Preserved custom keys in RetrievalQA

[Atom-101]

Description

Currently RetrievalQA drops any extra keys in the prompt apart from the memory and user input. So if I want a prompt like this:

"""You are a friendly bot talking to a human.
Use the following pieces of context to answer the users question.
----------------
Context relevant to the question is: \n{context}
----------------
Recent conversation history is: \n{history}
----------------
You also have the following information:  \n{extra_context1}\n{extra_context2}
----------------

Human: {query}
AI:
"""

Here context is retrieved from the database, history comes from memory and query is the last input provided by the user. If some extra context has to be provided apart from conversation history and database, like extra_context1 and extra_context2, RetrievalQA cannot currently handle it.

With this change one can simply provide extra keys in in the .run() function of RetrievalQA. Like so:

qa = KwargsRetrievalQA.from_chain_type(
    llm=ChatOpenAI(),
    chain_type='stuff',
    retriever=retriever,
    verbose=False,
    chain_type_kwargs={
        "verbose": True,
        "prompt": prompt_template,
        "memory": memory,
    }
)

inp = input("Human: ")
query = inp

response = qa.run(
    query=query,
    history=memory,
    extra_context1=extra_context1,
    extra_context2=extra_context2,
)

Maintainer

@rlancemartin

[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	Jun 27, 2023 7:42am

[Atom-101]

@rlancemartin, @eyurtsev can you take a look?

[leo-gan]

@Atom-101 Hi , could you, please, resolve the merging issues? After that ping me and I push this PR for the review. Thanks!

[Atom-101]

This has been fixed in the current version of langchain. I am closing this. Thanks!