Search two ways
Keyword search catches exact terms like daemon.json; meaning-based search catches paraphrases.
Results are merged and re-ranked by a third model.
An AI assistant that knows when it doesn't know
Most chatbots guess.
This one checks its evidence first.
Ask anything about Docker. The AI searches the official docs, a second AI
fact-checks what it found, and you get either an answer with sources
or an honest that's not in the docs
.
Or try one:
You asked
- 🔎SearchLook through 2,594 doc snippets
- 🧑⚖️Fact-checkDoes the evidence really answer it?
- ✏️RephraseOnly if the evidence was weak
- 💬AnswerAnswer with sources, or refuse
For engineers: the raw agent trace
Does the fact-checking actually help?
We tested it on 70 questions, including 20 trick ones.
Same documents, same answer-writing model. Plain AI search is the standard recipe: search once, then answer from whatever comes back. Polygraph adds a sharper search and the fact-check-and-retry loop.
How it works
Three ideas, one honest assistant.
A second AI checks the evidence
Before anything is written, a fast "grader" model decides whether the snippets actually contain the answer, and says why.
Retry, then refuse
Weak evidence? The agent rewrites its search (up to twice). Still nothing? It says so, instead of inventing an answer.
LangGraphGroq · gpt-oss-120b / 20bBM25 + bge-small embeddings
Cross-encoder rerankerONNX RuntimeRagas · DeepEvalVercel