13115725275

Committed 03 Feb 2025 02:28PM UTC coverage: 91.383% (+0.02%) from 91.359%

Build # 13115725275

Build Type

Pull #8794

github

Committed by

web-flow

Commit Message

Merge f9479e038 into 503d275ad

Pull Request Pull Request #8794: refactor: HF API Embedders - use `InferenceClient.feature_extraction` instead of `InferenceClient.post`

Run Details

8898 of 9737 relevant lines covered (91.38%)

0.91 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

96.63

haystack/components/embedders/hugging_face_api_document_embedder.py

# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
#
# SPDX-License-Identifier: Apache-2.0

import warnings
from typing import Any, Dict, List, Optional, Union

from tqdm import tqdm

from haystack import component, default_from_dict, default_to_dict, logging
from haystack.dataclasses import Document
from haystack.lazy_imports import LazyImport
from haystack.utils import Secret, deserialize_secrets_inplace
from haystack.utils.hf import HFEmbeddingAPIType, HFModelType, check_valid_model
from haystack.utils.url_validation import is_valid_http_url

with LazyImport(message="Run 'pip install \"huggingface_hub>=0.27.0\"'") as huggingface_hub_import:
    from huggingface_hub import InferenceClient

logger = logging.getLogger(__name__)


@component
class HuggingFaceAPIDocumentEmbedder:
    """
    Embeds documents using Hugging Face APIs.

    Use it with the following Hugging Face APIs:
    - [Free Serverless Inference API](https://huggingface.co/inference-api)
    - [Paid Inference Endpoints](https://huggingface.co/inference-endpoints)
    - [Self-hosted Text Embeddings Inference](https://github.com/huggingface/text-embeddings-inference)


    ### Usage examples

    #### With free serverless inference API

    ```python
    from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
    from haystack.utils import Secret
    from haystack.dataclasses import Document

    doc = Document(content="I love pizza!")

    doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="serverless_inference_api",
                                                  api_params={"model": "BAAI/bge-small-en-v1.5"},
                                                  token=Secret.from_token("<your-api-key>"))

    result = document_embedder.run([doc])
    print(result["documents"][0].embedding)

    # [0.017020374536514282, -0.023255806416273117, ...]
    ```

    #### With paid inference endpoints

    ```python
    from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
    from haystack.utils import Secret
    from haystack.dataclasses import Document

    doc = Document(content="I love pizza!")

    doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="inference_endpoints",
                                                  api_params={"url": "<your-inference-endpoint-url>"},
                                                  token=Secret.from_token("<your-api-key>"))

    result = document_embedder.run([doc])
    print(result["documents"][0].embedding)

    # [0.017020374536514282, -0.023255806416273117, ...]
    ```

    #### With self-hosted text embeddings inference

    ```python
    from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
    from haystack.dataclasses import Document

    doc = Document(content="I love pizza!")

    doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="text_embeddings_inference",
                                                  api_params={"url": "http://localhost:8080"})

    result = document_embedder.run([doc])
    print(result["documents"][0].embedding)

    # [0.017020374536514282, -0.023255806416273117, ...]
    ```
    """

    def __init__(
        self,
        api_type: Union[HFEmbeddingAPIType, str],
        api_params: Dict[str, str],
        token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
        prefix: str = "",
        suffix: str = "",
        truncate: Optional[bool] = True,
        normalize: Optional[bool] = False,
        batch_size: int = 32,
        progress_bar: bool = True,
        meta_fields_to_embed: Optional[List[str]] = None,
        embedding_separator: str = "\n",
    ):  # pylint: disable=too-many-positional-arguments
        """
        Creates a HuggingFaceAPIDocumentEmbedder component.

        :param api_type:
            The type of Hugging Face API to use.
        :param api_params:
            A dictionary with the following keys:
            - `model`: Hugging Face model ID. Required when `api_type` is `SERVERLESS_INFERENCE_API`.
            - `url`: URL of the inference endpoint. Required when `api_type` is `INFERENCE_ENDPOINTS` or
            `TEXT_EMBEDDINGS_INFERENCE`.
        :param token: The Hugging Face token to use as HTTP bearer authorization.
            Check your HF token in your [account settings](https://huggingface.co/settings/tokens).
        :param prefix:
            A string to add at the beginning of each text.
        :param suffix:
            A string to add at the end of each text.
        :param truncate:
            Truncates the input text to the maximum length supported by the model.
            Applicable when `api_type` is `TEXT_EMBEDDINGS_INFERENCE`, or `INFERENCE_ENDPOINTS`
            if the backend uses Text Embeddings Inference.
            If `api_type` is `SERVERLESS_INFERENCE_API`, this parameter is ignored.
        :param normalize:
            Normalizes the embeddings to unit length.
            Applicable when `api_type` is `TEXT_EMBEDDINGS_INFERENCE`, or `INFERENCE_ENDPOINTS`
            if the backend uses Text Embeddings Inference.
            If `api_type` is `SERVERLESS_INFERENCE_API`, this parameter is ignored.
        :param batch_size:
            Number of documents to process at once.
        :param progress_bar:
            If `True`, shows a progress bar when running.
        :param meta_fields_to_embed:
            List of metadata fields to embed along with the document text.
        :param embedding_separator:
            Separator used to concatenate the metadata fields to the document text.
        """
        huggingface_hub_import.check()

        if isinstance(api_type, str):
            api_type = HFEmbeddingAPIType.from_str(api_type)

        api_params = api_params or {}

        if api_type == HFEmbeddingAPIType.SERVERLESS_INFERENCE_API:
            model = api_params.get("model")
            if model is None:
                raise ValueError(
                    "To use the Serverless Inference API, you need to specify the `model` parameter in `api_params`."
                )
            check_valid_model(model, HFModelType.EMBEDDING, token)
            model_or_url = model
        elif api_type in [HFEmbeddingAPIType.INFERENCE_ENDPOINTS, HFEmbeddingAPIType.TEXT_EMBEDDINGS_INFERENCE]:
            url = api_params.get("url")
            if url is None:
                msg = (
                    "To use Text Embeddings Inference or Inference Endpoints, you need to specify the `url` "
                    "parameter in `api_params`."
                )
                raise ValueError(msg)
            if not is_valid_http_url(url):
                raise ValueError(f"Invalid URL: {url}")
            model_or_url = url
        else:
            msg = f"Unknown api_type {api_type}"
            raise ValueError(msg)

        self.api_type = api_type
        self.api_params = api_params
        self.token = token
        self.prefix = prefix
        self.suffix = suffix
        self.truncate = truncate
        self.normalize = normalize
        self.batch_size = batch_size
        self.progress_bar = progress_bar
        self.meta_fields_to_embed = meta_fields_to_embed or []
        self.embedding_separator = embedding_separator
        self._client = InferenceClient(model_or_url, token=token.resolve_value() if token else None)

    def to_dict(self) -> Dict[str, Any]:
        """
        Serializes the component to a dictionary.

        :returns:
            Dictionary with serialized data.
        """
        return default_to_dict(
            self,
            api_type=str(self.api_type),
            api_params=self.api_params,
            prefix=self.prefix,
            suffix=self.suffix,
            token=self.token.to_dict() if self.token else None,
            truncate=self.truncate,
            normalize=self.normalize,
            batch_size=self.batch_size,
            progress_bar=self.progress_bar,
            meta_fields_to_embed=self.meta_fields_to_embed,
            embedding_separator=self.embedding_separator,
        )

    @classmethod
    def from_dict(cls, data: Dict[str, Any]) -> "HuggingFaceAPIDocumentEmbedder":
        """
        Deserializes the component from a dictionary.

        :param data:
            Dictionary to deserialize from.
        :returns:
            Deserialized component.
        """
        deserialize_secrets_inplace(data["init_parameters"], keys=["token"])
        return default_from_dict(cls, data)

    def _prepare_texts_to_embed(self, documents: List[Document]) -> List[str]:
        """
        Prepare the texts to embed by concatenating the Document text with the metadata fields to embed.
        """
        texts_to_embed = []
        for doc in documents:
            meta_values_to_embed = [
                str(doc.meta[key]) for key in self.meta_fields_to_embed if key in doc.meta and doc.meta[key] is not None
            ]

            text_to_embed = (
                self.prefix + self.embedding_separator.join(meta_values_to_embed + [doc.content or ""]) + self.suffix
            )

            texts_to_embed.append(text_to_embed)
        return texts_to_embed

    def _embed_batch(self, texts_to_embed: List[str], batch_size: int) -> List[List[float]]:
        """
        Embed a list of texts in batches.
        """
        truncate = self.truncate
        normalize = self.normalize

        if self.api_type == HFEmbeddingAPIType.SERVERLESS_INFERENCE_API:
            if truncate is not None:
                msg = "`truncate` parameter is not supported for Serverless Inference API. It will be ignored."
                warnings.warn(msg)
                truncate = None
            if normalize is not None:
                msg = "`normalize` parameter is not supported for Serverless Inference API. It will be ignored."
                warnings.warn(msg)
                normalize = None

        all_embeddings = []
        for i in tqdm(
            range(0, len(texts_to_embed), batch_size), disable=not self.progress_bar, desc="Calculating embeddings"
        ):
            batch = texts_to_embed[i : i + batch_size]

            np_embeddings = self._client.feature_extraction(
                # this method does not officially support list of strings, but works as expected
                text=batch,  # type: ignore[arg-type]
                truncate=truncate,
                normalize=normalize,
            )

            if np_embeddings.ndim != 2 or np_embeddings.shape[0] != len(batch):
                raise ValueError(f"Expected embedding shape ({batch_size}, embedding_dim), got {np_embeddings.shape}")

            all_embeddings.extend(np_embeddings.tolist())

        return all_embeddings

    @component.output_types(documents=List[Document])
    def run(self, documents: List[Document]):
        """
        Embeds a list of documents.

        :param documents:
            Documents to embed.

        :returns:
            A dictionary with the following keys:
            - `documents`: A list of documents with embeddings.
        """
        if not isinstance(documents, list) or documents and not isinstance(documents[0], Document):
            raise TypeError(
                "HuggingFaceAPIDocumentEmbedder expects a list of Documents as input."
                " In case you want to embed a string, please use the HuggingFaceAPITextEmbedder."
            )

        texts_to_embed = self._prepare_texts_to_embed(documents=documents)

        embeddings = self._embed_batch(texts_to_embed=texts_to_embed, batch_size=self.batch_size)

        for doc, emb in zip(documents, embeddings):
            doc.embedding = emb

        return {"documents": documents}

1	# SPDX-FileCopyrightText: 2022-present deepset GmbH <info@deepset.ai>
2	#
3	# SPDX-License-Identifier: Apache-2.0
4
5	import warnings	1✔
6	from typing import Any, Dict, List, Optional, Union	1✔
7
8	from tqdm import tqdm	1✔
9
10	from haystack import component, default_from_dict, default_to_dict, logging	1✔
11	from haystack.dataclasses import Document	1✔
12	from haystack.lazy_imports import LazyImport	1✔
13	from haystack.utils import Secret, deserialize_secrets_inplace	1✔
14	from haystack.utils.hf import HFEmbeddingAPIType, HFModelType, check_valid_model	1✔
15	from haystack.utils.url_validation import is_valid_http_url	1✔
16
17	with LazyImport(message="Run 'pip install \"huggingface_hub>=0.27.0\"'") as huggingface_hub_import:	1✔
18	from huggingface_hub import InferenceClient	1✔
19
20	logger = logging.getLogger(__name__)	1✔
21
22
23	@component	1✔
24	class HuggingFaceAPIDocumentEmbedder:	1✔
25	"""
26	Embeds documents using Hugging Face APIs.
27
28	Use it with the following Hugging Face APIs:
29	- [Free Serverless Inference API](https://huggingface.co/inference-api)
30	- [Paid Inference Endpoints](https://huggingface.co/inference-endpoints)
31	- [Self-hosted Text Embeddings Inference](https://github.com/huggingface/text-embeddings-inference)
32
33
34	### Usage examples
35
36	#### With free serverless inference API
37
38	```python
39	from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
40	from haystack.utils import Secret
41	from haystack.dataclasses import Document
42
43	doc = Document(content="I love pizza!")
44
45	doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="serverless_inference_api",
46	api_params={"model": "BAAI/bge-small-en-v1.5"},
47	token=Secret.from_token("<your-api-key>"))
48
49	result = document_embedder.run([doc])
50	print(result["documents"][0].embedding)
51
52	# [0.017020374536514282, -0.023255806416273117, ...]
53	```
54
55	#### With paid inference endpoints
56
57	```python
58	from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
59	from haystack.utils import Secret
60	from haystack.dataclasses import Document
61
62	doc = Document(content="I love pizza!")
63
64	doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="inference_endpoints",
65	api_params={"url": "<your-inference-endpoint-url>"},
66	token=Secret.from_token("<your-api-key>"))
67
68	result = document_embedder.run([doc])
69	print(result["documents"][0].embedding)
70
71	# [0.017020374536514282, -0.023255806416273117, ...]
72	```
73
74	#### With self-hosted text embeddings inference
75
76	```python
77	from haystack.components.embedders import HuggingFaceAPIDocumentEmbedder
78	from haystack.dataclasses import Document
79
80	doc = Document(content="I love pizza!")
81
82	doc_embedder = HuggingFaceAPIDocumentEmbedder(api_type="text_embeddings_inference",
83	api_params={"url": "http://localhost:8080"})
84
85	result = document_embedder.run([doc])
86	print(result["documents"][0].embedding)
87
88	# [0.017020374536514282, -0.023255806416273117, ...]
89	```
90	"""
91
92	def __init__(	1✔
93	self,
94	api_type: Union[HFEmbeddingAPIType, str],
95	api_params: Dict[str, str],
96	token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
97	prefix: str = "",
98	suffix: str = "",
99	truncate: Optional[bool] = True,
100	normalize: Optional[bool] = False,
101	batch_size: int = 32,
102	progress_bar: bool = True,
103	meta_fields_to_embed: Optional[List[str]] = None,
104	embedding_separator: str = "\n",
105	): # pylint: disable=too-many-positional-arguments
106	"""
107	Creates a HuggingFaceAPIDocumentEmbedder component.
108
109	:param api_type:
110	The type of Hugging Face API to use.
111	:param api_params:
112	A dictionary with the following keys:
113	- `model`: Hugging Face model ID. Required when `api_type` is `SERVERLESS_INFERENCE_API`.
114	- `url`: URL of the inference endpoint. Required when `api_type` is `INFERENCE_ENDPOINTS` or
115	`TEXT_EMBEDDINGS_INFERENCE`.
116	:param token: The Hugging Face token to use as HTTP bearer authorization.
117	Check your HF token in your [account settings](https://huggingface.co/settings/tokens).
118	:param prefix:
119	A string to add at the beginning of each text.
120	:param suffix:
121	A string to add at the end of each text.
122	:param truncate:
123	Truncates the input text to the maximum length supported by the model.
124	Applicable when `api_type` is `TEXT_EMBEDDINGS_INFERENCE`, or `INFERENCE_ENDPOINTS`
125	if the backend uses Text Embeddings Inference.
126	If `api_type` is `SERVERLESS_INFERENCE_API`, this parameter is ignored.
127	:param normalize:
128	Normalizes the embeddings to unit length.
129	Applicable when `api_type` is `TEXT_EMBEDDINGS_INFERENCE`, or `INFERENCE_ENDPOINTS`
130	if the backend uses Text Embeddings Inference.
131	If `api_type` is `SERVERLESS_INFERENCE_API`, this parameter is ignored.
132	:param batch_size:
133	Number of documents to process at once.
134	:param progress_bar:
135	If `True`, shows a progress bar when running.
136	:param meta_fields_to_embed:
137	List of metadata fields to embed along with the document text.
138	:param embedding_separator:
139	Separator used to concatenate the metadata fields to the document text.
140	"""
141	huggingface_hub_import.check()	1✔
142
143	if isinstance(api_type, str):	1✔
144	api_type = HFEmbeddingAPIType.from_str(api_type)	1✔
145
146	api_params = api_params or {}	1✔
147
148	if api_type == HFEmbeddingAPIType.SERVERLESS_INFERENCE_API:	1✔
149	model = api_params.get("model")	1✔
150	if model is None:	1✔
151	raise ValueError(	1✔
152	"To use the Serverless Inference API, you need to specify the `model` parameter in `api_params`."
153	)
154	check_valid_model(model, HFModelType.EMBEDDING, token)	1✔
155	model_or_url = model	1✔
156	elif api_type in [HFEmbeddingAPIType.INFERENCE_ENDPOINTS, HFEmbeddingAPIType.TEXT_EMBEDDINGS_INFERENCE]:	1✔
157	url = api_params.get("url")	1✔
158	if url is None:	1✔
159	msg = (	1✔
160	"To use Text Embeddings Inference or Inference Endpoints, you need to specify the `url` "
161	"parameter in `api_params`."
162	)
163	raise ValueError(msg)	1✔
164	if not is_valid_http_url(url):	1✔
165	raise ValueError(f"Invalid URL: {url}")	1✔
166	model_or_url = url	1✔
167	else:
168	msg = f"Unknown api_type {api_type}"	×
169	raise ValueError(msg)	×
170
171	self.api_type = api_type	1✔
172	self.api_params = api_params	1✔
173	self.token = token	1✔
174	self.prefix = prefix	1✔
175	self.suffix = suffix	1✔
176	self.truncate = truncate	1✔
177	self.normalize = normalize	1✔
178	self.batch_size = batch_size	1✔
179	self.progress_bar = progress_bar	1✔
180	self.meta_fields_to_embed = meta_fields_to_embed or []	1✔
181	self.embedding_separator = embedding_separator	1✔
182	self._client = InferenceClient(model_or_url, token=token.resolve_value() if token else None)	1✔
183
184	def to_dict(self) -> Dict[str, Any]:	1✔
185	"""
186	Serializes the component to a dictionary.
187
188	:returns:
189	Dictionary with serialized data.
190	"""
191	return default_to_dict(	1✔
192	self,
193	api_type=str(self.api_type),
194	api_params=self.api_params,
195	prefix=self.prefix,
196	suffix=self.suffix,
197	token=self.token.to_dict() if self.token else None,
198	truncate=self.truncate,
199	normalize=self.normalize,
200	batch_size=self.batch_size,
201	progress_bar=self.progress_bar,
202	meta_fields_to_embed=self.meta_fields_to_embed,
203	embedding_separator=self.embedding_separator,
204	)
205
206	@classmethod	1✔
207	def from_dict(cls, data: Dict[str, Any]) -> "HuggingFaceAPIDocumentEmbedder":	1✔
208	"""
209	Deserializes the component from a dictionary.
210
211	:param data:
212	Dictionary to deserialize from.
213	:returns:
214	Deserialized component.
215	"""
216	deserialize_secrets_inplace(data["init_parameters"], keys=["token"])	1✔
217	return default_from_dict(cls, data)	1✔
218
219	def _prepare_texts_to_embed(self, documents: List[Document]) -> List[str]:	1✔
220	"""
221	Prepare the texts to embed by concatenating the Document text with the metadata fields to embed.
222	"""
223	texts_to_embed = []	1✔
224	for doc in documents:	1✔
225	meta_values_to_embed = [	1✔
226	str(doc.meta[key]) for key in self.meta_fields_to_embed if key in doc.meta and doc.meta[key] is not None
227	]
228
229	text_to_embed = (	1✔
230	self.prefix + self.embedding_separator.join(meta_values_to_embed + [doc.content or ""]) + self.suffix
231	)
232
233	texts_to_embed.append(text_to_embed)	1✔
234	return texts_to_embed	1✔
235
236	def _embed_batch(self, texts_to_embed: List[str], batch_size: int) -> List[List[float]]:	1✔
237	"""
238	Embed a list of texts in batches.
239	"""
240	truncate = self.truncate	1✔
241	normalize = self.normalize	1✔
242
243	if self.api_type == HFEmbeddingAPIType.SERVERLESS_INFERENCE_API:	1✔
244	if truncate is not None:	1✔
245	msg = "`truncate` parameter is not supported for Serverless Inference API. It will be ignored."	1✔
246	warnings.warn(msg)	1✔
247	truncate = None	1✔
248	if normalize is not None:	1✔
249	msg = "`normalize` parameter is not supported for Serverless Inference API. It will be ignored."	1✔
250	warnings.warn(msg)	1✔
251	normalize = None	1✔
252
253	all_embeddings = []	1✔
254	for i in tqdm(	1✔
255	range(0, len(texts_to_embed), batch_size), disable=not self.progress_bar, desc="Calculating embeddings"
256	):
257	batch = texts_to_embed[i : i + batch_size]	1✔
258
259	np_embeddings = self._client.feature_extraction(	1✔
260	# this method does not officially support list of strings, but works as expected
261	text=batch, # type: ignore[arg-type]
262	truncate=truncate,
263	normalize=normalize,
264	)
265
266	if np_embeddings.ndim != 2 or np_embeddings.shape[0] != len(batch):	1✔
267	raise ValueError(f"Expected embedding shape ({batch_size}, embedding_dim), got {np_embeddings.shape}")	1✔
268
269	all_embeddings.extend(np_embeddings.tolist())	1✔
270
271	return all_embeddings	1✔
272
273	@component.output_types(documents=List[Document])	1✔
274	def run(self, documents: List[Document]):	1✔
275	"""
276	Embeds a list of documents.
277
278	:param documents:
279	Documents to embed.
280
281	:returns:
282	A dictionary with the following keys:
283	- `documents`: A list of documents with embeddings.
284	"""
285	if not isinstance(documents, list) or documents and not isinstance(documents[0], Document):	1✔
286	raise TypeError(	×
287	"HuggingFaceAPIDocumentEmbedder expects a list of Documents as input."
288	" In case you want to embed a string, please use the HuggingFaceAPITextEmbedder."
289	)
290
291	texts_to_embed = self._prepare_texts_to_embed(documents=documents)	1✔
292
293	embeddings = self._embed_batch(texts_to_embed=texts_to_embed, batch_size=self.batch_size)	1✔
294
295	for doc, emb in zip(documents, embeddings):	1✔
296	doc.embedding = emb	1✔
297
298	return {"documents": documents}	1✔

deepset-ai / haystack / 13115725275

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous