Open a recorded query. Trace the returned sentence to a source hash. Then inspect the baselines and the cases the support guard missed.
This explorer loads checked-in evaluation results. It does not run a search service or a model. The actual Python service runs on local loopback; its source and tests are public.
Top three chunks, one expected source file per query. Overlapping documents and author-written labels limit interpretation.
| Method | Expected-file hit | MRR @ 3 |
|---|
Character TF-IDF is lexical matching. Hybrid fusion does not establish general superiority; neither ranking is an independently validated semantic search system.
The guard flags low lexical coverage, unmatched numbers or missing citations. It does not understand a changed meaning reliably.
These probes are designed by the project author. Low risk is a review signal, not a factual verdict. Exact extract spans elsewhere are true by construction, not a separate faithfulness benchmark.
Loading recorded evidence.
Incremental Markdown ingestion, SQLite BM25, character TF-IDF, reciprocal-rank fusion and extractive citations. No model credential is required.
Source, one-command demo and CI ↗Ten owner-authored documents are a development corpus. No independent relevance labels, free-form model generation, public backend or production adoption is claimed.
Read the failure analysis ↗ · Raw recorded output ↗