/llm-world-models

Make sure LLMs say true things.

Primary LanguagePythonMIT LicenseMIT

llm-world-models

Make sure LLMs say true things.

This repo is inspired by the tweet and blog post by Evan Conrad on using bayesian theory to build "world models" that may be used by LLMs to coerce factual answers w/ associated confidence levels.

run pip install -r requirements.txt and create a .env file with your OPENAI_API_KEY for use with Langchain.

The example_*.py scripts can be used as an entry point.

The agent directory currently has an experimental agent that can incrementally update its world model as it chats with a user.

Preview results (WIP)

frontend preview