Below is a video that demonstrates an AI application called The Fact Checker. It identifies unique facts in arbitrary text and then finds public internet references confirming or disconfirming the individual facts. How? After identifying unique facts, the application generates Google search phrases to find relevant results. Then it loads the webpages in the search results and has the LLM decide whether the fact is verified by the web content. It continues checking web sources until a “Satisfaction Threshold” is reached, either confirming or disconfirming each fact.

The video has a demonstration and details the prompts and logic around the diagrams below.

First, Fact Identification. This takes an arbitrary block of text, and makes two LLM calls to extract unique facts, and then uses another LLM call to combine the lists. This is designed to help identify all of the facts.

Then there’s the Fact Verification process. This involves generating Google search phrases, then executing the searches to find relevant results. The HTML for the result pages are loaded and a prompt is used to determine relevance and whether the fact is confirmed or disconfirmed by the content. This is repeated until a Satisfaction Threshold is met.

That’s all for now. Next up, talking with databases.