Skip to main content
Check one claim against the pages that discuss it, and get back a verdict and a quoted sentence for each URL. Your code searches for the claim, then fetches each top result with a prompt. An LLM reads the page and returns only whether the page supports, contradicts or doesn’t mention the claim, with the sentence that decides it.

Recipe prompt

Paste this into your coding agent to have it build the recipe in your project:
Recipe prompt

What you build

When it runs

On demand, once per claim. A common place is after your model drafts an answer: split the draft into claims and check the ones that carry a number, a date or a name. Each claim costs 1 search and up to 3 fetches.

Code

This sample checks one claim against the top 3 pages and prints each URL with its verdict. It runs with no API key:
The TypeScript sample needs keenable 0.2.0 or later on npm, for the prompt option on fetch. The Go and Rust samples skip a page that fails to fetch and keep going. The curl tab runs the search, then checks each page the same way.

Example output

You see each URL, then the verdict and the sentence from that page:
Example output · results vary
The quotes keep the pages’ own typos. An LLM writes the verdict, so treat it as a pointer to the quote, and show the quote and URL to the reader.

Production notes

A fetch with a prompt sends the page to an LLM. If the page is longer than 500,000 characters, the LLM reads only the first 500,000. See Fetch parameters.

Next steps