Utilize RAPTOR on content retrieved?

[danieldekay]

I found out about RAPTOR which promises better contextual performance compared to basic chunked RAG.

https://github.com/parthsarthi03/raptor

Has anyone tried this yet?

[a-yuksel]

Thanks for sharing, it seems like a prominent RAG improvement.

I am using a similar idea manually for questions where I have a rough idea about the potential answer. The process is like this:

• When I want to ask a question to an LLM, I also generate a short context for RAG. It is typically a candidate answer to that question.
• Then rather than using the embedding of the original question, I use the embedding of the (question + answer) in RAG
• If you are asking same/similar questions repetitively, you can use the answers in the previous steps to create a RAG context
• This works well for me where the potential answers to the question can be identified beforehand, e.g. you have a question with a few potential answers, you have a rough idea about potential answers but don't know which one is the correct answer, plus don't know the details of the answer
• It will be restrictive and inefficient if the question is an open ended one. In that case, an LLM generated context like in RAPTOR can be more useful

I am a new user of GPTResearcher.
Does GPTResearcher allow adding a user-supplied context for RAG?

[ElishaKay]

Sup @a-yuksel

Correct you can set a DOC_PATH like so

This doc_path will be leveraged when report_type equals local or hybrid

If you run with docker, these report types will also be available on localhost:3000

A PR was opened this morning to persist the vectors generated by these local docs across runs 🥳

[a-yuksel]

Thanks