ResearchChain aims at achieving perpetual RAG building with the help of a self-prompting chain of AI researching agents.

Github: link

Abstract

Modern LLMs, including the +400B ones lack specialized knowledge required for solving complex issues, and answering deeply technical questions.
A comprehensive context-providing system should allow LLMs of all sizes to gain knowledge well outside the scope of the base language model.
A self-prompting approach will allow for any available downtime to be used for knowledge expansion in perpetuity.

Documentation

Running ResearchChain

Deploy every single worker, database and utility simultaneously

sudo docker-compose -f docker/docker-compose.yml up

Please note, that webui frontend has to be launched separately. see below

Running webui front-end user interaction

Frontend is launched separately to back end, run the following command to start it.

• go to frontend directory: cd webui/frontend/
• install dependencies: npm install
• start react project: npm run dev
• open http://localhost:3000/ in your browser

Accessing postgres database

• postgres can be accessed via pgAdmin, which is already included in the docker compose,
  so there is no need for any additional packages
• go to localhost:8081/browser/
• click add new server
• in name, write postgres
• go to connection tab
• in hostname/address write postgres
• in username write admin and in password write pass
• click save, the database should be immediately available
• there, you'll see connection statistics as well as the entire schema

Other notes

• The default models can be seen in the core/models/configurations folder.
  
• We're using python version 3.9
• environment.yml is the linux env, but for macOS (silicon) and windows there are other available
• Apple intel is not supported anymore, but you can still get it working by manually installing
  any missing package that comes up during the program execution.

This is a monorepo for both a tool, and apps for it:

Research Chain

A set of workers watching for new tasks 24/7.

This is the tool part of this project, alone it cannot schedule new tasks,
unless as a side effect of a previous task.
Consists of 3 infinitely scalable workers, which together with a 4th auto scheduling worker app,
create a complete loop.
This design allows for near infinite scalability potential.

Schedulers

Built in examples of how to interact with Research Chain, the ones
that we offer are WebUI, with plans for AutoScheduler and NewsScheduler coming soon.

These services automatically dispatch, analyze and manage Research Chain.

AutoScheduler and NewsScheduler should work alongside Web Interface,
to supply constant 24/7 knowledge and news analysis,
and to expand its knowledge base by scheduling crawls based on the provided areas of interest.

Link text
Link text