Enhanciar builds three-layer system to catch AI citations that mislead engineers
A tool called Enhanciar has been developed to address a common failure in AI coding assistants, where models cite real source files but paraphrase content those files do not actually contain. The system works in three layers: building a structured wiki from repo files at ingest, forcing the AI to cite specific wiki handles in every answer, and then running a deterministic verifier that checks each claim against the source before the response is delivered. A key design choice is that structural claims — such as whether a function exists or a route is exposed — are verified using AST lookups via tree-sitter, requiring no model calls or API tokens. Only prose claims that cannot be resolved structurally are escalated to a language model for re-reading and grading as supported, partial, unsupported, or unverifiable. The verifier is explicitly barred from marking a claim as supported simply because it could not complete a check, ensuring engineers can trust or openly question every cited line.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in