Skip to main content

Document Chains

The document chains in LangChain provide various ways to combine and operate over multiple input documents. They allow summarizing, answering questions, and extracting information across documents.

Overview​

The core document chains include:

  • Stuff: Concatenates documents into a prompt for an LLM
  • Refine: Iteratively refines a response by looping over documents
  • MapReduce: Applies a "map" LLMChain to each doc, then a "reduce" chain on the results
  • MapRerank: Ranks mapped doc responses by certainty score

Interface​

All document chains implement this base interface:

def combine_docs(docs: List[Document], **kwargs) -> Tuple[str, dict]

Which combines documents into a single string response.

Stuff Chain​

The Stuff chain directly concatenates documents into a prompt template and passes it to an LLMChain.

Key parameters:

  • llm_chain: The LLMChain called on the concatenated prompt
  • document_variable_name: Name for docs in the template

Refine Chain​

The Refine chain loops through documents, passing each to an LLMChain along with the latest answer. It iteratively refines the response.

Key parameters:

  • llm_chain: The LLMChain called per document
  • input_key: Key for non-document inputs
  • document_key: Key for the current document
  • output_key: Key for the latest answer

MapReduce Chain​

The MapReduce chain applies an LLMChain to each document (map step). It passes the mapped docs to a combine chain for the final output (reduce step).

Key parameters:

  • llm_chain: The mapping LLMChain
  • reduce_documents_chain: The reducing combine docs chain
  • document_variable_name: Doc variable name
  • collapse_documents_chain: Optional compression chain

MapRerank Chain​

The MapRerank chain runs an initial prompt on each document to get an answer and certainty score. It returns the top scoring response.

Key parameters:

  • llm_chain: The mapping LLMChain
  • reranker_chain: Generates certainty score
  • document_variable_name: Doc variable name