Build an AI Fact Checker That Finds Sources for Any Claim in Python

This AI fact checker is a small web app for anyone who wants to see the sources behind a claim before repeating it. Paste one claim and get web sources that back it or push back on it, each with a link, a short summary that keeps only sentences tied to a returned source, and how the topic is discussed on X. It helps you find sources; you still read them and decide.

The problem

A claim shows up in a chat, a post or an AI answer, and checking it means opening a search engine, reading several pages and looking for who said what. A summary without links is hard to trust, and a list of links without context takes time to read. Most people want the few sources that matter, with the sentence that backs each point.

What the app does

• Takes one claim, 8 to 200 characters.
• Returns web sources, deduplicated, each with a link, a snippet and a cue label: supports, contradicts or unclear.
• Shows a summary in which every sentence cites one of the returned sources. Sentences that cite nothing are hidden and counted.
• Shows "No strong sources found" when fewer than two sources match most of the claim's key words.
• Adds X posts about the topic that contain all the search words, with authors shown as "a user".
• Can also search Wikipedia, and serves the same result as JSON from GET /api/check.

How it works

1. Check the input. Blank, too short or junk claims get a 400 before any paid call.
2. Find web sources. POST /desearch/ai/search with tools: ["web"], result_type: "LINKS_WITH_FINAL_SUMMARY" and count: 10 returns sources plus a summary that cites them.
3. See the conversation on X. GET /twitter with the claim's first 4 key words, sort: "Top" and count: 10, run at the same time as step 2.
4. Label and filter. Each source gets a keyword cue label, summary sentences are kept only when every link they cite is a returned source, and X posts are kept only when they contain all the search words.
5. Serve. The page and the JSON endpoint share the same code, with a 6-hour cache header so repeat claims are answered from a CDN.

The code

The core call, copied from the example's README. The key is read from the DESEARCH_API_KEY env var:

curl -s https://api.desearch.ai/desearch/ai/search \
  -H "Authorization: $DESEARCH_API_KEY" -H "Content-Type: application/json" \
  -d '{"prompt": "Octopuses have three hearts and blue blood", "tools": ["web"], "streaming": false, "result_type": "LINKS_WITH_FINAL_SUMMARY", "count": 10}'

The response holds search (each source's title, link, snippet and highlights) and completion, a summary that cites sources as [n](url).

The folder layout:

• factcheck/core.py: input checks, request bodies, source cleanup, cue labels, citation check, X post filter, rate limiter, spend cap and cache. No network.
• factcheck/client.py: the plain-HTTP Desearch client.
• factcheck/service.py: the /api/check logic shared by the local server and the hosted function.
• api/check.py, public/index.html, vercel.json: the hosted version.
• app.py: local server and CLI. test_core.py: offline tests.

Endpoints

• POST /desearch/ai/search
• GET /twitter

Example output

A saved run on 2026-10-18 for the claim "The Great Wall of China is visible from space with the naked eye" made 2 calls and cost $0.0055.

The top sources included BBC Sky at Night Magazine, Wikipedia and Britannica. One cited summary line reads: "The myth that it is visible from space predates space travel and originated from a 1932 Ripley's Believe It or Not cartoon, but has been debunked by astronaut observations and scientific explanations."

A made-up claim in the same run came back with every source labelled unclear, and the app showed "No strong sources found" instead of a summary.

Run it yourself

python3 -m pip install -r requirements-dev.txt   # only pytest, for the tests
export DESEARCH_API_KEY_FILE=/path/to/key         # or DESEARCH_API_KEY=your_key, or --key-file
python3 app.py                                    # web app at http://127.0.0.1:8000

For a one-off check from the command line:

python3 app.py --cli "Humans only use 10 percent of their brain" --wiki --json out.json --html out.html

• --wiki (or wiki=1 on /api/check) also searches Wikipedia, with one more call.
• --json and --html save the full result as data or as a page.
• --log-dir runs/today saves each API call with the key removed.

What it costs

Measured on 2026-10-18 from the X-Desearch-Cost-Usd header, and the same on 2026-10-19: $0.004 for the web search and $0.0015 for the X search, so $0.0055 per check, or $0.0095 with Wikipedia. Checks served from the cache cost nothing. The hosted version has a daily spend cap, $1.00 by default. See pricing.

Demo

The live demo is paused, and the sample shows a real saved run. Demo. Saved sample run.

Ideas to extend it

• Limit sources to sites you trust with include_domains, for example a science magazine and an encyclopedia.
• Add a browser button that sends the selected text on any page to /api/check.
• Check every factual sentence in a draft article and list the ones with no strong sources.
• Give an agent /api/check as a tool so it attaches sources before it answers.