deepset-ai--haystack/haystack/components/extractors/llm_metadata_extractor.py

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

import copy
import json
from asyncio import Semaphore, gather
from collections.abc import Iterable
from concurrent.futures import ThreadPoolExecutor
from dataclasses import replace
from typing import Any

from jinja2 import meta
from jinja2.sandbox import SandboxedEnvironment

from haystack import Document, component, default_from_dict, default_to_dict, logging
from haystack.components.builders import PromptBuilder
from haystack.components.generators.chat.types import ChatGenerator
from haystack.components.preprocessors import DocumentSplitter
from haystack.core.serialization import component_to_dict
from haystack.dataclasses import ChatMessage
from haystack.utils import deserialize_chatgenerator_inplace, expand_page_range
from haystack.utils.async_utils import _execute_component_async
from haystack.utils.misc import _parse_dict_from_json

logger = logging.getLogger(__name__)


@component
class LLMMetadataExtractor:
    """
    Extracts metadata from documents using a Large Language Model (LLM).

    The metadata is extracted by providing a prompt to an LLM that generates the metadata.

    This component expects as input a list of documents and a prompt. The prompt must have exactly one variable, called
    `document`, that points to a single document in the list of documents. So to access the content of the document,
    you can use `{{ document.content }}` in the prompt.

    The component will run the LLM on each document in the list and extract metadata from the document. The metadata
    will be added to the document's metadata field. If the LLM fails to extract metadata from a document, the document
    will be added to the `failed_documents` list. The failed documents will have the keys `metadata_extraction_error` and
    `metadata_extraction_response` in their metadata. These documents can be re-run with another extractor to
    extract metadata by using the `metadata_extraction_response` and `metadata_extraction_error` in the prompt.

    ```python
    from haystack import Document
    from haystack.components.extractors.llm_metadata_extractor import LLMMetadataExtractor
    from haystack.components.generators.chat import OpenAIChatGenerator

    NER_PROMPT = '''
    -Goal-
    Given text and a list of entity types, identify all entities of those types from the text.

    -Steps-
    1. Identify all entities. For each identified entity, extract the following information:
    - entity: Name of the entity
    - entity_type: One of the following types: [organization, product, service, industry]
    Format each entity as a JSON like: {"entity": <entity_name>, "entity_type": <entity_type>}

    2. Return output in a single list with all the entities identified in steps 1.

    -Examples-
    ######################
    Example 1:
    entity_types: [organization, person, partnership, financial metric, product, service, industry, investment strategy, market trend]
    text: Another area of strength is our co-brand issuance. Visa is the primary network partner for eight of the top
    10 co-brand partnerships in the US today and we are pleased that Visa has finalized a multi-year extension of
    our successful credit co-branded partnership with Alaska Airlines, a portfolio that benefits from a loyal customer
    base and high cross-border usage.
    We have also had significant co-brand momentum in CEMEA. First, we launched a new co-brand card in partnership
    with Qatar Airways, British Airways and the National Bank of Kuwait. Second, we expanded our strong global
    Marriott relationship to launch Qatar's first hospitality co-branded card with Qatar Islamic Bank. Across the
    United Arab Emirates, we now have exclusive agreements with all the leading airlines marked by a recent
    agreement with Emirates Skywards.
    And we also signed an inaugural Airline co-brand agreement in Morocco with Royal Air Maroc. Now newer digital
    issuers are equally
    ------------------------
    output:
    {"entities": [{"entity": "Visa", "entity_type": "company"}, {"entity": "Alaska Airlines", "entity_type": "company"}, {"entity": "Qatar Airways", "entity_type": "company"}, {"entity": "British Airways", "entity_type": "company"}, {"entity": "National Bank of Kuwait", "entity_type": "company"}, {"entity": "Marriott", "entity_type": "company"}, {"entity": "Qatar Islamic Bank", "entity_type": "company"}, {"entity": "Emirates Skywards", "entity_type": "company"}, {"entity": "Royal Air Maroc", "entity_type": "company"}]}
    #############################
    -Real Data-
    ######################
    entity_types: [company, organization, person, country, product, service]
    text: {{ document.content }}
    ######################
    output:
    '''

    docs = [
        Document(content="deepset was founded in 2018 in Berlin, and is known for its Haystack framework"),
        Document(content="Hugging Face is a company that was founded in New York, USA and is known for its Transformers library")
    ]

    chat_generator = OpenAIChatGenerator(
        generation_kwargs={
            "max_completion_tokens": 500,
            "temperature": 0.0,
            "seed": 0,
            "response_format": {
                "type": "json_schema",
                "json_schema": {
                    "name": "entity_extraction",
                    "schema": {
                        "type": "object",
                        "properties": {
                            "entities": {
                                "type": "array",
                                "items": {
                                    "type": "object",
                                    "properties": {
                                        "entity": {"type": "string"},
                                        "entity_type": {"type": "string"}
                                    },
                                    "required": ["entity", "entity_type"],
                                    "additionalProperties": False
                                }
                            }
                        },
                        "required": ["entities"],
                        "additionalProperties": False
                    }
                }
            },
        },
        max_retries=1,
        timeout=60.0,
    )

    extractor = LLMMetadataExtractor(
        prompt=NER_PROMPT,
        chat_generator=chat_generator,
        expected_keys=["entities"],
        raise_on_failure=False,
    )

    extractor.run(documents=docs)
    # >> {'documents': [
    #     Document(id=.., content: 'deepset was founded in 2018 in Berlin, and is known for its Haystack framework',
    #     meta: {'entities': [{'entity': 'deepset', 'entity_type': 'company'}, {'entity': 'Berlin', 'entity_type': 'city'},
    #           {'entity': 'Haystack', 'entity_type': 'product'}]}),
    #     Document(id=.., content: 'Hugging Face is a company that was founded in New York, USA and is known for its Transformers library',
    #     meta: {'entities': [
    #             {'entity': 'Hugging Face', 'entity_type': 'company'}, {'entity': 'New York', 'entity_type': 'city'},
    #             {'entity': 'USA', 'entity_type': 'country'}, {'entity': 'Transformers', 'entity_type': 'product'}
    #             ]})
    #        ]
    #     'failed_documents': []
    #    }
    # >>
    ```
    """  # noqa: E501

    def __init__(
        self,
        prompt: str,
        chat_generator: ChatGenerator,
        expected_keys: list[str] | None = None,
        page_range: list[str | int] | None = None,
        raise_on_failure: bool = False,
        max_workers: int = 3,
    ) -> None:
        """
        Initializes the LLMMetadataExtractor.

        :param prompt: The prompt to be used for the LLM. It must contain exactly one variable, called `document`,
            which points to a single document in the list of documents. For example, to access the content of the
            document, use `{{ document.content }}` in the prompt.
        :param chat_generator: a ChatGenerator instance which represents the LLM. In order for the component to work,
            the LLM should be configured to return a JSON object. For example, when using the OpenAIChatGenerator, you
            should pass `{"response_format": {"type": "json_object"}}` in the `generation_kwargs`.
        :param expected_keys: The keys expected in the JSON output from the LLM.
        :param page_range: A range of pages to extract metadata from. For example, page_range=['1', '3'] will extract
            metadata from the first and third pages of each document. It also accepts printable range strings, e.g.:
            ['1-3', '5', '8', '10-12'] will extract metadata from pages 1, 2, 3, 5, 8, 10,11, 12.
            If None, metadata will be extracted from the entire document for each document in the documents list.
            This parameter is optional and can be overridden in the `run` method.
        :param raise_on_failure: Whether to raise an error on failure during the execution of the Generator or
            validation of the JSON output.
        :param max_workers: The maximum number of workers to use in the thread pool executor.
            This parameter is used limit the maximum number of requests that should be allowed to run concurrently
            when using the `run_async` method.
        """
        self.prompt = prompt
        ast = SandboxedEnvironment().parse(prompt)
        template_variables = meta.find_undeclared_variables(ast)
        variables = list(template_variables)
        if variables != ["document"]:
            raise ValueError(
                f"Prompt must have exactly one variable called 'document'. "
                f"Found {','.join(variables) or 'no variables'} in the prompt."
            )
        self.builder = PromptBuilder(prompt, required_variables=variables)
        self.raise_on_failure = raise_on_failure
        self.expected_keys = expected_keys or []
        self.splitter = DocumentSplitter(split_by="page", split_length=1)
        self.expanded_range = expand_page_range(page_range) if page_range else None
        self.max_workers = max_workers
        self._chat_generator = chat_generator

    def warm_up(self) -> None:
        """
        Warm up the underlying chat generator and splitter.
        """
        for inner in (self._chat_generator, self.splitter):
            if hasattr(inner, "warm_up"):
                inner.warm_up()

    async def warm_up_async(self) -> None:
        """
        Warm up the underlying chat generator and splitter on the serving event loop.
        """
        for inner in (self._chat_generator, self.splitter):
            if hasattr(inner, "warm_up_async"):
                await inner.warm_up_async()
            elif hasattr(inner, "warm_up"):
                inner.warm_up()

    def close(self) -> None:
        """
        Release the underlying chat generator's and splitter's resources.
        """
        for inner in (self._chat_generator, self.splitter):
            if hasattr(inner, "close"):
                inner.close()

    async def close_async(self) -> None:
        """
        Release the underlying chat generator's and splitter's async resources.
        """
        for inner in (self._chat_generator, self.splitter):
            if hasattr(inner, "close_async"):
                await inner.close_async()
            elif hasattr(inner, "close"):
                inner.close()

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

        :returns:
            Dictionary with serialized data.
        """

        return default_to_dict(
            self,
            prompt=self.prompt,
            chat_generator=component_to_dict(obj=self._chat_generator, name="chat_generator"),
            expected_keys=self.expected_keys,
            page_range=self.expanded_range,
            raise_on_failure=self.raise_on_failure,
            max_workers=self.max_workers,
        )

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

        :param data:
            Dictionary with serialized data.
        :returns:
            An instance of the component.
        """

        deserialize_chatgenerator_inplace(data["init_parameters"], key="chat_generator")
        return default_from_dict(cls, data)

    def _extract_metadata(self, llm_answer: str) -> dict[str, Any]:
        try:
            parsed_metadata = _parse_dict_from_json(llm_answer, expected_keys=self.expected_keys, raise_on_failure=True)
        except (ValueError, json.JSONDecodeError) as e:
            logger.warning(
                "Response from the LLM is not valid JSON or missing expected keys. Received output: {response}",
                response=llm_answer,
            )
            if self.raise_on_failure:
                raise e
            return {"error": "Response is not valid JSON or missing keys. Error: " + str(e)}

        return parsed_metadata

    def _prepare_prompts(
        self, documents: list[Document], expanded_range: list[int] | None = None
    ) -> list[ChatMessage | None]:
        all_prompts: list[ChatMessage | None] = []
        for document in documents:
            if not document.content:
                logger.warning("Document {doc_id} has no content. Skipping metadata extraction.", doc_id=document.id)
                all_prompts.append(None)
                continue

            if expanded_range:
                doc_copy = copy.deepcopy(document)
                pages = self.splitter.run(documents=[doc_copy])
                content = ""
                for idx, page in enumerate(pages["documents"]):
                    if idx + 1 in expanded_range and page.content is not None:
                        content += page.content
                doc_copy = replace(doc_copy, content=content)
            else:
                doc_copy = document

            prompt_with_doc = self.builder.run(template=self.prompt, template_variables={"document": doc_copy})

            # build a ChatMessage with the prompt
            message = ChatMessage.from_user(prompt_with_doc["prompt"])
            all_prompts.append(message)

        return all_prompts

    def _run_on_thread(self, prompt: ChatMessage | None) -> dict[str, Any]:
        # If prompt is None, return an error dictionary
        if prompt is None:
            return {"error": "Document has no content, skipping LLM call."}

        try:
            result = self._chat_generator.run(messages=[prompt])
        except Exception as e:
            if self.raise_on_failure:
                raise e
            logger.exception(
                "LLM {class_name} execution failed. Skipping metadata extraction. Failed with exception '{error}'.",
                class_name=self._chat_generator.__class__.__name__,
                error=e,
            )
            result = {"error": "LLM failed with exception: " + str(e)}
        return result

    async def _run_async(self, prompt: ChatMessage | None) -> dict[str, Any]:
        # If prompt is None, return an error dictionary
        if prompt is None:
            return {"error": "Document has no content, skipping LLM call."}

        try:
            result = await _execute_component_async(self._chat_generator, messages=[prompt])
        except Exception as e:
            if self.raise_on_failure:
                raise e
            logger.exception(
                "LLM {class_name} execution failed. Skipping metadata extraction. Failed with exception '{error}'.",
                class_name=self._chat_generator.__class__.__name__,
                error=e,
            )
            result = {"error": "LLM failed with exception: " + str(e)}
        return result

    def _process_results(
        self, documents: list[Document], results: Iterable[dict[str, Any]]
    ) -> tuple[list[Document], list[Document]]:
        successful_documents = []
        failed_documents = []
        for document, result in zip(documents, results, strict=True):
            new_meta = {**document.meta}
            if "error" in result:
                new_meta["metadata_extraction_error"] = result["error"]
                new_meta["metadata_extraction_response"] = None
                failed_documents.append(replace(document, meta=new_meta))
                continue

            parsed_metadata = self._extract_metadata(result["replies"][0].text)
            if "error" in parsed_metadata:
                new_meta["metadata_extraction_error"] = parsed_metadata["error"]
                new_meta["metadata_extraction_response"] = result["replies"][0]
                failed_documents.append(replace(document, meta=new_meta))
                continue

            for key in parsed_metadata:
                new_meta[key] = parsed_metadata[key]
                # Remove metadata_extraction_error and metadata_extraction_response if present from previous runs
                new_meta.pop("metadata_extraction_error", None)
                new_meta.pop("metadata_extraction_response", None)
            successful_documents.append(replace(document, meta=new_meta))
        return successful_documents, failed_documents

    @component.output_types(documents=list[Document], failed_documents=list[Document])
    def run(self, documents: list[Document], page_range: list[str | int] | None = None) -> dict[str, Any]:
        """
        Extract metadata from documents using a Large Language Model.

        If `page_range` is provided, the metadata will be extracted from the specified range of pages. This component
        will split the documents into pages and extract metadata from the specified range of pages. The metadata will be
        extracted from the entire document if `page_range` is not provided.

        The original documents will be returned  updated with the extracted metadata.

        :param documents: List of documents to extract metadata from.
        :param page_range: A range of pages to extract metadata from. For example, page_range=['1', '3'] will extract
                           metadata from the first and third pages of each document. It also accepts printable range
                           strings, e.g.: ['1-3', '5', '8', '10-12'] will extract metadata from pages 1, 2, 3, 5, 8, 10,
                           11, 12.
                           If None, metadata will be extracted from the entire document for each document in the
                           documents list.
        :returns:
            A dictionary with the keys:
            - "documents": A list of documents that were successfully updated with the extracted metadata.
            - "failed_documents": A list of documents that failed to extract metadata. These documents will have
            "metadata_extraction_error" and "metadata_extraction_response" in their metadata. These documents can be
            re-run with the extractor to extract metadata.
        """
        if len(documents) == 0:
            logger.warning("No documents provided. Skipping metadata extraction.")
            return {"documents": [], "failed_documents": []}

        self.warm_up()

        expanded_range = self.expanded_range
        if page_range:
            expanded_range = expand_page_range(page_range)

        # Create ChatMessage prompts for each document
        all_prompts = self._prepare_prompts(documents=documents, expanded_range=expanded_range)

        # Run the LLM on each prompt
        with ThreadPoolExecutor(max_workers=self.max_workers) as executor:
            results = executor.map(self._run_on_thread, all_prompts)

        successful_documents, failed_documents = self._process_results(documents, results)

        return {"documents": successful_documents, "failed_documents": failed_documents}

    @component.output_types(documents=list[Document], failed_documents=list[Document])
    async def run_async(self, documents: list[Document], page_range: list[str | int] | None = None) -> dict[str, Any]:
        """
        Asynchronously extract metadata from documents using a Large Language Model.

        If `page_range` is provided, the metadata will be extracted from the specified range of pages. This component
        will split the documents into pages and extract metadata from the specified range of pages. The metadata will be
        extracted from the entire document if `page_range` is not provided.

        The original documents will be returned  updated with the extracted metadata.

        This is the asynchronous version of the `run` method. It has the same parameters
        and return values but can be used with `await` in an async code.

        :param documents: List of documents to extract metadata from.
        :param page_range: A range of pages to extract metadata from. For example, page_range=['1', '3'] will extract
                           metadata from the first and third pages of each document. It also accepts printable range
                           strings, e.g.: ['1-3', '5', '8', '10-12'] will extract metadata from pages 1, 2, 3, 5, 8, 10,
                           11, 12.
                           If None, metadata will be extracted from the entire document for each document in the
                           documents list.
        :returns:
            A dictionary with the keys:
            - "documents": A list of documents that were successfully updated with the extracted metadata.
            - "failed_documents": A list of documents that failed to extract metadata. These documents will have
            "metadata_extraction_error" and "metadata_extraction_response" in their metadata. These documents can be
            re-run with the extractor to extract metadata.
        """
        if len(documents) == 0:
            logger.warning("No documents provided. Skipping metadata extraction.")
            return {"documents": [], "failed_documents": []}

        await self.warm_up_async()

        expanded_range = self.expanded_range
        if page_range:
            expanded_range = expand_page_range(page_range)

        # Create ChatMessage prompts for each document
        all_prompts = self._prepare_prompts(documents=documents, expanded_range=expanded_range)

        # Run the LLM on each prompt, bounding concurrency per task so max_workers is enforced.
        sem = Semaphore(max(1, self.max_workers))

        async def _bounded_run(prompt: ChatMessage | None) -> dict[str, Any]:
            async with sem:
                return await self._run_async(prompt)

        results = await gather(*[_bounded_run(prompt) for prompt in all_prompts])

        successful_documents, failed_documents = self._process_results(documents, results)

        return {"documents": successful_documents, "failed_documents": failed_documents}