RAG Silent Hallucinations: Agent Claims 'Not in Corpus' After Reading 0.6%

CupGlass540 · reddit · 2026-08-25

The author exposes a "silent hallucination" in RAG systems where an agent retrieves only a tiny fraction of the corpus (e.g., 0.6%) but answers as if it has read everything. This is critical for negative claims (e.g., "no policy exists"), where the model asserts facts about unread documents—a failure standard faithfulness scores miss because they only check answer grounding, not corpus coverage.

The author argues that claims of existence need only cite the relevant chunk, but claims of absence/universality require exhaustive coverage. They open-sourced notchecked, an MCP server that logs corpus coverage before answering and distinguishes between retrieval failures, assembly drops, and reasoning errors. The post highlights the need for "Context Recall" in production evals.

Original post →

More from coding & agent

coding & agent channel →