VeriQuote
VeriQuote - verifies LLM citations: quote-in-source and claim-support checks
Last verified:
What is VeriQuote?
VeriQuote is a tool for verifying citations in AI-generated answers by combining two checks: a matcher that confirms quotes actually appear in their cited sources, and an LLM-based judge that confirms the quotes support the claim being made. It targets developers and researchers building or evaluating RAG/answering systems with citations.
VeriQuote pricing
Pricing model: Freemium
VeriQuote pros
- Catches both invented quotes (matcher, 100% on tested set) and unsupported claims (judge, up to ~89% caught)
- Interactive demo with example citations, editable text, and adjustable matching threshold
- Benchmarked against human annotators and multiple judge models with reproducible repository
- Works with any OpenAI-compatible endpoint; API key stays in the browser tab
VeriQuote cons
- Judge check requires an external model endpoint and user-supplied API key
- Each check misses what the other catches; neither alone is sufficient
- No visible hosted API or production integration; primarily a demo/benchmark tool
Frequently asked questions about VeriQuote
How does VeriQuote check citations?
It uses a matcher that compares the quote against the source text and a judge model that assesses whether the quote supports the claim, catching different failure types.
Do I need an API key?
Yes, the judge runs on any OpenAI-compatible endpoint (such as OpenRouter) using your own API key, which stays local in your browser tab.
What is the default matcher threshold?
0.40, which in benchmarking kept 100% of faithful quotes while rejecting 100% of missing quotes.