Citation-first retrieval scaffold · Concept
TraceRAG
A self-checking .NET retrieval baseline that answers from local evidence, cites exact chunks and refuses unsupported questions.
Result — Every answer is backed by document- and chunk-level citations; unsupported questions return "insufficient evidence" rather than a guess.
Problem — retrieval systems confidently answer questions the evidence does not actually support, and the citation, if any, rarely maps to the exact text.
What I built — a compiling .NET concept scaffold for grounded document retrieval. It ingests bounded plain text, creates deterministic overlapping chunks, retrieves with transparent lexical scoring and returns a verbatim extract with document and chunk citations.
Architecture — a claim-level evaluator checks that every returned sentence exists in the cited evidence. When retrieval is weak or a claim is unsupported, the pipeline returns an explicit insufficient-evidence response instead of inventing an answer.
Tech — the baseline is intentionally local and inspectable: no hosted model, embedding service or hidden network call. Provider interfaces and a documented roadmap show where embeddings, richer generation and production controls belong next.
