Add option to RetrievalQA chains to use provided documents in call() and acall() methods

[benchello]

Description

This PR adds an option to the RetrievalQA chains to use provided documents in _call() and_acall() methods like: qa_chain({"query": question, "from_documents": documents}) where documents is List[Document]

Currently there is no way to reuse documents found from a similarity_search in a RetrievalQA. This PR enables the following sequence:

1. User embeds query and gets documents
2. User implements desired logic based on documents
3. User calls RetrievalQA without needing to re-embed query to find relevant documents.

This option will also be helpful if the user wants to gather documents in any custom way, like combining from different sources.

Usage

Example Usage:

qa_chain  =  RetrievalQA.from_chain_type(
     llm=llm,
     retriever=vector_store.as_retriever(),
)

documents = vector_store.similarity_search("Is the sky blue?")

if "blue" in documents[0].page_content:
	# Reuse documents in LLM call since they passed our condition
	response  =  qa_chain({"query": question, "from_documents": documents})
	
#True
response["source_documents"] == documents 

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Oct 17, 2023 0:31am

[baskaryan]

if you're not retrieving documents do you need to use the RetrievalQA chain at all or can you just use the combine_documents_chain directly?

[benchello]

Ah, sorry, I didn't see that as an option 😅 . Will close this and use that!