new: gpu package (#224)

* new: add embedding size property

* fix: format exception message

* new: replace embedding size property with get_embedding_size classmethod

* fix: fix missed parts

* chore: fix docstrings

* new: add embedding_size property

.github/workflows/python-tests.yml

@@ -1,4 +1,5 @@
 name: Tests
+run-name: Tests (gpu)
 
 on:
   pull_request:
@@ -22,8 +23,6 @@ jobs:
           - '3.13.x'
         os:
           - ubuntu-latest
-          - macos-latest
-          - windows-latest
 
     runs-on: ${{ matrix.os }}
 
@@ -42,5 +41,7 @@ jobs:
           poetry install --no-interaction --no-ansi --without dev,docs
 
       - name: Run pytest
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
-          poetry run pytest 
+          poetry run pytest

.github/workflows/type-checkers.yml

@@ -26,8 +26,6 @@ jobs:
           python -m pip install --upgrade pip poetry
           poetry install --no-interaction --no-ansi --without dev,docs,test
 
-          poetry run pip install "numpy<2.0.0"  # https://github.com/python/mypy/issues/17396
-
       - name: mypy
         run: |
             poetry run mypy fastembed \

fastembed/common/onnx_model.py

@@ -28,7 +28,16 @@ class OnnxModel(Generic[T]):
     def _get_worker_class(cls) -> Type["EmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:

fastembed/common/preprocessor_utils.py

@@ -36,9 +36,9 @@ def load_tokenizer(model_dir: Path) -> tuple[Tokenizer, dict[str, int]]:
 
     with open(str(tokenizer_config_path)) as tokenizer_config_file:
         tokenizer_config = json.load(tokenizer_config_file)
-        assert (
-            "model_max_length" in tokenizer_config or "max_length" in tokenizer_config
-        ), "Models without model_max_length or max_length are not supported."
+        assert "model_max_length" in tokenizer_config or "max_length" in tokenizer_config, (
+            "Models without model_max_length or max_length are not supported."
+        )
         if "model_max_length" not in tokenizer_config:
             max_context = tokenizer_config["max_length"]
         elif "max_length" not in tokenizer_config:

fastembed/common/types.py

@@ -16,6 +16,7 @@ ImageInput: TypeAlias = Union[PathInput, Image.Image]
 
 OnnxProvider: TypeAlias = Union[str, tuple[str, dict[Any, Any]]]
 NumpyArray = Union[
+    NDArray[np.float64],
     NDArray[np.float32],
     NDArray[np.float16],
     NDArray[np.int8],

fastembed/image/image_embedding.py

@@ -77,6 +77,40 @@ class ImageEmbedding(ImageEmbeddingBase):
             "Please check the supported models using `ImageEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: Optional[int] = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
         images: Union[ImageInput, Iterable[ImageInput]],

fastembed/image/image_embedding_base.py

@@ -18,6 +18,7 @@ class ImageEmbeddingBase(ModelManagement[DenseModelDescription]):
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: Optional[int] = None
 
     def embed(
         self,
@@ -42,3 +43,13 @@ class ImageEmbeddingBase(ModelManagement[DenseModelDescription]):
             Iterable[NdArray]: The embeddings.
         """
         raise NotImplementedError()
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/image/onnx_embedding.py

@@ -112,11 +112,12 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -177,6 +178,8 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 
@@ -193,7 +196,9 @@ class OnnxImageEmbedding(ImageEmbeddingBase, OnnxImageModel[NumpyArray]):
 
         return onnx_input
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         return normalize(output.model_output)
 
 

fastembed/image/onnx_image_model.py

@@ -23,7 +23,16 @@ class OnnxImageModel(OnnxModel[T]):
     def _get_worker_class(cls) -> Type["ImageEmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:
@@ -88,6 +97,8 @@ class OnnxImageModel(OnnxModel[T]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_ids: Optional[list[int]] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -104,7 +115,7 @@ class OnnxImageModel(OnnxModel[T]):
                 self.load_onnx_model()
 
             for batch in iter_batch(images, batch_size):
-                yield from self._post_process_onnx_output(self.onnx_embed(batch))
+                yield from self._post_process_onnx_output(self.onnx_embed(batch), **kwargs)
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -114,6 +125,8 @@ class OnnxImageModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
@@ -125,7 +138,7 @@ class OnnxImageModel(OnnxModel[T]):
                 start_method=start_method,
             )
             for batch in pool.ordered_map(iter_batch(images, batch_size), **params):
-                yield from self._post_process_onnx_output(batch)  # type: ignore
+                yield from self._post_process_onnx_output(batch, **kwargs)  # type: ignore
 
 
 class ImageEmbeddingWorker(EmbeddingWorker[T]):

fastembed/image/transform/functional.py

@@ -72,26 +72,26 @@ def normalize(
     if not np.issubdtype(image.dtype, np.floating):
         image = image.astype(np.float32)
 
-    mean = mean if isinstance(mean, list) else [mean] * num_channels
+    mean_list = mean if isinstance(mean, list) else [mean] * num_channels
 
-    if len(mean) != num_channels:
+    if len(mean_list) != num_channels:
         raise ValueError(
             f"mean must have the same number of channels as the image, image has {num_channels} channels, got "
-            f"{len(mean)}"
+            f"{len(mean_list)}"
         )
 
-    mean_arr = np.array(mean, dtype=np.float32)
+    mean_arr = np.array(mean_list, dtype=np.float32)
 
-    std = std if isinstance(std, list) else [std] * num_channels
-    if len(std) != num_channels:
+    std_list = std if isinstance(std, list) else [std] * num_channels
+    if len(std_list) != num_channels:
         raise ValueError(
-            f"std must have the same number of channels as the image, image has {num_channels} channels, got {len(std)}"
+            f"std must have the same number of channels as the image, image has {num_channels} channels, got {len(std_list)}"
         )
 
-    std_arr = np.array(std, dtype=np.float32)
+    std_arr = np.array(std_list, dtype=np.float32)
 
-    image = ((image.T - mean_arr) / std_arr).T
-    return image
+    image_upd = ((image.T - mean_arr) / std_arr).T
+    return image_upd
 
 
 def resize(

fastembed/late_interaction/colbert.py

@@ -43,7 +43,7 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
     MASK_TOKEN = "[MASK]"
 
     def _post_process_onnx_output(
-        self, output: OnnxOutputContext, is_doc: bool = True
+        self, output: OnnxOutputContext, is_doc: bool = True, **kwargs: Any
     ) -> Iterable[NumpyArray]:
         if not is_doc:
             return output.model_output
@@ -169,11 +169,12 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
         self.mask_token_id: Optional[int] = None
         self.pad_token_id: Optional[int] = None
@@ -233,6 +234,8 @@ class Colbert(LateInteractionTextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 

fastembed/late_interaction/late_interaction_embedding_base.py

@@ -17,6 +17,7 @@ class LateInteractionTextEmbeddingBase(ModelManagement[DenseModelDescription]):
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: Optional[int] = None
 
     def embed(
         self,
@@ -58,3 +59,13 @@ class LateInteractionTextEmbeddingBase(ModelManagement[DenseModelDescription]):
             yield from self.embed([query], **kwargs)
         else:
             yield from self.embed(query, **kwargs)
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/late_interaction/late_interaction_text_embedding.py

@@ -80,6 +80,40 @@ class LateInteractionTextEmbedding(LateInteractionTextEmbeddingBase):
             "Please check the supported models using `LateInteractionTextEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: Optional[int] = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
         documents: Union[str, Iterable[str]],

fastembed/late_interaction/token_embeddings.py

@@ -0,0 +1,83 @@
+from dataclasses import asdict
+from typing import Union, Iterable, Optional, Any, Type
+
+from fastembed.common.model_description import DenseModelDescription, ModelSource
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common.types import NumpyArray
+from fastembed.late_interaction.late_interaction_embedding_base import (
+    LateInteractionTextEmbeddingBase,
+)
+from fastembed.text.onnx_embedding import OnnxTextEmbedding
+from fastembed.text.onnx_text_model import TextEmbeddingWorker
+
+
+supported_token_embeddings_models = [
+    DenseModelDescription(
+        model="jinaai/jina-embeddings-v2-small-en-tokens",
+        dim=512,
+        description="Text embeddings, Unimodal (text), English, 8192 input tokens truncation,"
+        " Prefixes for queries/documents: not necessary, 2023 year.",
+        license="apache-2.0",
+        size_in_GB=0.12,
+        sources=ModelSource(hf="xenova/jina-embeddings-v2-small-en"),
+        model_file="onnx/model.onnx",
+    ),
+]
+
+
+class TokenEmbeddingsModel(OnnxTextEmbedding, LateInteractionTextEmbeddingBase):
+    @classmethod
+    def _list_supported_models(cls) -> list[DenseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[DenseModelDescription]: A list of DenseModelDescription objects containing the model information.
+        """
+        return supported_token_embeddings_models
+
+    @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 [asdict(model) for model in cls._list_supported_models()]
+
+    @classmethod
+    def _get_worker_class(cls) -> Type[TextEmbeddingWorker[NumpyArray]]:
+        return TokensEmbeddingWorker
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
+        # Size: (batch_size, sequence_length, hidden_size)
+        embeddings = output.model_output
+        # Size: (batch_size, sequence_length)
+        assert output.attention_mask is not None
+        masks = output.attention_mask
+
+        # For each document we only select those embeddings that are not masked out
+        for i in range(embeddings.shape[0]):
+            yield embeddings[i, masks[i] == 1]
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> Iterable[NumpyArray]:
+        yield from super().embed(documents, batch_size=batch_size, parallel=parallel, **kwargs)
+
+
+class TokensEmbeddingWorker(TextEmbeddingWorker[NumpyArray]):
+    def init_embedding(
+        self, model_name: str, cache_dir: str, **kwargs: Any
+    ) -> TokenEmbeddingsModel:
+        return TokenEmbeddingsModel(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/late_interaction_multimodal/colpali.py

@@ -95,11 +95,12 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
         self.mask_token_id = None
         self.pad_token_id = None
@@ -174,7 +175,7 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
     ) -> dict[str, NumpyArray]:
         onnx_input["input_ids"] = np.array(
             [
-                self.QUERY_MARKER_TOKEN_ID + input_ids[2:].tolist()
+                self.QUERY_MARKER_TOKEN_ID + input_ids[2:].tolist()  # type: ignore[index]
                 for input_ids in onnx_input["input_ids"]
             ]
         )
@@ -235,6 +236,8 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 
@@ -268,6 +271,8 @@ class ColPali(LateInteractionMultimodalEmbeddingBase, OnnxMultimodalModel[NumpyA
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding.py

@@ -83,6 +83,40 @@ class LateInteractionMultimodalEmbedding(LateInteractionMultimodalEmbeddingBase)
             "Please check the supported models using `LateInteractionMultimodalEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: Optional[int] = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed_text(
         self,
         documents: Union[str, Iterable[str]],

fastembed/late_interaction_multimodal/late_interaction_multimodal_embedding_base.py

@@ -19,6 +19,7 @@ class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescripti
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: Optional[int] = None
 
     def embed_text(
         self,
@@ -65,3 +66,13 @@ class LateInteractionMultimodalEmbeddingBase(ModelManagement[DenseModelDescripti
             List of embeddings, one per image
         """
         raise NotImplementedError()
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the chosen model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")

fastembed/late_interaction_multimodal/onnx_multimodal_model.py

@@ -120,6 +120,8 @@ class OnnxMultimodalModel(OnnxModel[T]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_ids: Optional[list[int]] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -146,6 +148,8 @@ class OnnxMultimodalModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
@@ -183,6 +187,8 @@ class OnnxMultimodalModel(OnnxModel[T]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_ids: Optional[list[int]] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -209,6 +215,8 @@ class OnnxMultimodalModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 

fastembed/rerank/cross_encoder/onnx_text_cross_encoder.py

@@ -131,11 +131,12 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -189,6 +190,8 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 
@@ -196,7 +199,9 @@ class OnnxTextCrossEncoder(TextCrossEncoderBase, OnnxCrossEncoderModel):
     def _get_worker_class(cls) -> Type[TextRerankerWorker]:
         return TextCrossEncoderWorker
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[float]:
         return (float(elem) for elem in output.model_output)
 
 

fastembed/rerank/cross_encoder/onnx_text_model.py

@@ -94,6 +94,8 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_ids: Optional[list[int]] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
         **kwargs: Any,
     ) -> Iterable[float]:
         is_small = False
@@ -120,6 +122,8 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
@@ -133,7 +137,18 @@ class OnnxCrossEncoderModel(OnnxModel[float]):
             for batch in pool.ordered_map(iter_batch(pairs, batch_size), **params):
                 yield from self._post_process_onnx_output(batch)  # type: ignore
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[float]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[float]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[float]: Post-processed output as an iterable of float values.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def _preprocess_onnx_input(

fastembed/sparse/bm25.py

@@ -21,13 +21,9 @@ from fastembed.sparse.sparse_embedding_base import (
 from fastembed.sparse.utils.tokenizer import SimpleTokenizer
 from fastembed.common.model_description import SparseModelDescription, ModelSource
 
+
 supported_languages = [
     "arabic",
-    "azerbaijani",
-    "basque",
-    "bengali",
-    "catalan",
-    "chinese",
     "danish",
     "dutch",
     "english",
@@ -35,21 +31,15 @@ supported_languages = [
     "french",
     "german",
     "greek",
-    "hebrew",
-    "hinglish",
     "hungarian",
-    "indonesian",
     "italian",
-    "kazakh",
-    "nepali",
     "norwegian",
     "portuguese",
     "romanian",
     "russian",
-    "slovene",
     "spanish",
     "swedish",
-    "tajik",
+    "tamil",
     "turkish",
 ]
 
@@ -125,11 +115,12 @@ class Bm25(SparseTextEmbeddingBase):
         model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         self.token_max_length = token_max_length
@@ -170,6 +161,8 @@ class Bm25(SparseTextEmbeddingBase):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
     ) -> Iterable[SparseEmbedding]:
         is_small = False
 
@@ -198,6 +191,8 @@ class Bm25(SparseTextEmbeddingBase):
                 "language": self.language,
                 "token_max_length": self.token_max_length,
                 "disable_stemmer": self.disable_stemmer,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
             }
             pool = ParallelWorkerPool(
                 num_workers=parallel or 1,
@@ -236,6 +231,8 @@ class Bm25(SparseTextEmbeddingBase):
             documents=documents,
             batch_size=batch_size,
             parallel=parallel,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
         )
 
     def _stem(self, tokens: list[str]) -> list[str]:

fastembed/sparse/bm42.py

@@ -110,11 +110,12 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         self.invert_vocab: dict[int, str] = {}
@@ -217,7 +218,9 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
 
         return new_vector
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[SparseEmbedding]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         if output.input_ids is None:
             raise ValueError("input_ids must be provided for document post-processing")
 
@@ -299,6 +302,8 @@ class Bm42(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             cuda=self.cuda,
             device_ids=self.device_ids,
             alpha=self.alpha,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
         )
 
     @classmethod

fastembed/sparse/minicoil.py

@@ -0,0 +1,356 @@
+from pathlib import Path
+
+from typing import Any, Optional, Sequence, Iterable, Union, Type
+
+import numpy as np
+from numpy.typing import NDArray
+from py_rust_stemmers import SnowballStemmer
+from tokenizers import Tokenizer
+
+from fastembed.common.model_description import SparseModelDescription, ModelSource
+from fastembed.common.onnx_model import OnnxOutputContext
+from fastembed.common import OnnxProvider
+from fastembed.common.utils import define_cache_dir
+from fastembed.sparse.sparse_embedding_base import (
+    SparseEmbedding,
+    SparseTextEmbeddingBase,
+)
+from fastembed.sparse.utils.minicoil_encoder import Encoder
+from fastembed.sparse.utils.sparse_vectors_converter import SparseVectorConverter, WordEmbedding
+from fastembed.sparse.utils.vocab_resolver import VocabResolver, VocabTokenizer
+from fastembed.text.onnx_text_model import OnnxTextModel, TextEmbeddingWorker
+
+
+MINICOIL_MODEL_FILE = "minicoil.triplet.model.npy"
+MINICOIL_VOCAB_FILE = "minicoil.triplet.model.vocab"
+STOPWORDS_FILE = "stopwords.txt"
+
+
+supported_minicoil_models: list[SparseModelDescription] = [
+    SparseModelDescription(
+        model="Qdrant/minicoil-v1",
+        vocab_size=19125,
+        description="Sparse embedding model, that resolves semantic meaning of the words, "
+        "while keeping exact keyword match behavior. "
+        "Based on jinaai/jina-embeddings-v2-small-en-tokens",
+        license="apache-2.0",
+        size_in_GB=0.09,
+        sources=ModelSource(hf="Qdrant/minicoil-v1"),
+        model_file="onnx/model.onnx",
+        additional_files=[
+            STOPWORDS_FILE,
+            MINICOIL_MODEL_FILE,
+            MINICOIL_VOCAB_FILE,
+        ],
+        requires_idf=True,
+    ),
+]
+
+MODEL_TO_LANGUAGE = {
+    "Qdrant/minicoil-v1": "english",
+}
+
+
+class MiniCOIL(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
+    """
+        MiniCOIL is a sparse embedding model, that resolves semantic meaning of the words,
+        while keeping exact keyword match behavior.
+
+        Each vocabulary token is converted into 4d component of a sparse vector, which is then weighted by the token frequency in the corpus.
+        If the token is not found in the corpus, it is treated exactly like in BM25.
+    `
+        The model is based on `jinaai/jina-embeddings-v2-small-en-tokens`
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        cache_dir: Optional[str] = None,
+        threads: Optional[int] = None,
+        providers: Optional[Sequence[OnnxProvider]] = None,
+        k: float = 1.2,
+        b: float = 0.75,
+        avg_len: float = 150.0,
+        cuda: bool = False,
+        device_ids: Optional[list[int]] = None,
+        lazy_load: bool = False,
+        device_id: Optional[int] = None,
+        specific_model_path: Optional[str] = None,
+        **kwargs: Any,
+    ):
+        """
+        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.
+            providers (Optional[Sequence[OnnxProvider]], optional): The providers to use for onnxruntime.
+            k (float, optional): The k parameter in the BM25 formula. Defines the saturation of the term frequency.
+                I.e. defines how fast the moment when additional terms stop to increase the score. Defaults to 1.2.
+            b (float, optional): The b parameter in the BM25 formula. Defines the importance of the document length.
+                Defaults to 0.75.
+            avg_len (float, optional): The average length of the documents in the corpus. Defaults to 150.0.
+            cuda (bool, optional): Whether to use cuda for inference. Mutually exclusive with `providers`
+                Defaults to False.
+            device_ids (Optional[list[int]], optional): The list of device ids to use for data parallel processing in
+                workers. Should be used with `cuda=True`, mutually exclusive with `providers`. Defaults to None.
+            lazy_load (bool, optional): Whether to load the model during class initialization or on demand.
+                Should be set to True when using multiple-gpu and parallel encoding. Defaults to False.
+            device_id (Optional[int], optional): The device id to use for loading the model in the worker process.
+            specific_model_path (Optional[str], optional): The specific path to the onnx model dir if it should be imported from somewhere else
+
+        Raises:
+            ValueError: If the model_name is not in the format <org>/<model> e.g. BAAI/bge-base-en.
+        """
+
+        super().__init__(model_name, cache_dir, threads, **kwargs)
+        self.providers = providers
+        self.lazy_load = lazy_load
+        self.device_ids = device_ids
+        self.cuda = cuda
+        self.device_id = device_id
+        self.k = k
+        self.b = b
+        self.avg_len = avg_len
+
+        # Initialize class attributes
+        self.tokenizer: Optional[Tokenizer] = None
+        self.invert_vocab: dict[int, str] = {}
+        self.special_tokens: set[str] = set()
+        self.special_tokens_ids: set[int] = set()
+        self.stopwords: set[str] = set()
+        self.vocab_resolver: Optional[VocabResolver] = None
+        self.encoder: Optional[Encoder] = None
+        self.output_dim: Optional[int] = None
+        self.sparse_vector_converter: Optional[SparseVectorConverter] = None
+
+        self.model_description = self._get_model_description(model_name)
+        self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
+        self._model_dir = self.download_model(
+            self.model_description,
+            self.cache_dir,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+        )
+
+        if not self.lazy_load:
+            self.load_onnx_model()
+
+    def load_onnx_model(self) -> None:
+        self._load_onnx_model(
+            model_dir=self._model_dir,
+            model_file=self.model_description.model_file,
+            threads=self.threads,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_id=self.device_id,
+        )
+
+        assert self.tokenizer is not None
+
+        for token, idx in self.tokenizer.get_vocab().items():  # type: ignore[union-attr]
+            self.invert_vocab[idx] = token
+        self.special_tokens = set(self.special_token_to_id.keys())
+        self.special_tokens_ids = set(self.special_token_to_id.values())
+        self.stopwords = set(self._load_stopwords(self._model_dir))
+
+        stemmer = SnowballStemmer(MODEL_TO_LANGUAGE[self.model_name])
+
+        self.vocab_resolver = VocabResolver(
+            tokenizer=VocabTokenizer(self.tokenizer),
+            stopwords=self.stopwords,
+            stemmer=stemmer,
+        )
+        self.vocab_resolver.load_json_vocab(str(self._model_dir / MINICOIL_VOCAB_FILE))
+
+        weights = np.load(str(self._model_dir / MINICOIL_MODEL_FILE), mmap_mode="r")
+        self.encoder = Encoder(weights)
+        self.output_dim = self.encoder.output_dim
+
+        self.sparse_vector_converter = SparseVectorConverter(
+            stopwords=self.stopwords,
+            stemmer=stemmer,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+        )
+
+    def embed(
+        self,
+        documents: Union[str, Iterable[str]],
+        batch_size: int = 256,
+        parallel: Optional[int] = None,
+        **kwargs: Any,
+    ) -> 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,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+            is_query=False,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            **kwargs,
+        )
+
+    def query_embed(
+        self, query: Union[str, Iterable[str]], **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
+        """
+        Encode a list of queries into list of embeddings.
+        """
+        yield from self._embed_documents(
+            model_name=self.model_name,
+            cache_dir=str(self.cache_dir),
+            documents=query,
+            providers=self.providers,
+            cuda=self.cuda,
+            device_ids=self.device_ids,
+            k=self.k,
+            b=self.b,
+            avg_len=self.avg_len,
+            is_query=True,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
+            **kwargs,
+        )
+
+    @classmethod
+    def _load_stopwords(cls, model_dir: Path) -> list[str]:
+        stopwords_path = model_dir / STOPWORDS_FILE
+        if not stopwords_path.exists():
+            return []
+
+        with open(stopwords_path, "r") as f:
+            return f.read().splitlines()
+
+    @classmethod
+    def _list_supported_models(cls) -> list[SparseModelDescription]:
+        """Lists the supported models.
+
+        Returns:
+            list[SparseModelDescription]: A list of SparseModelDescription objects containing the model information.
+        """
+        return supported_minicoil_models
+
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, is_query: bool = False, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
+        if output.input_ids is None:
+            raise ValueError("input_ids must be provided for document post-processing")
+
+        assert self.vocab_resolver is not None
+        assert self.encoder is not None
+        assert self.sparse_vector_converter is not None
+
+        # Size: (batch_size, sequence_length, hidden_size)
+        embeddings = output.model_output
+        # Size: (batch_size, sequence_length)
+        assert output.attention_mask is not None
+        masks = output.attention_mask
+
+        vocab_size = self.vocab_resolver.vocab_size()
+        embedding_size = self.encoder.output_dim
+
+        # For each document we only select those embeddings that are not masked out
+
+        for i in range(embeddings.shape[0]):
+            # Size: (sequence_length, hidden_size)
+            token_embeddings = embeddings[i, masks[i] == 1]
+
+            # Size: (sequence_length)
+            token_ids: NDArray[np.int64] = output.input_ids[i, masks[i] == 1]
+
+            word_ids_array, counts, oov, forms = self.vocab_resolver.resolve_tokens(token_ids)
+
+            # Size: (1, words)
+            word_ids_array_expanded: NDArray[np.int64] = np.expand_dims(word_ids_array, axis=0)
+
+            # Size: (1, words, embedding_size)
+            token_embeddings_array: NDArray[np.float32] = np.expand_dims(token_embeddings, axis=0)
+
+            assert word_ids_array_expanded.shape[1] == token_embeddings_array.shape[1]
+
+            # Size of word_ids_mapping: (unique_words, 2) - [vocab_id, batch_id]
+            # Size of embeddings: (unique_words, embedding_size)
+            ids_mapping, minicoil_embeddings = self.encoder.forward(
+                word_ids_array_expanded, token_embeddings_array
+            )
+
+            # Size of counts: (unique_words)
+            words_ids: list[int] = ids_mapping[:, 0].tolist()  # type: ignore[assignment]
+
+            sentence_result: dict[str, WordEmbedding] = {}
+
+            words = [self.vocab_resolver.lookup_word(word_id) for word_id in words_ids]
+
+            for word, word_id, emb in zip(words, words_ids, minicoil_embeddings.tolist()):  # type: ignore[arg-type]
+                if word_id == 0:
+                    continue
+
+                sentence_result[word] = WordEmbedding(
+                    word=word,
+                    forms=forms[word],
+                    count=int(counts[word_id]),
+                    word_id=int(word_id),
+                    embedding=emb,  # type: ignore[arg-type]
+                )
+
+            for oov_word, count in oov.items():
+                # {
+                #     "word": oov_word,
+                #     "forms": [oov_word],
+                #     "count": int(count),
+                #     "word_id": -1,
+                #     "embedding": [1]
+                # }
+                sentence_result[oov_word] = WordEmbedding(
+                    word=oov_word, forms=[oov_word], count=int(count), word_id=-1, embedding=[1]
+                )
+
+            if not is_query:
+                yield self.sparse_vector_converter.embedding_to_vector(
+                    sentence_result, vocab_size=vocab_size, embedding_size=embedding_size
+                )
+            else:
+                yield self.sparse_vector_converter.embedding_to_vector_query(
+                    sentence_result, vocab_size=vocab_size, embedding_size=embedding_size
+                )
+
+    @classmethod
+    def _get_worker_class(cls) -> Type["MiniCoilTextEmbeddingWorker"]:
+        return MiniCoilTextEmbeddingWorker
+
+
+class MiniCoilTextEmbeddingWorker(TextEmbeddingWorker[SparseEmbedding]):
+    def init_embedding(self, model_name: str, cache_dir: str, **kwargs: Any) -> MiniCOIL:
+        return MiniCOIL(
+            model_name=model_name,
+            cache_dir=cache_dir,
+            threads=1,
+            **kwargs,
+        )

fastembed/sparse/sparse_text_embedding.py

@@ -4,6 +4,7 @@ from dataclasses import asdict
 from fastembed.common import OnnxProvider
 from fastembed.sparse.bm25 import Bm25
 from fastembed.sparse.bm42 import Bm42
+from fastembed.sparse.minicoil import MiniCOIL
 from fastembed.sparse.sparse_embedding_base import (
     SparseEmbedding,
     SparseTextEmbeddingBase,
@@ -14,7 +15,7 @@ from fastembed.common.model_description import SparseModelDescription
 
 
 class SparseTextEmbedding(SparseTextEmbeddingBase):
-    EMBEDDINGS_REGISTRY: list[Type[SparseTextEmbeddingBase]] = [SpladePP, Bm42, Bm25]
+    EMBEDDINGS_REGISTRY: list[Type[SparseTextEmbeddingBase]] = [SpladePP, Bm42, Bm25, MiniCOIL]
 
     @classmethod
     def list_supported_models(cls) -> list[dict[str, Any]]:
@@ -61,7 +62,7 @@ class SparseTextEmbedding(SparseTextEmbeddingBase):
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
-        if model_name == "prithvida/Splade_PP_en_v1":
+        if model_name.lower() == "prithvida/Splade_PP_en_v1".lower():
             warnings.warn(
                 "The right spelling is prithivida/Splade_PP_en_v1. "
                 "Support of this name will be removed soon, please fix the model_name",

fastembed/sparse/splade_pp.py

@@ -18,7 +18,7 @@ supported_splade_models: list[SparseModelDescription] = [
         description="Independent Implementation of SPLADE++ Model for English.",
         license="apache-2.0",
         size_in_GB=0.532,
-        sources=ModelSource(hf="Qdrant/SPLADE_PP_en_v1"),
+        sources=ModelSource(hf="Qdrant/Splade_PP_en_v1"),
         model_file="model.onnx",
     ),
     SparseModelDescription(
@@ -27,14 +27,16 @@ supported_splade_models: list[SparseModelDescription] = [
         description="Independent Implementation of SPLADE++ Model for English.",
         license="apache-2.0",
         size_in_GB=0.532,
-        sources=ModelSource(hf="Qdrant/SPLADE_PP_en_v1"),
+        sources=ModelSource(hf="Qdrant/Splade_PP_en_v1"),
         model_file="model.onnx",
     ),
 ]
 
 
 class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[SparseEmbedding]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[SparseEmbedding]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 
@@ -112,11 +114,12 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
 
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -163,6 +166,8 @@ class SpladePP(SparseTextEmbeddingBase, OnnxTextModel[SparseEmbedding]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 

fastembed/sparse/utils/minicoil_encoder.py

@@ -0,0 +1,146 @@
+"""
+Pure numpy implementation of encoder model for a single word.
+
+This model is not trainable, and should only be used for inference.
+"""
+
+import numpy as np
+from fastembed.common.types import NumpyArray
+
+
+class Encoder:
+    """
+    Encoder(768, 4, 10000)
+
+    Will look like this:
+
+
+                                         Per-word
+                                         Encoder Matrix
+     ┌─────────────────────┐
+     │ Token Embedding(768)├──────┐      (10k, 768, 4)
+     └─────────────────────┘      │         ┌─────────┐
+                                  │         │         │
+     ┌─────────────────────┐      │       ┌─┴───────┐ │
+     │                     │      │       │         │ │
+     └─────────────────────┘      │     ┌─┴───────┐ │ │      ┌─────────┐
+                                  └────►│         │ │ ├─────►│Tanh     │
+     ┌─────────────────────┐            │         │ │ │      └─────────┘
+     │                     │            │         │ ├─┘
+     └─────────────────────┘            │         ├─┘
+                                        │         │
+     ┌─────────────────────┐            └─────────┘
+     │                     │
+     └─────────────────────┘
+
+     Final linear transformation is accompanied by a non-linear activation function: Tanh.
+
+     Tanh is used to ensure that the output is in the range [-1, 1].
+     It would be easier to visually interpret the output of the model, assuming that each dimension
+     would need to encode a type of semantic cluster.
+    """
+
+    def __init__(
+        self,
+        weights: NumpyArray,
+    ):
+        self.weights = weights
+        self.vocab_size, self.input_dim, self.output_dim = weights.shape
+
+        self.encoder_weights: NumpyArray = weights
+
+        # Activation function
+        self.activation = np.tanh
+
+    @staticmethod
+    def convert_vocab_ids(vocab_ids: NumpyArray) -> NumpyArray:
+        """
+        Convert vocab_ids of shape (batch_size, seq_len) into (batch_size, seq_len, 2)
+        by appending batch_id alongside each vocab_id.
+        """
+        batch_size, seq_len = vocab_ids.shape
+        batch_ids = np.arange(batch_size, dtype=vocab_ids.dtype).reshape(batch_size, 1)
+        batch_ids = np.repeat(batch_ids, seq_len, axis=1)
+        # Stack vocab_ids and batch_ids along the last dimension
+        combined: NumpyArray = np.stack((vocab_ids, batch_ids), axis=2).astype(np.int32)
+        return combined
+
+    @classmethod
+    def avg_by_vocab_ids(
+        cls, vocab_ids: NumpyArray, embeddings: NumpyArray
+    ) -> tuple[NumpyArray, NumpyArray]:
+        """
+        Takes:
+            vocab_ids: (batch_size, seq_len) int array
+            embeddings: (batch_size, seq_len, input_dim) float array
+
+        Returns:
+            unique_flattened_vocab_ids: (total_unique, 2) array of [vocab_id, batch_id]
+            unique_flattened_embeddings: (total_unique, input_dim) averaged embeddings
+        """
+        input_dim = embeddings.shape[2]
+
+        # Flatten vocab_ids and embeddings
+        # flattened_vocab_ids: (batch_size*seq_len, 2)
+        flattened_vocab_ids = cls.convert_vocab_ids(vocab_ids).reshape(-1, 2)
+
+        # flattened_embeddings: (batch_size*seq_len, input_dim)
+        flattened_embeddings = embeddings.reshape(-1, input_dim)
+
+        # Find unique (vocab_id, batch_id) pairs
+        unique_flattened_vocab_ids, inverse_indices = np.unique(
+            flattened_vocab_ids, axis=0, return_inverse=True
+        )
+
+        # Prepare arrays to accumulate sums
+        unique_count = unique_flattened_vocab_ids.shape[0]
+        unique_flattened_embeddings = np.zeros((unique_count, input_dim), dtype=np.float32)
+        unique_flattened_count = np.zeros(unique_count, dtype=np.int32)
+
+        # Use np.add.at to accumulate sums based on inverse indices
+        np.add.at(unique_flattened_embeddings, inverse_indices, flattened_embeddings)
+        np.add.at(unique_flattened_count, inverse_indices, 1)
+
+        # Compute averages
+        unique_flattened_embeddings /= unique_flattened_count[:, None]
+
+        return unique_flattened_vocab_ids.astype(np.int32), unique_flattened_embeddings.astype(
+            np.float32
+        )
+
+    def forward(
+        self, vocab_ids: NumpyArray, embeddings: NumpyArray
+    ) -> tuple[NumpyArray, NumpyArray]:
+        """
+        Args:
+            vocab_ids: (batch_size, seq_len) int array
+            embeddings: (batch_size, seq_len, input_dim) float array
+
+        Returns:
+            unique_flattened_vocab_ids_and_batch_ids: (total_unique, 2)
+            unique_flattened_encoded: (total_unique, output_dim)
+        """
+        # Average embeddings for duplicate vocab_ids
+        unique_flattened_vocab_ids_and_batch_ids, unique_flattened_embeddings = (
+            self.avg_by_vocab_ids(vocab_ids, embeddings)
+        )
+
+        # Select the encoder weights for each unique vocab_id
+        unique_flattened_vocab_ids = unique_flattened_vocab_ids_and_batch_ids[:, 0].astype(
+            np.int32
+        )
+
+        # unique_encoder_weights: (total_unique, input_dim, output_dim)
+        unique_encoder_weights = self.encoder_weights[unique_flattened_vocab_ids]
+
+        # Compute linear transform: (total_unique, output_dim)
+        # Using Einstein summation for matrix multiplication:
+        # 'bi,bio->bo' means: for each "b" (batch element), multiply embeddings (b,i) by weights (b,i,o) -> (b,o)
+        unique_flattened_encoded = np.einsum(
+            "bi,bio->bo", unique_flattened_embeddings, unique_encoder_weights
+        )
+
+        # Apply Tanh activation and ensure float32 type
+        unique_flattened_encoded = self.activation(unique_flattened_encoded).astype(np.float32)
+
+        return unique_flattened_vocab_ids_and_batch_ids.astype(np.int32), unique_flattened_encoded

fastembed/sparse/utils/sparse_vectors_converter.py

@@ -0,0 +1,247 @@
+from typing import Dict, List, Set
+from py_rust_stemmers import SnowballStemmer
+from fastembed.common.utils import get_all_punctuation, remove_non_alphanumeric
+import mmh3
+import copy
+from dataclasses import dataclass
+
+import numpy as np
+
+from fastembed.sparse.sparse_embedding_base import SparseEmbedding
+
+GAP = 32000
+INT32_MAX = 2**31 - 1
+
+
+@dataclass
+class WordEmbedding:
+    word: str
+    forms: List[str]
+    count: int
+    word_id: int
+    embedding: List[float]
+
+
+class SparseVectorConverter:
+    def __init__(
+        self,
+        stopwords: Set[str],
+        stemmer: SnowballStemmer,
+        k: float = 1.2,
+        b: float = 0.75,
+        avg_len: float = 150.0,
+    ):
+        punctuation = set(get_all_punctuation())
+        special_tokens = {"[CLS]", "[SEP]", "[PAD]", "[UNK]", "[MASK]"}
+
+        self.stemmer = stemmer
+        self.unwanted_tokens = punctuation | special_tokens | stopwords
+
+        self.k = k
+        self.b = b
+        self.avg_len = avg_len
+
+    @classmethod
+    def unkn_word_token_id(
+        cls, word: str, shift: int
+    ) -> int:  # 2-3 words can collide in 1 index with this mapping, not considering mm3 collisions
+        token_hash = abs(mmh3.hash(word))
+
+        range_size = INT32_MAX - shift
+        remapped_hash = shift + (token_hash % range_size)
+
+        return remapped_hash
+
+    def bm25_tf(self, num_occurrences: int, sentence_len: int) -> float:
+        res = num_occurrences * (self.k + 1)
+        res /= num_occurrences + self.k * (1 - self.b + self.b * sentence_len / self.avg_len)
+        return res
+
+    @classmethod
+    def normalize_vector(cls, vector: List[float]) -> List[float]:
+        norm = sum([x**2 for x in vector]) ** 0.5
+        if norm < 1e-8:
+            return vector
+        return [x / norm for x in vector]
+
+    def clean_words(
+        self, sentence_embedding: Dict[str, WordEmbedding], token_max_length: int = 40
+    ) -> Dict[str, WordEmbedding]:
+        """
+        Clean miniCOIL-produced sentence_embedding, as unknown to the miniCOIL's stemmer tokens should fully resemble
+        our BM25 token representation.
+
+        sentence_embedding = {"9°": {"word": "9°", "word_id": -1, "count": 2, "embedding": [1], "forms": ["9°"]},
+                "9": {"word": "9", "word_id": -1, "count": 2, "embedding": [1], "forms": ["9"]},
+                "bat": {"word": "bat", "word_id": 2, "count": 3, "embedding": [0.2, 0.1, -0.2, -0.2], "forms": ["bats", "bat"]},
+                "9°9": {"word": "9°9", "word_id": -1, "count": 1, "embedding": [1], "forms": ["9°9"]},
+                "screech": {"word": "screech", "word_id": -1, "count": 1, "embedding": [1], "forms": ["screech"]},
+                "screeched": {"word": "screeched", "word_id": -1, "count": 1, "embedding": [1], "forms": ["screeched"]}
+                }
+        cleaned_embedding_ground_truth = {
+                "9": {"word": "9", "word_id": -1, "count": 6, "embedding": [1], "forms": ["9°", "9", "9°9", "9°9"]},
+                "bat": {"word": "bat", "word_id": 2, "count": 3, "embedding": [0.2, 0.1, -0.2, -0.2], "forms": ["bats", "bat"]},
+                "screech": {"word": "screech", "word_id": -1, "count": 2, "embedding": [1], "forms": ["screech", "screeched"]}
+                }
+        """
+
+        new_sentence_embedding: Dict[str, WordEmbedding] = {}
+
+        for word, embedding in sentence_embedding.items():
+            # embedding = {
+            #     "word": "vector",
+            #     "forms": ["vector", "vectors"],
+            #     "count": 2,
+            #     "word_id": 1231,
+            #     "embedding": [0.1, 0.2, 0.3, 0.4]
+            # }
+            if embedding.word_id > 0:
+                # Known word, no need to clean
+                new_sentence_embedding[word] = embedding
+            else:
+                # Unknown word
+                if word in self.unwanted_tokens:
+                    continue
+
+                # Example complex word split:
+                # word = `word^vec`
+                word_cleaned = remove_non_alphanumeric(word).strip()
+                # word_cleaned = `word vec`
+
+                if len(word_cleaned) > 0:
+                    # Subwords: ['word', 'vec']
+                    for subword in word_cleaned.split():
+                        stemmed_subword: str = self.stemmer.stem_word(subword)
+                        if (
+                            len(stemmed_subword) <= token_max_length
+                            and stemmed_subword not in self.unwanted_tokens
+                        ):
+                            if stemmed_subword not in new_sentence_embedding:
+                                new_sentence_embedding[stemmed_subword] = copy.deepcopy(embedding)
+                                new_sentence_embedding[stemmed_subword].word = stemmed_subword
+                            else:
+                                new_sentence_embedding[stemmed_subword].count += embedding.count
+                                new_sentence_embedding[stemmed_subword].forms += embedding.forms
+
+        return new_sentence_embedding
+
+    def embedding_to_vector(
+        self,
+        sentence_embedding: Dict[str, WordEmbedding],
+        embedding_size: int,
+        vocab_size: int,
+    ) -> SparseEmbedding:
+        """
+        Convert miniCOIL sentence embedding to Qdrant sparse vector
+
+        Example input:
+
+        ```
+        {
+            "vector": WordEmbedding({ // Vocabulary word, encoded with miniCOIL normally
+                "word": "vector",
+                "forms": ["vector", "vectors"],
+                "count": 2,
+                "word_id": 1231,
+                "embedding": [0.1, 0.2, 0.3, 0.4]
+            }),
+            "axiotic": WordEmbedding({ // Out-of-vocabulary word, fallback to BM25
+                "word": "axiotic",
+                "forms": ["axiotics"],
+                "count": 1,
+                "word_id": -1,
+            })
+        }
+        ```
+
+        """
+
+        indices: List[int] = []
+        values: List[float] = []
+        
+        # Example:
+        # vocab_size = 10000
+        # embedding_size = 4
+        # GAP = 32000
+        # 
+        # We want to start random words section from the bucket, that is guaranteed to not
+        # include any vocab words.
+        # We need (vocab_size * embedding_size) slots for vocab words.
+        # Therefore we need (vocab_size * embedding_size) // GAP + 1 buckets for vocab words.
+        # Therefore, we can start random words from bucket (vocab_size * embedding_size) // GAP + 1 + 1
+
+        # ID at which the scope of OOV words starts
+        unknown_words_shift = (
+            (vocab_size * embedding_size) // GAP + 2
+        ) * GAP
+        sentence_embedding_cleaned = self.clean_words(sentence_embedding)
+
+        # Calculate sentence length after cleaning
+        sentence_len = 0
+        for embedding in sentence_embedding_cleaned.values():
+            sentence_len += embedding.count
+
+        for embedding in sentence_embedding_cleaned.values():
+            word_id = embedding.word_id
+            num_occurrences = embedding.count
+            tf = self.bm25_tf(num_occurrences, sentence_len)
+            if (
+                word_id > 0
+            ):  # miniCOIL starts with ID 1, we generally won't have word_id == 0 (UNK), as we don't add
+                # these words to sentence_embedding
+                embedding_values = embedding.embedding
+                normalized_embedding = self.normalize_vector(embedding_values)
+
+                for val_id, value in enumerate(normalized_embedding):
+                    indices.append(
+                        word_id * embedding_size + val_id
+                    )  # since miniCOIL IDs start with 1
+                    values.append(value * tf)
+            else:
+                indices.append(self.unkn_word_token_id(embedding.word, unknown_words_shift))
+                values.append(tf)
+
+        return SparseEmbedding(
+            indices=np.array(indices, dtype=np.int32),
+            values=np.array(values, dtype=np.float32),
+        )
+
+    def embedding_to_vector_query(
+        self,
+        sentence_embedding: Dict[str, WordEmbedding],
+        embedding_size: int,
+        vocab_size: int,
+    ) -> SparseEmbedding:
+        """
+        Same as `embedding_to_vector`, but no TF
+        """
+
+        indices: List[int] = []
+        values: List[float] = []
+
+        # ID at which the scope of OOV words starts
+        unknown_words_shift = ((vocab_size * embedding_size) // GAP + 2) * GAP
+
+        sentence_embedding_cleaned = self.clean_words(sentence_embedding)
+
+        for embedding in sentence_embedding_cleaned.values():
+            word_id = embedding.word_id
+            tf = 1.0
+
+            if word_id >= 0:  # miniCOIL starts with ID 1
+                embedding_values = embedding.embedding
+                normalized_embedding = self.normalize_vector(embedding_values)
+
+                for val_id, value in enumerate(normalized_embedding):
+                    indices.append(
+                        word_id * embedding_size + val_id
+                    )  # since miniCOIL IDs start with 1
+                    values.append(value * tf)
+            else:
+                indices.append(self.unkn_word_token_id(embedding.word, unknown_words_shift))
+                values.append(tf)
+
+        return SparseEmbedding(
+            indices=np.array(indices, dtype=np.int32),
+            values=np.array(values, dtype=np.float32),
+        )

fastembed/sparse/utils/vocab_resolver.py

@@ -0,0 +1,202 @@
+from collections import defaultdict
+from typing import Iterable
+
+from py_rust_stemmers import SnowballStemmer
+import numpy as np
+from tokenizers import Tokenizer
+from numpy.typing import NDArray
+
+from fastembed.common.types import NumpyArray
+
+
+class VocabTokenizerBase:
+    def tokenize(self, sentence: str) -> NumpyArray:
+        raise NotImplementedError()
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        raise NotImplementedError()
+
+
+class VocabTokenizer(VocabTokenizerBase):
+    def __init__(self, tokenizer: Tokenizer):
+        self.tokenizer = tokenizer
+
+    def tokenize(self, sentence: str) -> NumpyArray:
+        return np.array(self.tokenizer.encode(sentence).ids)
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        return [self.tokenizer.id_to_token(token_id) for token_id in token_ids]
+
+
+class VocabResolver:
+    def __init__(self, tokenizer: VocabTokenizerBase, stopwords: set[str], stemmer: SnowballStemmer):
+        # Word to id mapping
+        self.vocab: dict[str, int] = {}
+        # Id to word mapping
+        self.words: list[str] = []
+        # Lemma to word mapping
+        self.stem_mapping: dict[str, str] = {}
+        self.tokenizer: VocabTokenizerBase = tokenizer
+        self.stemmer = stemmer
+        self.stopwords: set[str] = stopwords
+
+    def tokenize(self, sentence: str) -> NumpyArray:
+        return self.tokenizer.tokenize(sentence)
+
+    def lookup_word(self, word_id: int) -> str:
+        if word_id == 0:
+            return "UNK"
+        return self.words[word_id - 1]
+
+    def convert_ids_to_tokens(self, token_ids: NumpyArray) -> list[str]:
+        return self.tokenizer.convert_ids_to_tokens(token_ids)
+
+    def vocab_size(self) -> int:
+        # We need +1 for UNK token
+        return len(self.vocab) + 1
+
+    def save_vocab(self, path: str) -> None:
+        with open(path, "w") as f:
+            for word in self.words:
+                f.write(word + "\n")
+
+    def save_json_vocab(self, path: str) -> None:
+        import json
+
+        with open(path, "w") as f:
+            json.dump({"vocab": self.words, "stem_mapping": self.stem_mapping}, f, indent=2)
+
+    def load_json_vocab(self, path: str) -> None:
+        import json
+
+        with open(path, "r") as f:
+            data = json.load(f)
+            self.words = data["vocab"]
+            self.vocab = {word: idx + 1 for idx, word in enumerate(self.words)}
+            self.stem_mapping = data["stem_mapping"]
+
+    def add_word(self, word: str) -> None:
+        if word not in self.vocab:
+            self.vocab[word] = len(self.vocab) + 1
+            self.words.append(word)
+            stem = self.stemmer.stem_word(word)
+            if stem not in self.stem_mapping:
+                self.stem_mapping[stem] = word
+            else:
+                existing_word = self.stem_mapping[stem]
+                if len(existing_word) > len(word):
+                    # Prefer shorter words for the same stem
+                    # Example: "swim" is preferred over "swimming"
+                    self.stem_mapping[stem] = word
+
+    def load_vocab(self, path: str) -> None:
+        with open(path, "r") as f:
+            for line in f:
+                self.add_word(line.strip())
+
+    @classmethod
+    def _reconstruct_bpe(
+        cls, bpe_tokens: Iterable[tuple[int, str]]
+    ) -> list[tuple[str, list[int]]]:
+        result: list[tuple[str, list[int]]] = []
+        acc: str = ""
+        acc_idx: list[int] = []
+
+        continuing_subword_prefix = "##"
+        continuing_subword_prefix_len = len(continuing_subword_prefix)
+
+        for idx, token in bpe_tokens:
+            if token.startswith(continuing_subword_prefix):
+                acc += token[continuing_subword_prefix_len:]
+                acc_idx.append(idx)
+            else:
+                if acc:
+                    result.append((acc, acc_idx))
+                    acc_idx = []
+                acc = token
+                acc_idx.append(idx)
+
+        if acc:
+            result.append((acc, acc_idx))
+        return result
+
+    def resolve_tokens(
+        self, token_ids: NDArray[np.int64]
+    ) -> tuple[NDArray[np.int64], dict[int, int], dict[str, int], dict[str, list[str]]]:
+        """
+        Mark known tokens (including composed tokens) with vocab ids.
+
+        Args:
+            token_ids: (seq_len) - list of ids of tokens
+                Example:
+                    [
+                        101,  3897, 19332, 12718, 23348,
+                        1010,  1996,  7151,  2296, 4845,
+                        2359,  2005,  4234,  1010,  4332,
+                        2871,  3191,  2062, 102
+                    ]
+
+            returns:
+                - token_ids with vocab ids
+                    [
+                        0,  151, 151, 0, 0,
+                        912,  0,  0,  0, 332,
+                        332,  332,  0,  7121,  191,
+                        0,  0,  332, 0
+                    ]
+                - counts of each token
+                    {
+                        151: 1,
+                        332: 3,
+                        7121: 1,
+                        191: 1,
+                        912: 1
+                    }
+                - oov counts of each token
+                    {
+                        "the": 1,
+                        "a": 1,
+                        "[CLS]": 1,
+                        "[SEP]": 1,
+                        ...
+                    }
+                - forms of each token
+                    {
+                        "hello": ["hello"],
+                        "world": ["worlds", "world", "worlding"],
+                    }
+
+        """
+        tokens = self.convert_ids_to_tokens(token_ids)
+        tokens_mapping = self._reconstruct_bpe(enumerate(tokens))
+
+        counts: dict[int, int] = defaultdict(int)
+        oov_count: dict[str, int] = defaultdict(int)
+
+        forms: dict[str, list[str]] = defaultdict(list)
+
+        for token, mapped_token_ids in tokens_mapping:
+            vocab_id = 0
+            if token in self.stopwords:
+                vocab_id = 0
+            elif token in self.vocab:
+                vocab_id = self.vocab[token]
+                forms[token].append(token)
+            elif token in self.stem_mapping:
+                vocab_id = self.vocab[self.stem_mapping[token]]
+                forms[self.stem_mapping[token]].append(token)
+            else:
+                stem = self.stemmer.stem_word(token)
+                if stem in self.stem_mapping:
+                    vocab_id = self.vocab[self.stem_mapping[stem]]
+                    forms[self.stem_mapping[stem]].append(token)
+
+            for token_id in mapped_token_ids:
+                token_ids[token_id] = vocab_id
+
+            if vocab_id == 0:
+                oov_count[token] += 1
+            else:
+                counts[vocab_id] += 1
+        return token_ids, counts, oov_count, forms
+

fastembed/text/clip_embedding.py

@@ -35,7 +35,9 @@ class CLIPOnnxEmbedding(OnnxTextEmbedding):
         """
         return supported_clip_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         return output.model_output
 
 

fastembed/text/custom_text_embedding.py

@@ -58,7 +58,9 @@ class CustomTextEmbedding(OnnxTextEmbedding):
     def _list_supported_models(cls) -> list[DenseModelDescription]:
         return cls.SUPPORTED_MODELS
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         return self._normalize(self._pool(output.model_output, output.attention_mask))
 
     def _pool(
@@ -75,6 +77,11 @@ class CustomTextEmbedding(OnnxTextEmbedding):
         if self._pooling == PoolingType.DISABLED:
             return embeddings
 
+        raise ValueError(
+            f"Unsupported pooling type {self._pooling}. "
+            f"Supported types are: {PoolingType.CLS}, {PoolingType.MEAN}, {PoolingType.DISABLED}."
+        )
+
     def _normalize(self, embeddings: NumpyArray) -> NumpyArray:
         return normalize(embeddings) if self._normalization else embeddings
 

fastembed/text/multitask_embedding.py

@@ -3,6 +3,7 @@ from typing import Any, Type, Iterable, Union, Optional
 
 import numpy as np
 
+from fastembed.common.onnx_model import OnnxOutputContext
 from fastembed.common.types import NumpyArray
 from fastembed.text.pooled_normalized_embedding import PooledNormalizedEmbedding
 from fastembed.text.onnx_embedding import OnnxTextEmbeddingWorker
@@ -44,9 +45,11 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
     PASSAGE_TASK = Task.RETRIEVAL_PASSAGE
     QUERY_TASK = Task.RETRIEVAL_QUERY
 
-    def __init__(self, *args: Any, **kwargs: Any):
+    def __init__(self, *args: Any, task_id: Optional[int] = None, **kwargs: Any):
         super().__init__(*args, **kwargs)
-        self.current_task_id: Union[Task, int] = self.PASSAGE_TASK
+        self.default_task_id: Union[Task, int] = (
+            task_id if task_id is not None else self.PASSAGE_TASK
+        )
 
     @classmethod
     def _get_worker_class(cls) -> Type[OnnxTextEmbeddingWorker]:
@@ -57,9 +60,14 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
         return supported_multitask_models
 
     def _preprocess_onnx_input(
-        self, onnx_input: dict[str, NumpyArray], **kwargs: Any
+        self,
+        onnx_input: dict[str, NumpyArray],
+        task_id: Optional[Union[int, Task]] = None,
+        **kwargs: Any,
     ) -> dict[str, NumpyArray]:
-        onnx_input["task_id"] = np.array(self.current_task_id, dtype=np.int64)
+        if task_id is None:
+            raise ValueError(f"task_id must be provided for JinaEmbeddingV3, got <{task_id}>")
+        onnx_input["task_id"] = np.array(task_id, dtype=np.int64)
         return onnx_input
 
     def embed(
@@ -67,20 +75,19 @@ class JinaEmbeddingV3(PooledNormalizedEmbedding):
         documents: Union[str, Iterable[str]],
         batch_size: int = 256,
         parallel: Optional[int] = None,
-        task_id: int = PASSAGE_TASK,
+        task_id: Optional[int] = None,
         **kwargs: Any,
     ) -> Iterable[NumpyArray]:
-        self.current_task_id = task_id
-        kwargs["task_id"] = task_id
-        yield from super().embed(documents, batch_size, parallel, **kwargs)
+        task_id = (
+            task_id if task_id is not None else self.default_task_id
+        )  # required for multiprocessing
+        yield from super().embed(documents, batch_size, parallel, task_id=task_id, **kwargs)
 
     def query_embed(self, query: Union[str, Iterable[str]], **kwargs: Any) -> Iterable[NumpyArray]:
-        self.current_task_id = self.QUERY_TASK
-        yield from super().embed(query, **kwargs)
+        yield from super().embed(query, task_id=self.QUERY_TASK, **kwargs)
 
     def passage_embed(self, texts: Iterable[str], **kwargs: Any) -> Iterable[NumpyArray]:
-        self.current_task_id = self.PASSAGE_TASK
-        yield from super().embed(texts, **kwargs)
+        yield from super().embed(texts, task_id=self.PASSAGE_TASK, **kwargs)
 
 
 class JinaEmbeddingV3Worker(OnnxTextEmbeddingWorker):
@@ -90,11 +97,15 @@ class JinaEmbeddingV3Worker(OnnxTextEmbeddingWorker):
         cache_dir: str,
         **kwargs: Any,
     ) -> JinaEmbeddingV3:
-        model = JinaEmbeddingV3(
+        return JinaEmbeddingV3(
             model_name=model_name,
             cache_dir=cache_dir,
             threads=1,
             **kwargs,
         )
-        model.current_task_id = kwargs["task_id"]
-        return model
+
+    def process(self, items: Iterable[tuple[int, Any]]) -> Iterable[tuple[int, OnnxOutputContext]]:
+        self.model: JinaEmbeddingV3  # mypy complaints `self.model` does not have `default_task_id`
+        for idx, batch in items:
+            onnx_output = self.model.onnx_embed(batch, task_id=self.model.default_task_id)
+            yield idx, onnx_output

fastembed/text/onnx_embedding.py

@@ -247,11 +247,12 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
 
         self.model_description = self._get_model_description(model_name)
         self.cache_dir = str(define_cache_dir(cache_dir))
+        self._specific_model_path = specific_model_path
         self._model_dir = self.download_model(
             self.model_description,
             self.cache_dir,
             local_files_only=self._local_files_only,
-            specific_model_path=specific_model_path,
+            specific_model_path=self._specific_model_path,
         )
 
         if not self.lazy_load:
@@ -288,6 +289,8 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
             providers=self.providers,
             cuda=self.cuda,
             device_ids=self.device_ids,
+            local_files_only=self._local_files_only,
+            specific_model_path=self._specific_model_path,
             **kwargs,
         )
 
@@ -303,7 +306,9 @@ class OnnxTextEmbedding(TextEmbeddingBase, OnnxTextModel[NumpyArray]):
         """
         return onnx_input
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         embeddings = output.model_output
 
         if embeddings.ndim == 3:  # (batch_size, seq_len, embedding_dim)

fastembed/text/onnx_text_model.py

@@ -21,7 +21,16 @@ class OnnxTextModel(OnnxModel[T]):
     def _get_worker_class(cls) -> Type["TextEmbeddingWorker[T]"]:
         raise NotImplementedError("Subclasses must implement this method")
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[T]:
+    def _post_process_onnx_output(self, output: OnnxOutputContext, **kwargs: Any) -> Iterable[T]:
+        """Post-process the ONNX model output to convert it into a usable format.
+
+        Args:
+            output (OnnxOutputContext): The raw output from the ONNX model.
+            **kwargs: Additional keyword arguments that may be needed by specific implementations.
+
+        Returns:
+            Iterable[T]: Post-processed output as an iterable of type T.
+        """
         raise NotImplementedError("Subclasses must implement this method")
 
     def __init__(self) -> None:
@@ -99,6 +108,8 @@ class OnnxTextModel(OnnxModel[T]):
         providers: Optional[Sequence[OnnxProvider]] = None,
         cuda: bool = False,
         device_ids: Optional[list[int]] = None,
+        local_files_only: bool = False,
+        specific_model_path: Optional[str] = None,
         **kwargs: Any,
     ) -> Iterable[T]:
         is_small = False
@@ -115,7 +126,9 @@ class OnnxTextModel(OnnxModel[T]):
             if not hasattr(self, "model") or self.model is None:
                 self.load_onnx_model()
             for batch in iter_batch(documents, batch_size):
-                yield from self._post_process_onnx_output(self.onnx_embed(batch))
+                yield from self._post_process_onnx_output(
+                    self.onnx_embed(batch, **kwargs), **kwargs
+                )
         else:
             if parallel == 0:
                 parallel = os.cpu_count()
@@ -125,6 +138,8 @@ class OnnxTextModel(OnnxModel[T]):
                 "model_name": model_name,
                 "cache_dir": cache_dir,
                 "providers": providers,
+                "local_files_only": local_files_only,
+                "specific_model_path": specific_model_path,
                 **kwargs,
             }
 
@@ -136,7 +151,7 @@ class OnnxTextModel(OnnxModel[T]):
                 start_method=start_method,
             )
             for batch in pool.ordered_map(iter_batch(documents, batch_size), **params):
-                yield from self._post_process_onnx_output(batch)  # type: ignore
+                yield from self._post_process_onnx_output(batch, **kwargs)  # type: ignore
 
 
 class TextEmbeddingWorker(EmbeddingWorker[T]):

fastembed/text/pooled_embedding.py

@@ -110,7 +110,9 @@ class PooledEmbedding(OnnxTextEmbedding):
         """
         return supported_pooled_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 

fastembed/text/pooled_normalized_embedding.py

@@ -138,7 +138,9 @@ class PooledNormalizedEmbedding(PooledEmbedding):
         """
         return supported_pooled_normalized_models
 
-    def _post_process_onnx_output(self, output: OnnxOutputContext) -> Iterable[NumpyArray]:
+    def _post_process_onnx_output(
+        self, output: OnnxOutputContext, **kwargs: Any
+    ) -> Iterable[NumpyArray]:
         if output.attention_mask is None:
             raise ValueError("attention_mask must be provided for document post-processing")
 

fastembed/text/text_embedding.py

@@ -55,7 +55,7 @@ class TextEmbedding(TextEmbeddingBase):
     ) -> None:
         registered_models = cls._list_supported_models()
         for registered_model in registered_models:
-            if model == registered_model.model:
+            if model.lower() == registered_model.model.lower():
                 raise ValueError(
                     f"Model {model} is already registered in TextEmbedding, if you still want to add this model, "
                     f"please use another model name"
@@ -88,18 +88,18 @@ class TextEmbedding(TextEmbeddingBase):
         **kwargs: Any,
     ):
         super().__init__(model_name, cache_dir, threads, **kwargs)
-        if model_name == "nomic-ai/nomic-embed-text-v1.5-Q":
+        if model_name.lower() == "nomic-ai/nomic-embed-text-v1.5-Q".lower():
             warnings.warn(
                 "The model 'nomic-ai/nomic-embed-text-v1.5-Q' has been updated on HuggingFace. Please review "
                 "the latest documentation on HF and release notes to ensure compatibility with your workflow. ",
                 UserWarning,
                 stacklevel=2,
             )
-        if model_name in {
-            "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
-            "thenlper/gte-large",
-            "intfloat/multilingual-e5-large",
-            "sentence-transformers/paraphrase-multilingual-mpnet-base-v2",
+        if model_name.lower() in {
+            "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2".lower(),
+            "thenlper/gte-large".lower(),
+            "intfloat/multilingual-e5-large".lower(),
+            "sentence-transformers/paraphrase-multilingual-mpnet-base-v2".lower(),
         }:
             warnings.warn(
                 f"The model {model_name} now uses mean pooling instead of CLS embedding. "
@@ -128,6 +128,40 @@ class TextEmbedding(TextEmbeddingBase):
             "Please check the supported models using `TextEmbedding.list_supported_models()`"
         )
 
+    @property
+    def embedding_size(self) -> int:
+        """Get the embedding size of the current model"""
+        if self._embedding_size is None:
+            self._embedding_size = self.get_embedding_size(self.model_name)
+        return self._embedding_size
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Get the embedding size of the passed model
+
+        Args:
+            model_name (str): The name of the model to get embedding size for.
+
+        Returns:
+            int: The size of the embedding.
+
+        Raises:
+            ValueError: If the model name is not found in the supported models.
+        """
+        descriptions = cls._list_supported_models()
+        embedding_size: Optional[int] = None
+        for description in descriptions:
+            if description.model.lower() == model_name.lower():
+                embedding_size = description.dim
+                break
+        if embedding_size is None:
+            model_names = [description.model for description in descriptions]
+            raise ValueError(
+                f"Embedding size for model {model_name} was None. "
+                f"Available model names: {model_names}"
+            )
+        return embedding_size
+
     def embed(
         self,
         documents: Union[str, Iterable[str]],

fastembed/text/text_embedding_base.py

@@ -17,6 +17,7 @@ class TextEmbeddingBase(ModelManagement[DenseModelDescription]):
         self.cache_dir = cache_dir
         self.threads = threads
         self._local_files_only = kwargs.pop("local_files_only", False)
+        self._embedding_size: Optional[int] = None
 
     def embed(
         self,
@@ -58,3 +59,13 @@ class TextEmbeddingBase(ModelManagement[DenseModelDescription]):
             yield from self.embed([query], **kwargs)
         else:
             yield from self.embed(query, **kwargs)
+
+    @classmethod
+    def get_embedding_size(cls, model_name: str) -> int:
+        """Returns embedding size of the passed model."""
+        raise NotImplementedError("Subclasses must implement this method")
+
+    @property
+    def embedding_size(self) -> int:
+        """Returns embedding size for the current model"""
+        raise NotImplementedError("Subclasses must implement this method")

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
-name = "fastembed"
-version = "0.6.0"
+name = "fastembed-gpu"
+version = "0.7.1"
 description = "Fast, light, accurate library built for retrieval embedding generation"
 authors = ["Qdrant Team <info@qdrant.tech>", "NirantK <nirant.bits@gmail.com>"]
 license = "Apache License"
@@ -18,9 +18,9 @@ numpy = [
     { version = ">=2.1.0", python = ">=3.13" },
     { version = ">=1.21,<2.1.0", python = "<3.10" },
 ]
-onnxruntime = [
-    { version = ">1.20.0", python = ">=3.13" },
+onnxruntime-gpu = [
     { version = ">=1.17.0,<1.20.0", python = "<3.10" },
+    { version = ">1.20.0", python = ">=3.13" },
     { version = ">=1.17.0,!=1.20.0", python = ">=3.10,<3.13" },
 ]
 tqdm = "^4.66"

tests/test_image_onnx_embeddings.py

@@ -127,3 +127,21 @@ def test_lazy_load(model_name: str) -> None:
     assert hasattr(model.model, "model")
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size() -> None:
+    assert ImageEmbedding.get_embedding_size(model_name="Qdrant/clip-ViT-B-32-vision") == 512
+    assert ImageEmbedding.get_embedding_size(model_name="Qdrant/clip-vit-b-32-vision") == 512
+
+
+def test_embedding_size() -> None:
+    is_ci = os.getenv("CI")
+    model_name = "Qdrant/clip-ViT-B-32-vision"
+    model = ImageEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 512
+
+    model_name = "Qdrant/clip-vit-b-32-vision"
+    model = ImageEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 512
+    if is_ci:
+        delete_model_cache(model.model._model_dir)

tests/test_late_interaction_embeddings.py

@@ -254,3 +254,24 @@ def test_lazy_load(model_name: str):
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size():
+    model_name = "answerdotai/answerai-colbert-small-v1"
+    assert LateInteractionTextEmbedding.get_embedding_size(model_name) == 96
+
+    model_name = "answerdotai/answerai-ColBERT-small-v1"
+    assert LateInteractionTextEmbedding.get_embedding_size(model_name) == 96
+
+
+def test_embedding_size():
+    is_ci = os.getenv("CI")
+    model_name = "answerdotai/answerai-colbert-small-v1"
+    model = LateInteractionTextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 96
+
+    model_name = "answerdotai/answerai-ColBERT-small-v1"
+    model = LateInteractionTextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 96
+    if is_ci:
+        delete_model_cache(model.model._model_dir)

tests/test_late_interaction_multimodal.py

@@ -81,3 +81,23 @@ def test_single_embedding_query():
         result = next(iter(model.embed_text(queries)))
         token_num, abridged_dim = expected_result.shape
         assert np.allclose(result[:token_num, :abridged_dim], expected_result, atol=2e-3)
+
+
+def test_get_embedding_size():
+    model_name = "Qdrant/colpali-v1.3-fp16"
+    assert LateInteractionMultimodalEmbedding.get_embedding_size(model_name) == 128
+
+    model_name = "Qdrant/ColPali-v1.3-fp16"
+    assert LateInteractionMultimodalEmbedding.get_embedding_size(model_name) == 128
+
+
+def test_embedding_size():
+    if os.getenv("CI"):
+        pytest.skip("Colpali is too large to test in CI")
+    model_name = "Qdrant/colpali-v1.3-fp16"
+    model = LateInteractionMultimodalEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 128
+
+    model_name = "Qdrant/ColPali-v1.3-fp16"
+    model = LateInteractionMultimodalEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 128

tests/test_sparse_embeddings.py

@@ -43,13 +43,46 @@ CANONICAL_COLUMN_VALUES = {
             2.1904349327087402,
             1.0531445741653442,
         ],
-    }
+    },
+    "Qdrant/minicoil-v1": {
+        "indices": [80, 81, 82, 83, 6664, 6665, 6666, 6667],
+        "values": [
+            0.52634597,
+            0.8711344,
+            1.2264385,
+            0.52123857,
+            0.974713,
+            -0.97803956,
+            -0.94312465,
+            -0.12508166,
+        ],
+    },
 }
 
+CANONICAL_QUERY_VALUES = {
+    "Qdrant/minicoil-v1": {
+        "indices": [80, 81, 82, 83, 6664, 6665, 6666, 6667],
+        "values": [
+            0.31389374,
+            0.5195128,
+            0.7314033,
+            0.3108479,
+            0.5812834,
+            -0.5832673,
+            -0.5624452,
+            -0.0745942,
+        ],
+    },
+}
+
+
 docs = ["Hello World"]
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
+@pytest.mark.parametrize(
+    "model_name",
+    ["prithivida/Splade_PP_en_v1", "Qdrant/minicoil-v1"],
+)
 def test_batch_embedding(model_name: str) -> None:
     is_ci = os.getenv("CI")
     docs_to_embed = docs * 10
@@ -65,7 +98,7 @@ def test_batch_embedding(model_name: str) -> None:
         delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
+@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1", "Qdrant/minicoil-v1"])
 def test_single_embedding(model_name: str) -> None:
     is_ci = os.getenv("CI")
     is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
@@ -84,16 +117,23 @@ def test_single_embedding(model_name: str) -> None:
         passage_result = next(iter(model.embed(docs, batch_size=6)))
         query_result = next(iter(model.query_embed(docs)))
         expected_result = CANONICAL_COLUMN_VALUES[model_name]
-        for result in [passage_result, query_result]:
-            assert result.indices.tolist() == expected_result["indices"]
+        expected_query_result = CANONICAL_QUERY_VALUES.get(model_name, expected_result)
+        assert passage_result.indices.tolist() == expected_result["indices"]
+        for i, value in enumerate(passage_result.values):
+            assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
+
+        assert query_result.indices.tolist() == expected_query_result["indices"]
+        for i, value in enumerate(query_result.values):
+            assert pytest.approx(value, abs=0.001) == expected_query_result["values"][i]
 
-            for i, value in enumerate(result.values):
-                assert pytest.approx(value, abs=0.001) == expected_result["values"][i]
         if is_ci:
             delete_model_cache(model.model._model_dir)
 
 
-@pytest.mark.parametrize("model_name", ["prithivida/Splade_PP_en_v1"])
+@pytest.mark.parametrize(
+    "model_name",
+    ["prithivida/Splade_PP_en_v1", "Qdrant/minicoil-v1"],
+)
 def test_parallel_processing(model_name: str) -> None:
     is_ci = os.getenv("CI")
     model = SparseTextEmbedding(model_name=model_name)

tests/test_text_multitask_embeddings.py

@@ -109,9 +109,25 @@ def test_single_embedding():
 
             canonical_vector = task["vectors"]
             assert np.allclose(
-                embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+                embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
             ), model_desc.model
 
+        classification_embeddings = list(model.embed(documents=docs, task_id=Task.CLASSIFICATION))
+        classification_embeddings = np.stack(classification_embeddings, axis=0)
+
+        assert classification_embeddings.shape == (len(docs), dim)
+
+        model = TextEmbedding(model_name=model_name, task_id=Task.CLASSIFICATION)
+        default_embeddings = list(model.embed(documents=docs))
+        default_embeddings = np.stack(default_embeddings, axis=0)
+
+        assert default_embeddings.shape == (len(docs), dim)
+
+        assert np.allclose(
+            classification_embeddings,
+            default_embeddings,
+            atol=1e-4,
+        ), model_desc.model
         if is_ci:
             delete_model_cache(model.model._model_dir)
 
@@ -140,7 +156,7 @@ def test_single_embedding_query():
 
         canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
         assert np.allclose(
-            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+            embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
         ), model_desc.model
 
         if is_ci:
@@ -172,7 +188,7 @@ def test_single_embedding_passage():
 
         canonical_vector = CANONICAL_VECTOR_VALUES[model_name][task_id]["vectors"]
         assert np.allclose(
-            embeddings[: len(docs), : canonical_vector.shape[1]], canonical_vector, atol=1e-4
+            embeddings[:, : canonical_vector.shape[1]], canonical_vector, atol=1e-4
         ), model_desc.model
 
         if is_ci:
@@ -207,27 +223,6 @@ def test_parallel_processing(dim: int, model_name: str):
         delete_model_cache(model.model._model_dir)
 
 
-def test_task_assignment():
-    is_ci = os.getenv("CI")
-    is_manual = os.getenv("GITHUB_EVENT_NAME") == "workflow_dispatch"
-
-    if is_ci and not is_manual:
-        pytest.skip("Skipping in CI non-manual mode")
-
-    for model_desc in JinaEmbeddingV3._list_supported_models():
-        # todo: once we add more models, we should not test models >1GB size locally
-        model_name = model_desc.model
-
-        model = TextEmbedding(model_name=model_name)
-
-        for i, task_id in enumerate(Task):
-            _ = list(model.embed(documents=docs, batch_size=1, task_id=i))
-            assert model.model.current_task_id == task_id
-
-        if is_ci:
-            delete_model_cache(model.model._model_dir)
-
-
 @pytest.mark.parametrize("model_name", ["jinaai/jina-embeddings-v3"])
 def test_lazy_load(model_name: str):
     is_ci = os.getenv("CI")

tests/test_text_onnx_embeddings.py

@@ -156,3 +156,22 @@ def test_lazy_load(model_name: str) -> None:
 
     if is_ci:
         delete_model_cache(model.model._model_dir)
+
+
+def test_get_embedding_size() -> None:
+    assert TextEmbedding.get_embedding_size("sentence-transformers/all-MiniLM-L6-v2") == 384
+    assert TextEmbedding.get_embedding_size("sentence-transformers/all-minilm-l6-v2") == 384
+
+
+def test_embedding_size() -> None:
+    is_ci = os.getenv("CI")
+    model_name = "sentence-transformers/all-MiniLM-L6-v2"
+    model = TextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 384
+
+    model_name = "sentence-transformers/all-minilm-l6-v2"
+    model = TextEmbedding(model_name=model_name, lazy_load=True)
+    assert model.embedding_size == 384
+
+    if is_ci:
+        delete_model_cache(model.model._model_dir)