Changing the system Prompt for a ConversationalRetrievalChain

[Bennoo]

Hello,

I understand that a ConversationalRetrievalChain is doing a 2 steps processing.

1. Based on the question and the chat history, the chain is asking a LLM to rephrase the question including chat history elements
2. Adding to the context parts of the doc found in the documents and asking the question

I know how to change the rephrasing prompt phase (1st action) but I would like to change the way the doc and the question is submitted to the LLM (phase 2)
I would like to add some information in addition to the Retrieved documents and rephrase the prompt.

Does someone knows how to do it?

Many thanks

[simonfromla]

I think what you are looking for may be solved by passing the prompt in a dict object {"prompt": PROMPT} to the combine_docs_chain_kwargs parameter of ConversationalRetrievalChain.

Follow-up question: in "step 1", are you able to override the default behavior of passing in chat history?