[Experimental] Chat with history

[khoangothe]

I am trying to implement chat history to chat with the final report. The code is just experimental since I'm not sure whether this is where I should put the chat agent. The code will put the report and vectorize in a inmemory vectorstore to get the relevant chunks and I used default LangGraph React Agent. Would be cool if I could have some feedbacks on how to proceed! Thanks guys

Here's the UI (default to chatting after having a report, pressing the search symbol will conduct another research)

[ElishaKay]

Pretty awesome @khoangothe!

a) You'll want to think about the text in the homepage, as defined in "frontend/nextjs/components/InputArea.tsx".

Perhaps conditional text based on the current state?

b) I've played with the branch a bit & the results are pretty solid for the follow up questions

I'm curious whether we should leverage the new GPTR vector_store parameter:

researcher = GPTResearcher(
            query=query,
            report_type="research_report",
            report_source="langchain_vectorstore",
            vector_store=vector_store,
        )

[khoangothe]

@ElishaKay Got it thanks for the feedback! I'll see if I can push a change soon. Hoping can change the answer to some chat-like interface too.

Regarding the Vector Store, I saw your ideas in the VectorStore PR #781 and I think it would be very cool to have those implemented as well. I think we can store all of the related contents into a vectorstore and retrieve the information when ask for more context. But I struggle in how best to store new data since the vectorstore is inputted by the user which makes it a bit complicated

Case 1: Since we are storing every crawled info/documents in "vector_store", ideally the "vector_store" should be empty or contain related information so the new page won't contaminate the existing data and provide better queries.
Case 2: We just want to store all the results, so user can have an all knowing bot that can answers related to his/her previous reports.
Case 3: Should we add a default vectorDB (InMemory or any local vector DB) in case the user does not provide a vectorstore.

I would love to hear your opinion. Not sure which case the user will want, but I think I can quickly change the InMemoryVectorStore in the code to the inputted VectorStore after we got this features! Would be cool if we can have a discussion or you could help with some ideas about this!

[ElishaKay]

Good points @khoangothe

Case 1: Agreed. If we add a chat_id or report_id to the vectorstore records, I guess we'd be able to filter vectorstore searches by chat_id

Case 2: personalization would be great - but it depends on whether we want to have user authentication & multi-user logic within this repo, or whether the GPTR community should build that out independently on a case by case basis - not yet decided - therefore less relevant for this PR

Case 3: Faiss seems like a good candidate since it seems to come out of the box with a pip package. Postgres could be good as well for SQL querying. Have a look here at the Langchain Vector Stores

[ElishaKay]

P.S. @assafelovic is the lead on backend integrations, so would be worthwhile to add him into the loop regarding the architecture plan on this PR.

@assafelovic:

a) are we ready for the Postgres integration?

b) if so, regarding the data model (table structure) I'm thinking of:

• Chat

• Message (1 or more within chats)

• • chat_id
• • type = enum(query, logs, report, followup_question...)

c) would love to get your input regarding the backend logic for followup questions (see chats above)

Jah bless ✌

[khoangothe]

@ElishaKay Thanks man, I just added a PR for VectorStore #838 . I think if we want to proceed with the vector_store ideas, we need to implement this first. Then we can continue with this branch to chat with the crawled information in the vector db

[ElishaKay]

@khoangothe

Good thinking!
I've merged #838 into the AI Dev Team flow in #819
I'd like to also merge this PR into the same AI Dev Team branch & do extensive testing.

The main use cases I'd like to test & verify are:

• leveraging these additions in all the report_types supported by main NextJS frontend on port 3000
• the current plan of the AI Dev Team flow

I've also invited you as a collaborator on my GPTR fork, so you can feel free to push up commits to the AI Dev Team branch if inspiration sparks

[khoangothe]

you can feel free to push up commits to the AI Dev Team branch if inspiration sparks

Thanks man, I do have some ideas on how to leverage the vector store. I'll read through the AI Dev Team branch and see what I can add

[assafelovic]

Hey @khoangothe any update on this? Would love see this merged!

[khoangothe]

Thank @assafelovic , my plan was to have the vector_store pr merge first and then utilize the vector_store for this pr. With that functionality, I can store crawled data into an in memory_db, then we can chat with the data in DB (not just the report).
If you prefer, I can also merge the vector_store to this pr and then push a commit altogether to chat with the crawled data. I'm working on it right now and can push a commit in either case soon!

[assafelovic]

Got it @khoangothe apologize in advance that we've refactored quite a lot of the codebase to be better written so it created some conflicts with the recent PRs

[khoangothe]

@assafelovic no worries! I think I'll just close this PR and make another one instead of resolving the merge conflict so it's easier to keep track of the change history. Will do when I finish merging