haystack/haystack/components/classifiers/zero_shot_document_classifier.py

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

from typing import Any, Dict, List, Optional

from haystack import Document, component, default_from_dict, default_to_dict
from haystack.lazy_imports import LazyImport
from haystack.utils import ComponentDevice, Secret, deserialize_secrets_inplace
from haystack.utils.hf import deserialize_hf_model_kwargs, resolve_hf_pipeline_kwargs, serialize_hf_model_kwargs

with LazyImport(message="Run 'pip install transformers[torch,sentencepiece]'") as torch_and_transformers_import:
    from transformers import Pipeline as HfPipeline
    from transformers import pipeline


@component
class TransformersZeroShotDocumentClassifier:
    """
    Performs zero-shot classification of documents based on given labels and adds the predicted label to their metadata.

    The component uses a Hugging Face pipeline for zero-shot classification.
    Provide the model and the set of labels to be used for categorization during initialization.
    Additionally, you can configure the component to allow multiple labels to be true.

    Classification is run on the document's content field by default. If you want it to run on another field, set the
    `classification_field` to one of the document's metadata fields.

    Available models for the task of zero-shot-classification include:
        - `valhalla/distilbart-mnli-12-3`
        - `cross-encoder/nli-distilroberta-base`
        - `cross-encoder/nli-deberta-v3-xsmall`

    ### Usage example

    The following is a pipeline that classifies documents based on predefined classification labels
    retrieved from a search pipeline:

    ```python
    from haystack import Document
    from haystack.components.retrievers.in_memory import InMemoryBM25Retriever
    from haystack.document_stores.in_memory import InMemoryDocumentStore
    from haystack.core.pipeline import Pipeline
    from haystack.components.classifiers import TransformersZeroShotDocumentClassifier

    documents = [Document(id="0", content="Today was a nice day!"),
                 Document(id="1", content="Yesterday was a bad day!")]

    document_store = InMemoryDocumentStore()
    retriever = InMemoryBM25Retriever(document_store=document_store)
    document_classifier = TransformersZeroShotDocumentClassifier(
        model="cross-encoder/nli-deberta-v3-xsmall",
        labels=["positive", "negative"],
    )

    document_store.write_documents(documents)

    pipeline = Pipeline()
    pipeline.add_component(instance=retriever, name="retriever")
    pipeline.add_component(instance=document_classifier, name="document_classifier")
    pipeline.connect("retriever", "document_classifier")

    queries = ["How was your day today?", "How was your day yesterday?"]
    expected_predictions = ["positive", "negative"]

    for idx, query in enumerate(queries):
        result = pipeline.run({"retriever": {"query": query, "top_k": 1}})
        assert result["document_classifier"]["documents"][0].to_dict()["id"] == str(idx)
        assert (result["document_classifier"]["documents"][0].to_dict()["classification"]["label"]
                == expected_predictions[idx])
    ```
    """

    def __init__(  # pylint: disable=too-many-positional-arguments
        self,
        model: str,
        labels: List[str],
        multi_label: bool = False,
        classification_field: Optional[str] = None,
        device: Optional[ComponentDevice] = None,
        token: Optional[Secret] = Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False),
        huggingface_pipeline_kwargs: Optional[Dict[str, Any]] = None,
    ):
        """
        Initializes the TransformersZeroShotDocumentClassifier.

        See the Hugging Face [website](https://huggingface.co/models?pipeline_tag=zero-shot-classification&sort=downloads&search=nli)
        for the full list of zero-shot classification models (NLI) models.

        :param model:
            The name or path of a Hugging Face model for zero shot document classification.
        :param labels:
            The set of possible class labels to classify each document into, for example,
            ["positive", "negative"]. The labels depend on the selected model.
        :param multi_label:
            Whether or not multiple candidate labels can be true.
            If `False`, the scores are normalized such that
            the sum of the label likelihoods for each sequence is 1. If `True`, the labels are considered
            independent and probabilities are normalized for each candidate by doing a softmax of the entailment
            score vs. the contradiction score.
        :param classification_field:
            Name of document's meta field to be used for classification.
            If not set, `Document.content` is used by default.
        :param device:
            The device on which the model is loaded. If `None`, the default device is automatically
            selected. If a device/device map is specified in `huggingface_pipeline_kwargs`, it overrides this parameter.
        :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 huggingface_pipeline_kwargs:
            Dictionary containing keyword arguments used to initialize the
            Hugging Face pipeline for text classification.
        """

        torch_and_transformers_import.check()

        self.classification_field = classification_field

        self.token = token
        self.labels = labels
        self.multi_label = multi_label

        huggingface_pipeline_kwargs = resolve_hf_pipeline_kwargs(
            huggingface_pipeline_kwargs=huggingface_pipeline_kwargs or {},
            model=model,
            task="zero-shot-classification",
            supported_tasks=["zero-shot-classification"],
            device=device,
            token=token,
        )

        self.huggingface_pipeline_kwargs = huggingface_pipeline_kwargs
        self.pipeline: Optional[HfPipeline] = None

    def _get_telemetry_data(self) -> Dict[str, Any]:
        """
        Data that is sent to Posthog for usage analytics.
        """
        if isinstance(self.huggingface_pipeline_kwargs["model"], str):
            return {"model": self.huggingface_pipeline_kwargs["model"]}
        return {"model": f"[object of type {type(self.huggingface_pipeline_kwargs['model'])}]"}

    def warm_up(self):
        """
        Initializes the component.
        """
        if self.pipeline is None:
            self.pipeline = pipeline(**self.huggingface_pipeline_kwargs)

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

        :returns:
            Dictionary with serialized data.
        """
        serialization_dict = default_to_dict(
            self,
            labels=self.labels,
            model=self.huggingface_pipeline_kwargs["model"],
            huggingface_pipeline_kwargs=self.huggingface_pipeline_kwargs,
            token=self.token.to_dict() if self.token else None,
        )

        huggingface_pipeline_kwargs = serialization_dict["init_parameters"]["huggingface_pipeline_kwargs"]
        huggingface_pipeline_kwargs.pop("token", None)

        serialize_hf_model_kwargs(huggingface_pipeline_kwargs)
        return serialization_dict

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

        :param data:
            Dictionary to deserialize from.
        :returns:
            Deserialized component.
        """
        deserialize_secrets_inplace(data["init_parameters"], keys=["token"])
        if data["init_parameters"].get("huggingface_pipeline_kwargs") is not None:
            deserialize_hf_model_kwargs(data["init_parameters"]["huggingface_pipeline_kwargs"])
        return default_from_dict(cls, data)

    @component.output_types(documents=List[Document])
    def run(self, documents: List[Document], batch_size: int = 1):
        """
        Classifies the documents based on the provided labels and adds them to their metadata.

        The classification results are stored in the `classification` dict within
        each document's metadata. If `multi_label` is set to `True`, the scores for each label are available under
        the `details` key within the `classification` dictionary.

        :param documents:
            Documents to process.
        :param batch_size:
            Batch size used for processing the content in each document.
        :returns:
            A dictionary with the following key:
            - `documents`: A list of documents with an added metadata field called `classification`.
        """

        if self.pipeline is None:
            raise RuntimeError(
                "The component TransformerZeroShotDocumentClassifier wasn't warmed up. "
                "Run 'warm_up()' before calling 'run()'."
            )

        if not isinstance(documents, list) or documents and not isinstance(documents[0], Document):
            raise TypeError(
                "TransformerZeroShotDocumentClassifier expects a list of documents as input. "
                "In case you want to classify and route a text, please use the TransformersZeroShotTextRouter."
            )

        invalid_doc_ids = []

        for doc in documents:
            if self.classification_field is not None and self.classification_field not in doc.meta:
                invalid_doc_ids.append(doc.id)

        if invalid_doc_ids:
            raise ValueError(
                f"The following documents do not have the classification field '{self.classification_field}': "
                f"{', '.join(invalid_doc_ids)}"
            )

        texts = [
            (doc.content if self.classification_field is None else doc.meta[self.classification_field])
            for doc in documents
        ]

        predictions = self.pipeline(texts, self.labels, multi_label=self.multi_label, batch_size=batch_size)

        for prediction, document in zip(predictions, documents):
            formatted_prediction = {
                "label": prediction["labels"][0],
                "score": prediction["scores"][0],
                "details": dict(zip(prediction["labels"], prediction["scores"])),
            }
            document.meta["classification"] = formatted_prediction

        return {"documents": documents}