle special case for exporting text and vision models separately
-            from transformers.models.clip import CLIPTextModelWithProjection, CLIPVisionModelWithProjection
-
-            from .extra.clip import CLIPTextModelWithProjectionOnnxConfig, CLIPVisionModelWithProjectionOnnxConfig
-
-            text_model = CLIPTextModelWithProjection.from_pretrained(model_id, **from_pretrained_kwargs)
-            vision_model = CLIPVisionModelWithProjection.from_pretrained(model_id, **from_pretrained_kwargs)
-
-            export_models(
-                models_and_onnx_configs={
-                    "text_model": (text_model, CLIPTextModelWithProjectionOnnxConfig(text_model.config)),
-                    "vision_model": (vision_model, CLIPVisionModelWithProjectionOnnxConfig(vision_model.config)),
-                },
-                **custom_export_kwargs,
-            )
-
-        elif config.model_type == "siglip":
-            # Handle special case for exporting text and vision models separately
-            from transformers.models.siglip import SiglipTextModel, SiglipVisionModel
-
-            from .extra.siglip import SiglipTextModelOnnxConfig, SiglipVisionModelOnnxConfig
-
-            text_model = SiglipTextModel.from_pretrained(model_id, **from_pretrained_kwargs)
-            vision_model = SiglipVisionModel.from_pretrained(model_id, **from_pretrained_kwargs)
-
-            export_models(
-                models_and_onnx_configs={
-                    "text_model": (text_model, SiglipTextModelOnnxConfig(text_model.config)),
-                    "vision_model": (vision_model, SiglipVisionModelOnnxConfig(vision_model.config)),
-                },
-                **custom_export_kwargs,
-            )
-
-        # TODO: Enable once https://github.com/huggingface/optimum/pull/1552 is merged
-        # elif config.model_type == 'clap':
-        #     # Handle special case for exporting text and audio models separately
-        #     from .extra.clap import ClapTextModelWithProjectionOnnxConfig, ClapAudioModelWithProjectionOnnxConfig
-        #     from transformers.models.clap import ClapTextModelWithProjection, ClapAudioModelWithProjection
-
-        #     text_model = ClapTextModelWithProjection.from_pretrained(model_id, **from_pretrained_kwargs)
-        #     audio_model = ClapAudioModelWithProjection.from_pretrained(model_id, **from_pretrained_kwargs)
-
-        #     export_models(
-        #         models_and_onnx_configs={
-        #             "text_model": (text_model, ClapTextModelWithProjectionOnnxConfig(text_model.config)),
-        #             "audio_model": (audio_model, ClapAudioModelWithProjectionOnnxConfig(audio_model.config)),
-        #         },
-        #         **custom_export_kwargs,
-        #     )
-
-        else:
-            raise Exception(f"Unable to export {config.model_type} model with `--split_modalities`.")
-
-    # Step 2. (optional, recommended) quantize the converted model for fast inference and to reduce model size.
-    if conv_args.quantize:
-        # Update quantize config with model specific defaults
-        quantize_config = MODEL_SPECIFIC_QUANTIZE_PARAMS.get(config.model_type, DEFAULT_QUANTIZE_PARAMS)
-
-        # Update if user specified values
-        if conv_args.per_channel is not None:
-            quantize_config["per_channel"] = conv_args.per_channel
-
-        if conv_args.reduce_range is not None:
-            quantize_config["reduce_range"] = conv_args.reduce_range
-
-        quantize(
-            [
-                os.path.join(output_model_folder, x)
-                for x in os.listdir(output_model_folder)
-                if x.endswith(".onnx") and not x.endswith("_quantized.onnx")
-            ],
-            **quantize_config,
-        )
-
-    # Step 3. Move .onnx files to the 'onnx' subfolder
-    os.makedirs(os.path.join(output_model_folder, "onnx"), exist_ok=True)
-    for file in os.listdir(output_model_folder):
-        if file.endswith((".onnx", ".onnx_data")):
-            shutil.move(os.path.join(output_model_folder, file), os.path.join(output_model_folder, "onnx", file))
-
-    # Step 4. Update the generation config if necessary
-    if config.model_type == "whisper":
-        from transformers import GenerationConfig
-
-        from .extra.whisper import get_alignment_heads
-
-        generation_config = GenerationConfig.from_pretrained(model_id, **from_pretrained_kwargs)
-        generation_config.alignment_heads = get_alignment_heads(config)
-        generation_config.save_pretrained(output_model_folder)
-
-
-if __name__ == "__main__":
-    main()

experiments/try_attention_export.py

@@ -0,0 +1,29 @@
+import numpy as np
+import onnx
+import onnxruntime
+from transformers import AutoTokenizer
+
+model_id = "sentence-transformers/paraphrase-MiniLM-L6-v2"
+output_dir = f"models/{model_id.replace('/', '_')}"
+model_kwargs = {"output_attentions": True, "return_dict": True}
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+model_path = f"{output_dir}/model.onnx"
+onnx_model = onnx.load(model_path)
+ort_session = onnxruntime.InferenceSession(model_path)
+text = "This is a test sentence"
+tokenizer_output = tokenizer(text, return_tensors="np")
+input_ids = tokenizer_output["input_ids"]
+attention_mask = tokenizer_output["attention_mask"]
+print(attention_mask)
+# Prepare the input
+input_ids = np.array(input_ids).astype(np.int64)  # Replace your_input_ids with actual input data
+
+# Run the ONNX model
+outputs = ort_session.run(None, {"input_ids": input_ids, "attention_mask": attention_mask})
+
+# Get the attention weights
+attentions = outputs[-1]
+
+# Print the attention weights for the first layer and first head
+print(attentions[0][0])

experiments/try_quantized_model.py

@@ -1,56 +0,0 @@
-from pathlib import Path
-from typing import List
-
-import click
-import numpy as np
-import torch
-import torch.nn.functional as F
-from optimum.onnxruntime import ORTModelForFeatureExtraction
-from optimum.pipelines import pipeline
-from torch import Tensor
-from transformers import AutoModel, AutoTokenizer
-
-
-def average_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor:
-    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
-    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
-
-
-def hf_embed(model_id: str, texts: List[str], tokenizer):
-    # Tokenize the input texts
-    model = AutoModel.from_pretrained(model_id)
-    model.eval()
-    encoded_input = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
-
-    model_output = model(**encoded_input)
-    sentence_embeddings = model_output[0][:, 0]
-    sentence_embeddings = torch.nn.functional.normalize(sentence_embeddings, p=2, dim=1)
-    return sentence_embeddings
-
-
-@click.command()
-@click.option("--model_id", help="model id from huggingface.co/models")
-@click.option("--model_dir", help="The person to greet.")
-def setup(model_id, model_dir):
-    text = "This is a test sentence"
-    tokenizer = AutoTokenizer.from_pretrained(model_id)
-    output_dir = Path(model_dir)
-    model = ORTModelForFeatureExtraction.from_pretrained(output_dir)
-    onnx_quant_embed = pipeline(
-        "feature-extraction", model=model, accelerator="ort", tokenizer=tokenizer, return_tensors=True
-    )
-    quant_embeddings = onnx_quant_embed([text])
-    quant_embeddings = F.normalize(quant_embeddings[0][:,0], p=2, dim=1)
-    quant_embeddings = quant_embeddings.detach().numpy()
-    print(quant_embeddings.shape)
-
-    torch_embeddings = hf_embed(model_id, texts=[text], tokenizer=tokenizer)
-    torch_embeddings = F.normalize(torch_embeddings, p=2, dim=1)
-    torch_embeddings = torch_embeddings.detach().numpy()
-    print(torch_embeddings.shape)
-    assert quant_embeddings.shape == torch_embeddings.shape
-    print(np.allclose(quant_embeddings, torch_embeddings, atol=1e-5))
-
-
-if __name__ == "__main__":
-    setup()

fastembed/__init__.py

@@ -1,3 +1,7 @@
-from fastembed.text.text_embedding import TextEmbedding
+import importlib.metadata
 
-__all__ = ["TextEmbedding"]
+from fastembed.text import TextEmbedding
+from fastembed.sparse import SparseTextEmbedding, SparseEmbedding
+
+__version__ = importlib.metadata.version("fastembed")
+__all__ = ["TextEmbedding", "SparseTextEmbedding", "SparseEmbedding"]

fastembed/common/model_management.py

@@ -11,24 +11,36 @@ from tqdm import tqdm
 from loguru import logger
 
 
-def locate_model_file(model_dir: Path, file_names: List[str]) -> Path:
-    """
-    Find model path for both TransformerJS style `onnx`  subdirectory structure and direct model weights structure used
-    by Optimum and Qdrant
-    """
-    if not model_dir.is_dir():
-        raise ValueError(f"Provided model path '{model_dir}' is not a directory.")
-
-    for file_name in file_names:
-        file_paths = [path for path in model_dir.rglob(file_name) if path.is_file()]
-
-        if file_paths:
-            return file_paths[0]
-
-    raise ValueError(f"Could not find either of {', '.join(file_names)} in {model_dir}")
-
-
 class ModelManagement:
+    @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.
+        """
+        raise NotImplementedError()
+
+    @classmethod
+    def _get_model_description(cls, model_name: str) -> Dict[str, Any]:
+        """
+        Gets the model description from the model_name.
+
+        Args:
+            model_name (str): The name of the model.
+
+        raises:
+            ValueError: If the model_name is not supported.
+
+        Returns:
+            Dict[str, Any]: The model description.
+        """
+        for model in cls.list_supported_models():
+            if model_name.lower() == model["model"].lower():
+                return model
+
+        raise ValueError(f"Model {model_name} is not supported in {cls.__name__}.")
+
     @classmethod
     def download_file_from_gcs(cls, url: str, output_path: str, show_progress: bool = True) -> str:
         """
@@ -63,7 +75,9 @@ class ModelManagement:
 
         show_progress = total_size_in_bytes and show_progress
 
-        with tqdm(total=total_size_in_bytes, unit="iB", unit_scale=True, disable=not show_progress) as progress_bar:
+        with tqdm(
+            total=total_size_in_bytes, unit="iB", unit_scale=True, disable=not show_progress
+        ) as progress_bar:
             with open(output_path, "wb") as file:
                 for chunk in response.iter_content(chunk_size=1024):
                     if chunk:  # Filter out keep-alive new chunks
@@ -72,19 +86,34 @@ class ModelManagement:
         return output_path
 
     @classmethod
-    def download_files_from_huggingface(cls, hf_source_repo: str, cache_dir: Optional[str] = None) -> str:
+    def download_files_from_huggingface(
+        cls,
+        hf_source_repo: str,
+        cache_dir: Optional[str] = None,
+        extra_patterns: Optional[List[str]] = None,
+    ) -> 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
+                includes the required model files.
         Returns:
             Path: The path to the model directory.
         """
+        allow_patterns = [
+            "config.json",
+            "tokenizer.json",
+            "tokenizer_config.json",
+            "special_tokens_map.json",
+        ]
+        if extra_patterns is not None:
+            allow_patterns.extend(extra_patterns)
 
         return snapshot_download(
             repo_id=hf_source_repo,
-            ignore_patterns=["model.safetensors", "pytorch_model.bin"],
+            allow_patterns=allow_patterns,
             cache_dir=cache_dir,
         )
 
@@ -142,6 +171,9 @@ class ModelManagement:
 
         model_tar_gz = Path(cache_dir) / f"{fast_model_name}.tar.gz"
 
+        if model_tar_gz.exists():
+            model_tar_gz.unlink()
+
         cls.download_file_from_gcs(
             source_url,
             output_path=str(model_tar_gz),
@@ -186,10 +218,20 @@ class ModelManagement:
         url_source = model.get("sources", {}).get("url")
 
         if hf_source:
+            extra_patterns = [model["model_file"]]
+            extra_patterns.extend(model.get("additional_files", []))
+
             try:
-                return Path(cls.download_files_from_huggingface(hf_source, cache_dir=str(cache_dir)))
+                return Path(
+                    cls.download_files_from_huggingface(
+                        hf_source, cache_dir=str(cache_dir), extra_patterns=extra_patterns
+                    )
+                )
             except (EnvironmentError, RepositoryNotFoundError, ValueError) as e:
-                logger.error(f"Could not download model from HuggingFace: {e}" "Falling back to other sources.")
+                logger.error(
+                    f"Could not download model from HuggingFace: {e}"
+                    "Falling back to other sources."
+                )
 
         if url_source:
             return cls.retrieve_model_gcs(model["model"], url_source, str(cache_dir))

fastembed/common/models.py

@@ -33,7 +33,9 @@ def load_tokenizer(model_dir: Path, max_length: int = 512) -> Tokenizer:
 
     tokenizer = Tokenizer.from_file(str(tokenizer_path))
     tokenizer.enable_truncation(max_length=min(tokenizer_config["model_max_length"], max_length))
-    tokenizer.enable_padding(pad_id=config.get("pad_token_id", 0), pad_token=tokenizer_config["pad_token"])
+    tokenizer.enable_padding(
+        pad_id=config.get("pad_token_id", 0), pad_token=tokenizer_config["pad_token"]
+    )
 
     for token in tokens_map.values():
         if isinstance(token, str):

fastembed/common/onnx_model.py

@@ -0,0 +1,163 @@
+import os
+from multiprocessing import get_all_start_methods
+from pathlib import Path
+from typing import (
+    Any,
+    Dict,
+    Generic,
+    Iterable,
+    List,
+    Optional,
+    Tuple,
+    Type,
+    TypeVar,
+    Union,
+    Sequence,
+)
+
+import numpy as np
+import onnxruntime as ort
+
+from fastembed.common.models import load_tokenizer
+from fastembed.common.utils import iter_batch
+from fastembed.parallel_processor import ParallelWorkerPool, Worker
+
+
+# Holds type of the embedding result
+T = TypeVar("T")
+
+
+class OnnxModel(Generic[T]):
+    @classmethod
+    def _get_worker_class(cls) -> Type["EmbeddingWorker"]:
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @classmethod
+    def _post_process_onnx_output(cls, output: Tuple[np.ndarray, np.ndarray]) -> Iterable[T]:
+        raise NotImplementedError("Subclasses must implement this method")
+
+    def __init__(self) -> None:
+        self.model = None
+        self.tokenizer = None
+
+    def _preprocess_onnx_input(self, onnx_input: Dict[str, np.ndarray]) -> Dict[str, np.ndarray]:
+        """
+        Preprocess the onnx input.
+        """
+        return onnx_input
+
+    def load_onnx_model(
+        self,
+        model_dir: Path,
+        model_file: str,
+        threads: Optional[int],
+        providers: Optional[Sequence[Union[str, Tuple[str, Dict[Any, Any]]]]] = None,
+    ) -> None:
+        model_path = model_dir / model_file
+
+        # List of Execution Providers: https://onnxruntime.ai/docs/execution-providers
+
+        onnx_providers = ["CPUExecutionProvider"] if providers is None else list(providers)
+        available_providers = ort.get_available_providers()
+        for provider in onnx_providers:
+            # check providers available
+            provider_name = provider if isinstance(provider, str) else provider[0]
+            if provider_name not in available_providers:
+                raise ValueError(
+                    f"Provider {provider_name} is not available. Available providers: {available_providers}"
+                )
+
+        so = ort.SessionOptions()
+        so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+
+        if threads is not None:
+            so.intra_op_num_threads = threads
+            so.inter_op_num_threads = threads
+
+        self.tokenizer = load_tokenizer(model_dir=model_dir)
+        self.model = ort.InferenceSession(
+            str(model_path), providers=onnx_providers, sess_options=so
+        )
+
+    def onnx_embed(self, documents: List[str]) -> Tuple[np.ndarray, np.ndarray]:
+        encoded = self.tokenizer.encode_batch(documents)
+        input_ids = np.array([e.ids for e in encoded])
+        attention_mask = np.array([e.attention_mask for e in encoded])
+
+        onnx_input = {
+            "input_ids": np.array(input_ids, dtype=np.int64),
+            "attention_mask": np.array(attention_mask, dtype=np.int64),
+            "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)
+
+        model_output = self.model.run(None, onnx_input)
+        embeddings = model_output[0]
+        return embeddings, attention_mask
+
+    def _embed_documents(
+        self,
+        model_name: str,
+        cache_dir: str,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+    ) -> Iterable[T]:
+        is_small = False
+
+        if isinstance(documents, str):
+            documents = [documents]
+            is_small = True
+
+        if isinstance(documents, list):
+            if len(documents) < batch_size:
+                is_small = True
+
+        if parallel == 0:
+            parallel = os.cpu_count()
+
+        if parallel is None or is_small:
+            for batch in iter_batch(documents, batch_size):
+                yield from self._post_process_onnx_output(self.onnx_embed(batch))
+        else:
+            start_method = "forkserver" if "forkserver" in get_all_start_methods() else "spawn"
+            params = {
+                "model_name": model_name,
+                "cache_dir": cache_dir,
+            }
+            pool = ParallelWorkerPool(
+                parallel, self._get_worker_class(), start_method=start_method
+            )
+            for batch in pool.ordered_map(iter_batch(documents, batch_size), **params):
+                yield from self._post_process_onnx_output(batch)
+
+
+class EmbeddingWorker(Worker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+    ) -> OnnxModel:
+        raise NotImplementedError()
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: str,
+    ):
+        self.model = self.init_embedding(model_name, cache_dir)
+
+    @classmethod
+    def start(cls, model_name: str, cache_dir: str, **kwargs: Any) -> "EmbeddingWorker":
+        return cls(
+            model_name=model_name,
+            cache_dir=cache_dir,
+        )
+
+    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
+        for idx, batch in items:
+            embeddings, attn_mask = self.model.onnx_embed(batch)
+            yield idx, (embeddings, attn_mask)

fastembed/embedding.py

@@ -2,9 +2,12 @@ from typing import Optional
 
 from loguru import logger
 
-from fastembed.text.text_embedding import TextEmbedding
+from fastembed import TextEmbedding
 
-logger.warning("DefaultEmbedding, FlagEmbedding, JinaEmbedding are deprecated." "Use from fastembed import TextEmbedding instead.")
+logger.warning(
+    "DefaultEmbedding, FlagEmbedding, JinaEmbedding are deprecated."
+    "Use from fastembed import TextEmbedding instead."
+)
 
 DefaultEmbedding = TextEmbedding
 FlagEmbedding = TextEmbedding

fastembed/parallel_processor.py

@@ -128,7 +128,9 @@ class ParallelWorkerPool:
                 yield buffer.pop(next_expected)
                 next_expected += 1
 
-    def semi_ordered_map(self, stream: Iterable[Any], *args: Any, **kwargs: Any) -> Iterable[Tuple[int, Any]]:
+    def semi_ordered_map(
+        self, stream: Iterable[Any], *args: Any, **kwargs: Any
+    ) -> Iterable[Tuple[int, Any]]:
         try:
             self.start(**kwargs)
 

fastembed/sparse/__init__.py

@@ -0,0 +1,4 @@
+from fastembed.sparse.sparse_embedding_base import SparseEmbedding
+from fastembed.sparse.sparse_text_embedding import SparseTextEmbedding
+
+__all__ = ["SparseEmbedding", "SparseTextEmbedding"]

fastembed/sparse/sparse_embedding_base.py

@@ -0,0 +1,43 @@
+from dataclasses import dataclass
+from typing import Dict, Iterable, Optional, Union
+
+import numpy as np
+
+from fastembed.common.model_management import ModelManagement
+
+
+@dataclass
+class SparseEmbedding:
+    values: np.ndarray
+    indices: np.ndarray
+
+    def as_object(self) -> Dict[str, np.ndarray]:
+        return {
+            "values": self.values,
+            "indices": self.indices,
+        }
+
+    def as_dict(self) -> Dict[int, float]:
+        return {i: v for i, v in zip(self.indices, self.values)}
+
+
+class SparseTextEmbeddingBase(ModelManagement):
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        **kwargs,
+    ):
+        self.model_name = model_name
+        self.cache_dir = cache_dir
+        self.threads = threads
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs,
+    ) -> Iterable[SparseEmbedding]:
+        raise NotImplementedError()

fastembed/sparse/sparse_text_embedding.py

@@ -0,0 +1,82 @@
+from typing import List, Type, Dict, Any, Union, Iterable, Optional
+
+from fastembed.sparse.sparse_embedding_base import SparseTextEmbeddingBase, SparseEmbedding
+from fastembed.sparse.splade_pp import SpladePP
+
+
+class SparseTextEmbedding(SparseTextEmbeddingBase):
+    EMBEDDINGS_REGISTRY: List[Type[SparseTextEmbeddingBase]] = [
+        SpladePP,
+    ]
+
+    @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.
+
+            Example:
+                ```
+                [
+                    {
+                        "model": "prithvida/SPLADE_PP_en_v1",
+                        "vocab_size": 30522,
+                        "description": "Independent Implementation of SPLADE++ Model for English",
+                        "size_in_GB": 0.532,
+                        "sources": {
+                            "hf": "qdrant/SPLADE_PP_en_v1",
+                        },
+                    }
+                ]
+                ```
+        """
+        result = []
+        for embedding in cls.EMBEDDINGS_REGISTRY:
+            result.extend(embedding.list_supported_models())
+        return result
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        **kwargs,
+    ):
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+
+        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):
+                self.model = EMBEDDING_MODEL_TYPE(model_name, cache_dir, threads, **kwargs)
+                return
+
+        raise ValueError(
+            f"Model {model_name} is not supported in SparseTextEmbedding."
+            "Please check the supported models using `SparseTextEmbedding.list_supported_models()`"
+        )
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs,
+    ) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of documents into list of embeddings.
+        We use mean pooling with attention so that the model can handle variable-length inputs.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            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.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self.model.embed(documents, batch_size, parallel, **kwargs)

fastembed/sparse/splade_pp.py

@@ -0,0 +1,134 @@
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union, Type
+
+import numpy as np
+
+from fastembed.common.onnx_model import EmbeddingWorker, OnnxModel
+from fastembed.common.utils import define_cache_dir
+from fastembed.sparse.sparse_embedding_base import SparseEmbedding, SparseTextEmbeddingBase
+
+supported_splade_models = [
+    {
+        "model": "prithvida/Splade_PP_en_v1",
+        "vocab_size": 30522,
+        "description": "Misspelled version of the model. Retained for backward compatibility. Independent Implementation of SPLADE++ Model for English",
+        "size_in_GB": 0.532,
+        "sources": {
+            "hf": "Qdrant/SPLADE_PP_en_v1",
+        },
+        "model_file": "model.onnx",
+    },
+    {
+        "model": "prithivida/Splade_PP_en_v1",
+        "vocab_size": 30522,
+        "description": "Independent Implementation of SPLADE++ Model for English",
+        "size_in_GB": 0.532,
+        "sources": {
+            "hf": "Qdrant/SPLADE_PP_en_v1",
+        },
+        "model_file": "model.onnx",
+    },
+]
+
+
+class SpladePP(SparseTextEmbeddingBase, OnnxModel[SparseEmbedding]):
+    @classmethod
+    def _post_process_onnx_output(
+        cls, output: Tuple[np.ndarray, np.ndarray]
+    ) -> Iterable[SparseEmbedding]:
+        logits, attention_mask = output
+        relu_log = np.log(1 + np.maximum(logits, 0))
+
+        weighted_log = relu_log * np.expand_dims(attention_mask, axis=-1)
+
+        scores = np.max(weighted_log, axis=1)
+
+        # Score matrix of shape (batch_size, vocab_size)
+        # Most of the values are 0, only a few are non-zero
+        for row_scores in scores:
+            indices = row_scores.nonzero()[0]
+            scores = row_scores[indices]
+            yield SparseEmbedding(values=scores, indices=indices)
+
+    @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_splade_models
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        **kwargs,
+    ):
+        """
+        Args:
+            model_name (str): The name of the model to use.
+            cache_dir (str, optional): The path to the cache directory.
+                                       Can be set using the `FASTEMBED_CACHE_PATH` env variable.
+                                       Defaults to `fastembed_cache` in the system's temp directory.
+            threads (int, optional): The number of threads single onnxruntime session can use. Defaults to None.
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+
+        model_description = self._get_model_description(model_name)
+        cache_dir = define_cache_dir(cache_dir)
+
+        model_dir = self.download_model(model_description, cache_dir)
+
+        self.load_onnx_model(
+            model_dir=model_dir,
+            model_file=model_description["model_file"],
+            threads=threads,
+        )
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs,
+    ) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of documents into list of embeddings.
+        We use mean pooling with attention so that the model can handle variable-length inputs.
+
+        Args:
+            documents: Iterator of documents or single document to embed
+            batch_size: Batch size for encoding -- higher values will use more memory, but be faster
+            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.
+
+        Returns:
+            List of embeddings, one per document
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+        )
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[EmbeddingWorker]:
+        return SpladePPEmbeddingWorker
+
+
+class SpladePPEmbeddingWorker(EmbeddingWorker):
+    def init_embedding(
+        self,
+        model_name: str,
+        cache_dir: str,
+    ) -> SpladePP:
+        return SpladePP(model_name=model_name, cache_dir=cache_dir, threads=1)

fastembed/text/__init__.py

@@ -0,0 +1,3 @@
+from fastembed.text.text_embedding import TextEmbedding
+
+__all__ = ["TextEmbedding"]

fastembed/text/e5_onnx_embedding.py

@@ -2,7 +2,8 @@ from typing import Type, List, Dict, Any
 
 import numpy as np
 
-from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker, EmbeddingWorker
+from fastembed.common.onnx_model import EmbeddingWorker
+from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
 
 supported_multilingual_e5_models = [
     {
@@ -14,15 +15,18 @@ supported_multilingual_e5_models = [
             "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": "Sentence-transformers model for tasks like clustering or semantic search",
-        "size_in_GB": 1.11,
+        "size_in_GB": 1.00,
         "sources": {
             "hf": "xenova/paraphrase-multilingual-mpnet-base-v2",
         },
+        "model_file": "onnx/model.onnx",
     },
 ]
 

fastembed/text/jina_onnx_embedding.py

@@ -1,24 +1,27 @@
-from typing import Type, List, Dict, Any, Tuple
+from typing import Type, List, Dict, Any, Tuple, Iterable
 
 import numpy as np
 
 from fastembed.common.models import normalize
-from fastembed.text.onnx_embedding import OnnxTextEmbedding, EmbeddingWorker, OnnxTextEmbeddingWorker
+from fastembed.common.onnx_model import EmbeddingWorker
+from fastembed.text.onnx_embedding import OnnxTextEmbedding, OnnxTextEmbeddingWorker
 
 supported_jina_models = [
     {
         "model": "jinaai/jina-embeddings-v2-base-en",
         "dim": 768,
         "description": "English embedding model supporting 8192 sequence length",
-        "size_in_GB": 0.55,
+        "size_in_GB": 0.52,
         "sources": {"hf": "xenova/jina-embeddings-v2-base-en"},
+        "model_file": "onnx/model.onnx",
     },
     {
         "model": "jinaai/jina-embeddings-v2-small-en",
         "dim": 512,
         "description": "English embedding model supporting 8192 sequence length",
-        "size_in_GB": 0.13,
+        "size_in_GB": 0.12,
         "sources": {"hf": "xenova/jina-embeddings-v2-small-en"},
+        "model_file": "onnx/model.onnx",
     },
 ]
 
@@ -48,7 +51,9 @@ class JinaOnnxEmbedding(OnnxTextEmbedding):
         return supported_jina_models
 
     @classmethod
-    def _post_process_onnx_output(cls, output: Tuple[np.ndarray, np.ndarray]) -> np.ndarray:
+    def _post_process_onnx_output(
+        cls, output: Tuple[np.ndarray, np.ndarray]
+    ) -> Iterable[np.ndarray]:
         embeddings, attn_mask = output
         return normalize(cls.mean_pooling(embeddings, attn_mask)).astype(np.float32)
 

fastembed/text/onnx_embedding.py

@@ -1,14 +1,10 @@
-import os
-from multiprocessing import get_all_start_methods
-from typing import List, Dict, Any, Optional, Tuple, Union, Iterable, Type
+from typing import Dict, Optional, Tuple, Union, Iterable, Type, List, Any
 
 import numpy as np
-import onnxruntime as ort
 
-from fastembed.common.model_management import locate_model_file
-from fastembed.common.models import load_tokenizer, normalize
-from fastembed.common.utils import define_cache_dir, iter_batch
-from fastembed.parallel_processor import ParallelWorkerPool, Worker
+from fastembed.common.onnx_model import OnnxModel, EmbeddingWorker
+from fastembed.common.models import normalize
+from fastembed.common.utils import define_cache_dir
 from fastembed.text.text_embedding_base import TextEmbeddingBase
 
 supported_onnx_models = [
@@ -16,79 +12,64 @@ supported_onnx_models = [
         "model": "BAAI/bge-base-en",
         "dim": 768,
         "description": "Base English model",
-        "size_in_GB": 0.5,
+        "size_in_GB": 0.42,
         "sources": {
             "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en.tar.gz",
         },
+        "model_file": "model_optimized.onnx",
     },
     {
         "model": "BAAI/bge-base-en-v1.5",
         "dim": 768,
         "description": "Base English model, v1.5",
-        "size_in_GB": 0.44,
+        "size_in_GB": 0.21,
         "sources": {
             "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-base-en-v1.5.tar.gz",
             "hf": "qdrant/bge-base-en-v1.5-onnx-q",
         },
-    },
-    {
-        "model": "BAAI/bge-large-en-v1.5-quantized",
-        "dim": 1024,
-        "description": "Large English model, v1.5",
-        "size_in_GB": 1.34,
-        "sources": {
-            "hf": "qdrant/bge-large-en-v1.5-onnx-q",
-        },
+        "model_file": "model_optimized.onnx",
     },
     {
         "model": "BAAI/bge-large-en-v1.5",
         "dim": 1024,
         "description": "Large English model, v1.5",
-        "size_in_GB": 1.34,
+        "size_in_GB": 1.20,
         "sources": {
             "hf": "qdrant/bge-large-en-v1.5-onnx",
         },
+        "model_file": "model.onnx",
     },
     {
         "model": "BAAI/bge-small-en",
         "dim": 384,
         "description": "Fast English model",
-        "size_in_GB": 0.2,
+        "size_in_GB": 0.13,
         "sources": {
             "url": "https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz",
         },
+        "model_file": "model_optimized.onnx",
     },
-    # {
-    #     "model": "BAAI/bge-small-en",
-    #     "dim": 384,
-    #     "description": "Fast English model",
-    #     "size_in_GB": 0.2,
-    #     "hf_sources": [],
-    #     "compressed_url_sources": [
-    #         "https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-en.tar.gz",
-    #         "https://storage.googleapis.com/qdrant-fastembed/BAAI-bge-small-en.tar.gz"
-    #     ]
-    # },
     {
         "model": "BAAI/bge-small-en-v1.5",
         "dim": 384,
         "description": "Fast and Default English model",
-        "size_in_GB": 0.13,
+        "size_in_GB": 0.067,
         "sources": {
-            "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-en-v1.5.tar.gz",
             "hf": "qdrant/bge-small-en-v1.5-onnx-q",
         },
+        "model_file": "model_optimized.onnx",
     },
     {
         "model": "BAAI/bge-small-zh-v1.5",
         "dim": 512,
         "description": "Fast and recommended Chinese model",
-        "size_in_GB": 0.1,
+        "size_in_GB": 0.09,
         "sources": {
             "url": "https://storage.googleapis.com/qdrant-fastembed/fast-bge-small-zh-v1.5.tar.gz",
         },
+        "model_file": "model_optimized.onnx",
     },
-    {  # todo: it is not a flag embedding
+    {
         "model": "sentence-transformers/all-MiniLM-L6-v2",
         "dim": 384,
         "description": "Sentence Transformer model, MiniLM-L6-v2",
@@ -97,91 +78,134 @@ supported_onnx_models = [
             "url": "https://storage.googleapis.com/qdrant-fastembed/sentence-transformers-all-MiniLM-L6-v2.tar.gz",
             "hf": "qdrant/all-MiniLM-L6-v2-onnx",
         },
+        "model_file": "model.onnx",
     },
     {
         "model": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
         "dim": 384,
         "description": "Sentence Transformer model, paraphrase-multilingual-MiniLM-L12-v2",
-        "size_in_GB": 0.46,
+        "size_in_GB": 0.22,
         "sources": {
             "hf": "qdrant/paraphrase-multilingual-MiniLM-L12-v2-onnx-Q",
         },
+        "model_file": "model_optimized.onnx",
     },
     {
         "model": "nomic-ai/nomic-embed-text-v1",
         "dim": 768,
         "description": "8192 context length english model",
-        "size_in_GB": 0.54,
+        "size_in_GB": 0.52,
         "sources": {
             "hf": "nomic-ai/nomic-embed-text-v1",
         },
+        "model_file": "onnx/model.onnx",
     },
     {
         "model": "nomic-ai/nomic-embed-text-v1.5",
         "dim": 768,
         "description": "8192 context length english model",
-        "size_in_GB": 0.54,
+        "size_in_GB": 0.52,
         "sources": {
             "hf": "nomic-ai/nomic-embed-text-v1.5",
         },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "nomic-ai/nomic-embed-text-v1.5-Q",
+        "dim": 768,
+        "description": "Quantized 8192 context length english model",
+        "size_in_GB": 0.13,
+        "sources": {
+            "hf": "nomic-ai/nomic-embed-text-v1.5",
+        },
+        "model_file": "onnx/model_quantized.onnx",
     },
     {
         "model": "thenlper/gte-large",
         "dim": 1024,
         "description": "Large general text embeddings model",
-        "size_in_GB": 1.34,
+        "size_in_GB": 1.20,
         "sources": {
             "hf": "qdrant/gte-large-onnx",
         },
+        "model_file": "model.onnx",
+    },
+    {
+        "model": "mixedbread-ai/mxbai-embed-large-v1",
+        "dim": 1024,
+        "description": "MixedBread Base sentence embedding model, does well on MTEB",
+        "size_in_GB": 0.64,
+        "sources": {
+            "hf": "mixedbread-ai/mxbai-embed-large-v1",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "snowflake/snowflake-arctic-embed-xs",
+        "dim": 384,
+        "description": "Based on all-MiniLM-L6-v2 model with only 22m parameters, ideal for latency/TCO budgets.",
+        "size_in_GB": 0.09,
+        "sources": {
+            "hf": "snowflake/snowflake-arctic-embed-xs",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "snowflake/snowflake-arctic-embed-s",
+        "dim": 384,
+        "description": "Based on infloat/e5-small-unsupervised, does not trade off retrieval accuracy for its small size.",
+        "size_in_GB": 0.13,
+        "sources": {
+            "hf": "snowflake/snowflake-arctic-embed-s",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "snowflake/snowflake-arctic-embed-m",
+        "dim": 768,
+        "description": "Based on intfloat/e5-base-unsupervised model, provides the best retrieval without slowing down inference.",
+        "size_in_GB": 0.43,
+        "sources": {
+            "hf": "Snowflake/snowflake-arctic-embed-m",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "snowflake/snowflake-arctic-embed-m-long",
+        "dim": 768,
+        "description": "Based on nomic-ai/nomic-embed-text-v1-unsupervised model, 8192 context-length model",
+        "size_in_GB": 0.54,
+        "sources": {
+            "hf": "snowflake/snowflake-arctic-embed-m-long",
+        },
+        "model_file": "onnx/model.onnx",
+    },
+    {
+        "model": "snowflake/snowflake-arctic-embed-l",
+        "dim": 1024,
+        "description": "Based on intfloat/e5-large-unsupervised, large model for most accurate retrieval.",
+        "size_in_GB": 1.02,
+        "sources": {
+            "hf": "snowflake/snowflake-arctic-embed-l",
+        },
+        "model_file": "onnx/model.onnx",
     },
-    # {
-    #     "model": "sentence-transformers/all-MiniLM-L6-v2",
-    #     "dim": 384,
-    #     "description": "Sentence Transformer model, MiniLM-L6-v2",
-    #     "size_in_GB": 0.09,
-    #     "hf_sources": [
-    #         "qdrant/all-MiniLM-L6-v2-onnx"
-    #     ],
-    #     "compressed_url_sources": [
-    #         "https://storage.googleapis.com/qdrant-fastembed/fast-all-MiniLM-L6-v2.tar.gz",
-    #         "https://storage.googleapis.com/qdrant-fastembed/sentence-transformers-all-MiniLM-L6-v2.tar.gz"
-    #     ]
-    # }
 ]
 
 
-class OnnxTextEmbedding(TextEmbeddingBase):
+class OnnxTextEmbedding(TextEmbeddingBase, OnnxModel[np.ndarray]):
     """Implementation of the Flag Embedding model."""
 
     @classmethod
     def list_supported_models(cls) -> List[Dict[str, Any]]:
-        """Lists the supported models.
+        """
+        Lists the supported models.
 
         Returns:
             List[Dict[str, Any]]: A list of dictionaries containing the model information.
         """
         return supported_onnx_models
 
-    @classmethod
-    def _get_model_description(cls, model_name: str) -> Dict[str, Any]:
-        """
-        Gets the model description from the model_name.
-
-        Args:
-            model_name (str): The name of the model.
-
-        raises:
-            ValueError: If the model_name is not supported.
-
-        Returns:
-            Dict[str, Any]: The model description.
-        """
-        for model in cls.list_supported_models():
-            if model_name == model["model"]:
-                return model
-
-        raise ValueError(f"Model {model_name} is not supported in FlagEmbedding.")
-
     def __init__(
         self,
         model_name: str = "BAAI/bge-small-en-v1.5",
@@ -203,27 +227,15 @@ class OnnxTextEmbedding(TextEmbeddingBase):
 
         super().__init__(model_name, cache_dir, threads, **kwargs)
 
-        self.model_name = model_name
-        self._model_description = self._get_model_description(model_name)
+        model_description = self._get_model_description(model_name)
+        cache_dir = define_cache_dir(cache_dir)
+        model_dir = self.download_model(model_description, cache_dir)
 
-        self._cache_dir = define_cache_dir(cache_dir)
-        self._model_dir = self.download_model(self._model_description, self._cache_dir)
-        self._max_length = 512
-
-        model_path = locate_model_file(self._model_dir, ["model.onnx", "model_optimized.onnx"])
-
-        # List of Execution Providers: https://onnxruntime.ai/docs/execution-providers
-        onnx_providers = ["CPUExecutionProvider"]
-
-        so = ort.SessionOptions()
-        so.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-
-        if self.threads is not None:
-            so.intra_op_num_threads = self.threads
-            so.inter_op_num_threads = self.threads
-
-        self.tokenizer = load_tokenizer(model_dir=self._model_dir, max_length=self._max_length)
-        self.model = ort.InferenceSession(str(model_path), providers=onnx_providers, sess_options=so)
+        self.load_onnx_model(
+            model_dir=model_dir,
+            model_file=model_description["model_file"],
+            threads=threads,
+        )
 
     def embed(
         self,
@@ -247,31 +259,13 @@ class OnnxTextEmbedding(TextEmbeddingBase):
         Returns:
             List of embeddings, one per document
         """
-        is_small = False
-
-        if isinstance(documents, str):
-            documents = [documents]
-            is_small = True
-
-        if isinstance(documents, list):
-            if len(documents) < batch_size:
-                is_small = True
-
-        if parallel == 0:
-            parallel = os.cpu_count()
-
-        if parallel is None or is_small:
-            for batch in iter_batch(documents, batch_size):
-                yield from self._post_process_onnx_output(self.onnx_embed(batch))
-        else:
-            start_method = "forkserver" if "forkserver" in get_all_start_methods() else "spawn"
-            params = {
-                "model_name": self.model_name,
-                "cache_dir": str(self._cache_dir),
-            }
-            pool = ParallelWorkerPool(parallel, self._get_worker_class(), start_method=start_method)
-            for batch in pool.ordered_map(iter_batch(documents, batch_size), **params):
-                yield from self._post_process_onnx_output(batch)
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=documents,
+            batch_size=batch_size,
+            parallel=parallel,
+        )
 
     @classmethod
     def _get_worker_class(cls) -> Type["EmbeddingWorker"]:
@@ -284,55 +278,12 @@ class OnnxTextEmbedding(TextEmbeddingBase):
         return onnx_input
 
     @classmethod
-    def _post_process_onnx_output(cls, output: Tuple[np.ndarray, np.ndarray]):
+    def _post_process_onnx_output(
+        cls, output: Tuple[np.ndarray, np.ndarray]
+    ) -> Iterable[np.ndarray]:
         embeddings, _ = output
         return normalize(embeddings[:, 0]).astype(np.float32)
 
-    def onnx_embed(self, documents: List[str]) -> Tuple[np.ndarray, np.ndarray]:
-        encoded = self.tokenizer.encode_batch(documents)
-        input_ids = np.array([e.ids for e in encoded])
-        attention_mask = np.array([e.attention_mask for e in encoded])
-
-        onnx_input = {
-            "input_ids": np.array(input_ids, dtype=np.int64),
-            "attention_mask": np.array(attention_mask, dtype=np.int64),
-            "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)
-
-        model_output = self.model.run(None, onnx_input)
-        embeddings = model_output[0]
-        return embeddings, attention_mask
-
-
-class EmbeddingWorker(Worker):
-    def init_embedding(
-        self,
-        model_name: str,
-        cache_dir: str,
-    ) -> OnnxTextEmbedding:
-        raise NotImplementedError()
-
-    def __init__(
-        self,
-        model_name: str,
-        cache_dir: str,
-    ):
-        self.model = self.init_embedding(model_name, cache_dir)
-
-    @classmethod
-    def start(cls, model_name: str, cache_dir: str, **kwargs: Any) -> "EmbeddingWorker":
-        return cls(
-            model_name=model_name,
-            cache_dir=cache_dir,
-        )
-
-    def process(self, items: Iterable[Tuple[int, Any]]) -> Iterable[Tuple[int, Any]]:
-        for idx, batch in items:
-            embeddings, attn_mask = self.model.onnx_embed(batch)
-            yield idx, (embeddings, attn_mask)
-
 
 class OnnxTextEmbeddingWorker(EmbeddingWorker):
     def init_embedding(

fastembed/text/text_embedding.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union, Iterable, List, Dict, Any, Type
+from typing import Any, Dict, Iterable, List, Optional, Type, Union
 
 import numpy as np
 
@@ -53,10 +53,10 @@ class TextEmbedding(TextEmbeddingBase):
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
 
-        for embedding in self.EMBEDDINGS_REGISTRY:
-            supported_models = embedding.list_supported_models()
-            if any(model_name == model["model"] for model in supported_models):
-                self.model = embedding(model_name, cache_dir, threads, **kwargs)
+        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):
+                self.model = EMBEDDING_MODEL_TYPE(model_name, cache_dir, threads, **kwargs)
                 return
 
         raise ValueError(

fastembed/text/text_embedding_base.py

@@ -1,4 +1,4 @@
-from typing import Any, Dict, Iterable, List, Optional, Union
+from typing import Iterable, Optional, Union
 
 import numpy as np
 
@@ -6,11 +6,13 @@ from fastembed.common.model_management import ModelManagement
 
 
 class TextEmbeddingBase(ModelManagement):
-    @classmethod
-    def list_supported_models(cls) -> List[Dict[str, Any]]:
-        raise NotImplementedError()
-
-    def __init__(self, model_name: str, cache_dir: Optional[str] = None, threads: Optional[int] = None, **kwargs):
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        **kwargs,
+    ):
         self.model_name = model_name
         self.cache_dir = cache_dir
         self.threads = threads

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "fastembed"
-version = "0.2.2"
+version = "0.2.6"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["NirantK <nirant.bits@gmail.com>"]
 license = "Apache License"
@@ -12,8 +12,16 @@ keywords = ["vector", "embedding", "neural", "search", "qdrant", "sentence-trans
 
 [tool.poetry.dependencies]
 python = ">=3.8.0,<3.13"
-onnx = "^1.15.0"
-onnxruntime = "^1.17.0"
+onnx = [
+    {version = "^1.15.0", optional = true, markers = "extra != 'gpu'"}
+]
+onnxruntime = [
+    {version = "^1.17.0", optional = true, markers = "extra != 'gpu'"}
+]
+onnxruntime-gpu = [
+    { version = "^1.17.0", optional = true, python = "<3.13", markers = "extra == 'gpu'" }
+]
+
 tqdm = "^4.66"
 requests = "^2.31"
 tokenizers = "^0.15.1"
@@ -26,20 +34,23 @@ numpy = [
 
 [tool.poetry.group.dev.dependencies]
 pytest = "^7.4.2"
-ruff = "^0.2.2"
+ruff = "^0.3.1"
 notebook = ">=7.0.2"
+pre-commit = {version = "^3.6.2", python = ">=3.9,<3.12" }
+
+[tool.poetry.group.docs.dependencies]
 mkdocs-material = "^9.5.10"
 mkdocstrings = "^0.24.0"
 pillow = "^10.2.0"
 cairosvg = "^2.7.1"
 mknotebooks = "^0.8.0"
-pre-commit = {version = "^3.6.2", python = ">=3.9,<3.12" }
-click = "^8.1.7"
 
+[tool.poetry.extras]
+gpu = ["onnxruntime-gpu"]
 
 [build-system]
 requires = ["poetry-core"]
 build-backend = "poetry.core.masonry.api"
 
 [tool.ruff]
-line-length = 120
+line-length = 99

tests/test_sparse_embeddings.py

@@ -0,0 +1,98 @@
+import pytest
+from fastembed.sparse.sparse_text_embedding import SparseTextEmbedding
+
+CANONICAL_COLUMN_VALUES = {
+    "prithvida/Splade_PP_en_v1": {
+        "indices": [
+            2040,
+            2047,
+            2088,
+            2299,
+            2748,
+            3011,
+            3376,
+            3795,
+            4774,
+            5304,
+            5798,
+            6160,
+            7592,
+            7632,
+            8484,
+        ],
+        "values": [
+            0.4219532012939453,
+            0.4320072531700134,
+            2.766580104827881,
+            0.3314574658870697,
+            1.395172119140625,
+            0.021595917642116547,
+            0.43770670890808105,
+            0.0008370947907678783,
+            0.5187209844589233,
+            0.17124654352664948,
+            0.14742016792297363,
+            0.8142819404602051,
+            2.803262710571289,
+            2.1904349327087402,
+            1.0531445741653442,
+        ],
+    }
+}
+
+docs = ["Hello World"]
+
+
+def test_batch_embedding():
+    docs_to_embed = docs * 10
+
+    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
+        print("evaluating", model_name)
+        model = SparseTextEmbedding(model_name=model_name)
+        result = next(iter(model.embed(docs_to_embed, batch_size=6)))
+        print(result.indices)
+
+        assert result.indices.tolist() == expected_result["indices"]
+
+        for i, value in enumerate(result.values):
+            assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
+
+
+def test_single_embedding():
+    docs_to_embed = docs
+
+    for model_name, expected_result in CANONICAL_COLUMN_VALUES.items():
+        print("evaluating", model_name)
+        model = SparseTextEmbedding(model_name=model_name)
+        result = next(iter(model.embed(docs_to_embed, batch_size=6)))
+        print(result.indices)
+
+        assert result.indices.tolist() == expected_result["indices"]
+
+        for i, value in enumerate(result.values):
+            assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
+
+
+def test_parallel_processing():
+    import numpy as np
+
+    model = SparseTextEmbedding(
+        model_name="prithivida/Splade_PP_en_v1",
+    )
+    docs = ["hello world", "flag embedding"] * 30
+    sparse_embeddings_duo = list(model.embed(docs, batch_size=10, parallel=2))
+    sparse_embeddings_all = list(model.embed(docs, batch_size=10, parallel=0))
+    sparse_embeddings = list(model.embed(docs, batch_size=10, parallel=None))
+
+    assert len(sparse_embeddings) == len(sparse_embeddings_duo) == len(sparse_embeddings_all) == len(docs)
+
+    for sparse_embedding, sparse_embedding_duo, sparse_embedding_all in zip(
+        sparse_embeddings, sparse_embeddings_duo, sparse_embeddings_all
+    ):
+        assert (
+            sparse_embedding.indices.tolist()
+            == sparse_embedding_duo.indices.tolist()
+            == sparse_embedding_all.indices.tolist()
+        )
+        assert np.allclose(sparse_embedding.values, sparse_embedding_duo.values, atol=1e-3)
+        assert np.allclose(sparse_embedding.values, sparse_embedding_all.values, atol=1e-3)

tests/test_text_onnx_embeddings.py

@@ -8,6 +8,7 @@ from fastembed.text.text_embedding import TextEmbedding
 CANONICAL_VECTOR_VALUES = {
     "BAAI/bge-small-en": np.array([-0.0232, -0.0255, 0.0174, -0.0639, -0.0006]),
     "BAAI/bge-small-en-v1.5": np.array([0.01522374, -0.02271799, 0.00860278, -0.07424029, 0.00386434]),
+    "BAAI/bge-small-en-v1.5-quantized": np.array([0.01522374, -0.02271799, 0.00860278, -0.07424029, 0.00386434]),
     "BAAI/bge-small-zh-v1.5": np.array([-0.01023294, 0.07634465, 0.0691722, -0.04458365, -0.03160762]),
     "BAAI/bge-base-en": np.array([0.0115, 0.0372, 0.0295, 0.0121, 0.0346]),
     "BAAI/bge-base-en-v1.5": np.array([0.01129394, 0.05493144, 0.02615099, 0.00328772, 0.02996045]),
@@ -25,15 +26,26 @@ CANONICAL_VECTOR_VALUES = {
     "nomic-ai/nomic-embed-text-v1.5": np.array(
         [-1.6531514e-02, 8.5380634e-05, -1.8171231e-01, -3.9333291e-03, 1.2763254e-02]
     ),
+    "nomic-ai/nomic-embed-text-v1.5-Q": np.array(
+        [-0.01554983,  0.0129992 , -0.17909265, -0.01062993,  0.00512859]
+    ),
     "thenlper/gte-large": np.array([-0.01920587, 0.00113156, -0.00708992, -0.00632304, -0.04025577]),
+    "mixedbread-ai/mxbai-embed-large-v1": np.array([0.02295546, 0.03196154, 0.016512, -0.04031524, -0.0219634]),
+    "snowflake/snowflake-arctic-embed-xs": np.array([0.0092, 0.0619, 0.0196, 0.009, -0.0114]),
+    "snowflake/snowflake-arctic-embed-s": np.array([-0.0416, -0.0867, 0.0209, 0.0554, -0.0272]),
+    "snowflake/snowflake-arctic-embed-m": np.array([-0.0329, 0.0364, 0.0481, 0.0016, 0.0328]),
+    "snowflake/snowflake-arctic-embed-m-long": np.array(
+        [0.0080, -0.0266, -0.0335, 0.0282, 0.0143]
+    ),
+    "snowflake/snowflake-arctic-embed-l": np.array([0.0189, -0.0673, 0.0183, 0.0124, 0.0146]),
 }
 
 
 def test_embedding():
-    is_ubuntu_ci = os.getenv("IS_UBUNTU_CI")
+    is_ci = os.getenv("CI")
 
     for model_desc in TextEmbedding.list_supported_models():
-        if is_ubuntu_ci == "false" and model_desc["size_in_GB"] > 1:
+        if not is_ci and model_desc["size_in_GB"] > 1:
             continue
 
         dim = model_desc["dim"]