Search Improvements

[blc16]

I've got experience in the LLM Search Infra space and have some ideas on improving speed, costs, and relevance search infrastructure for gpt researcher. Would love to discuss some ideas and learn about what the big painpoints pertaining to search have been! Happy to help contribute here

[ElishaKay]

Welcome @blc16

You're in luck.

This sprint we're thinking about:

• ⁠Add mongodb or another provider to pull data from DB instead of local or web sources

This storage mechanism will be important both for the report context and chat context

More context:

Abstractions behind GPTR context

GPTR chat module

I've started to explore this concept in this repo so that I can test the relevant Langchain and GPTR modules:

Contextual Awareneness chatbot

The high-level flow we're thinking about is: a user wants to go deep into a given body of local or web data:

• Use GPTR to embed the data

• embeddings should be stored in persistent storage (I believe this should be built-in if we pass a vector store while generating a report)

• leverage those saved embeddings to write reports from a specific angle (depending on the use case)

• create a personalized chat experience based on those insights

Feel free to start playing with these abstractions in a PR on this repo or another or share ideas below - we're happy to collab on this 👍

Seems we'll need a persistent vector store module in the gpt_researcher folder.

[blc16]

Very interesting! Yes having a local memory storage would be helpful. I will take a look at these docs you shared and report back

[blc16]

It sounds like you are thinking about the scenario where the user knows the body of data ahead of time. I am also interested in optimizing data that gets dynamically used/discovered throughout the research process. Would love to learn more about how people are using the researcher and where I could help. Would help to have a better idea of this to see if the area I am thinking about is actually important to common usecases

[ElishaKay]

Status update: here's a working example of running GPTR with MongoDB

It saves the logs, embeddings & report in MongoDB

[pbaldridgework]

MongoDB would be awesome! Are you thinking about integrating it with the front end app?

[ElishaKay]

Welcome @pbaldridgework

Correct - the long-term goal is that users will have plug-n-play components for a full-stack experience.

We're building out all the required pieces in the meantime & welcome contributions around it.

We also have this react component which is an abstraction on top of the NextJS app.

There is a chat box for followup questions which I believe hits the ${GPTR_API_URL}/chat