Hallucination detector
Catch invented numbers, names and claims before your users do, with the sentence quoted.
Checks a model's output against the source it was given and quotes the sentence that went wrong. Built for support bots, summaries and RAG answers, where one invented price or policy is a real incident.
Try it live
Real calls to Jev. Pick an example or paste your own.
Pick an example or paste your own input, then run it.
Under the hood
Every question and rule, in the open. Fork it to change any of them.
What it asks
Reads source and answer, then asks 4 questions in one call.
- Q1States an unsupported factYes / no quotes evidence
Does the
answerstate a fact that thesourcedoes not support? (Rewording and common knowledge such as 'Paris is in France' don't count. Any specific number, name, date, price or policy that isn't in the source does.) - Q2Invented numbersYes / no quotes evidence
Does the
answerinclude a number, date, price or quantity that doesn't appear in thesource? - Q3Invented namesYes / no quotes evidence
Does the
answername a person, product, company, place or URL that thesourcenever mentions? - Q4Contradicts the sourceYes / no quotes evidence
Does the
answersay something thesourcecontradicts?
How it decides
Rules run in code, top to bottom. The first one that matches gives the verdict.
- 1If Contradicts the source ≥ 0.7Hallucinated
- 2If States an unsupported fact ≥ 0.75Hallucinated
- 3If Invented numbers ≥ 0.75Hallucinated
- 4If Invented names ≥ 0.75Hallucinated
- 5If Contradicts the source ≥ 0.4Suspectreview
- 6If States an unsupported fact ≥ 0.4Suspectreview
- 7If Invented numbers ≥ 0.45Suspectreview
- 8If Invented names ≥ 0.45Suspectreview
- ·DefaultGrounded
Call it from your code
Every template runs as-is with an API key, no setup: POST /api/v1/judges/hallucination-detector/run.
curl -X POST https://tryevally.com/api/v1/judges/hallucination-detector/run \ -H "Authorization: Bearer $EVALLY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "input": { "source": "The Pro plan includes 50,000 judgments a month and email support. Annual billing saves two months.", "answer": "The Pro plan includes 50,000 judgments a month and phone support, and it costs $19 a month if you pay yearly." }}'Seen in the wild
Community builds on Jev that use this pattern. Not affiliated with Evally.
More in AI output quality
- AI output quality
RAG answer grader
Is the answer grounded in the retrieved source, and does it answer the question?
PassNeeds reviewFail5 questions - AI output quality
Citation checker
Does the cited passage actually support the claim it's attached to?
SupportedPartly supportedUnsupported+12 questions - AI output quality
Summary checker
Faithful, complete and free of added opinions: the three ways summaries go wrong.
Good summaryNeeds editsUnfaithful5 questions - AI output quality
Support reply grader
Check an AI support reply against the customer's message and your policy before it's sent.
Ready to sendEdit firstEscalate+15 questions - AI output quality
Claim vs evidence
Does the evidence support the claim, contradict it, or not settle it? With the sentence that decides.
SupportedNot settledContradicted3 questions