LettuceDetect

Why we included this project

Most hallucination checkers return a single clean-or-flagged verdict, which is not much help when a long answer has one fabricated sentence buried in it. LettuceDetect works at the token level and returns exact character spans, so you can see precisely which claim in a RAG response, a coding agent's reply, or tool output is unsupported. It ships fast local encoder models for low-latency inference alongside generative detectors that can label the type of each flagged span, and it covers RAG prose, multilingual QA, and code-agent answers. For teams building grounded AI systems, it works as a verification step in evaluation pipelines or as a runtime guard that catches invented APIs, wrong values, and behavior the request never asked for.