fact checking with prompt chaining

This repo is a simple demonstration of using langchain to do fact-checking with prompt chaining. How it works:

• you ask your desired LLM a question
• the LLM generates an initial answer to the question
• the LLM self-interrogates what the assumptions were that went into that answer
• the LLM sequentially determines if each of these assumptions are true
• the LLM generates a new answer to the question, incorporating the new information

to run

Make sure you have langchain installed (pip install langchain). Then run

python3 fact_checker.py 'insert question here'

Be sure to wrap your question in quotes if you're passing it as a command line argument.

Alternatively, you can use the provided fact_checker.ipynb notebook.

example

Question: "What type of mammal lays the biggest eggs?"

Initial answer: The biggest eggs laid by any mammal belong to the elephant.

Assumptions made:

• The elephant is a mammal
• Mammals lay eggs
• Eggs come in different sizes
• Elephants lay bigger eggs than other mammals

Verification of assumptions:

• The elephant is a mammal: TRUE
• Mammals lay eggs: FALSE - Most mammals give birth to live young.
• Eggs come in different sizes: TRUE
• Elephants lay bigger eggs than other mammals: FALSE - Elephants do not lay eggs.

New answer: This question cannot be answered because elephants do not lay eggs and most mammals give birth to live young.

credits

Proof of concept by Jasper

Huge thanks to @hwchase17 and the langchain project!