_load_map_reduce_chain() — langchain Function Reference

Source Code

libs/langchain/langchain_classic/chains/summarize/chain.py lines 68–169

def _load_map_reduce_chain(
    llm: BaseLanguageModel,
    *,
    map_prompt: BasePromptTemplate = map_reduce_prompt.PROMPT,
    combine_prompt: BasePromptTemplate = map_reduce_prompt.PROMPT,
    combine_document_variable_name: str = "text",
    map_reduce_document_variable_name: str = "text",
    collapse_prompt: BasePromptTemplate | None = None,
    reduce_llm: BaseLanguageModel | None = None,
    collapse_llm: BaseLanguageModel | None = None,
    verbose: bool | None = None,
    token_max: int = 3000,
    callbacks: Callbacks = None,
    collapse_max_retries: int | None = None,
    **kwargs: Any,
) -> MapReduceDocumentsChain:
    map_chain = LLMChain(
        llm=llm,
        prompt=map_prompt,
        verbose=verbose,
        callbacks=callbacks,
    )
    _reduce_llm = reduce_llm or llm
    reduce_chain = LLMChain(
        llm=_reduce_llm,
        prompt=combine_prompt,
        verbose=verbose,
        callbacks=callbacks,
    )
    """Load a MapReduceDocumentsChain for summarization.

    This chain first applies a "map" step to summarize each document,
    then applies a "reduce" step to combine the summaries into a
    final result. Optionally, a "collapse" step can be used to handle
    long intermediate results.

    Args:
        llm: Language Model to use for map and reduce steps.
        map_prompt: Prompt used to summarize each document in the map step.
        combine_prompt: Prompt used to combine summaries in the reduce step.
        combine_document_variable_name: Variable name in the `combine_prompt` where
            the mapped summaries are inserted.
        map_reduce_document_variable_name: Variable name in the `map_prompt`
            where document text is inserted.
        collapse_prompt: Optional prompt used to collapse intermediate summaries
            if they exceed the token limit (`token_max`).
        reduce_llm: Optional separate LLM for the reduce step.
            which uses the same model as the map step.
        collapse_llm: Optional separate LLM for the collapse step.
            which uses the same model as the map step.
        verbose: Whether to log progress and intermediate steps.
        token_max: Token threshold that triggers the collapse step during reduction.
        callbacks: Optional callbacks for logging and tracing.
        collapse_max_retries: Maximum retries for the collapse step if it fails.

        **kwargs: Additional keyword arguments passed to the MapReduceDocumentsChain.

    Returns:
        A MapReduceDocumentsChain that maps each document to a summary,
        then reduces all summaries into a single cohesive result.
    """
    combine_documents_chain = StuffDocumentsChain(
        llm_chain=reduce_chain,
        document_variable_name=combine_document_variable_name,
        verbose=verbose,
        callbacks=callbacks,
    )
    if collapse_prompt is None:
        collapse_chain = None
        if collapse_llm is not None:
            msg = (
                "collapse_llm provided, but collapse_prompt was not: please "
                "provide one or stop providing collapse_llm."
            )
            raise ValueError(msg)
    else:
        _collapse_llm = collapse_llm or llm
        collapse_chain = StuffDocumentsChain(
            llm_chain=LLMChain(
                llm=_collapse_llm,
                prompt=collapse_prompt,