Finding a page isn’t finding all the evidence.

This browser-only tool scores page rankings you supply. Ideal, partial and miss presets are illustrative. Use “Load actual BM25 result” to inspect recorded retrieval output. No model, PDF upload or API key is used.

Reference relevance labels

Grade 2: complete answer · Grade 1: supporting information · Unlisted: treated as 0 by the scorer.

What each metric asks

This follows the project's linear gain convention: grade / log₂(rank + 1), normalized by the ideal top-five ranking. It does not use exponential gains. Repeated pages are collapsed at their first occurrence before scoring, as the source evaluator does.

Recorded BM25 baseline

Run on 12 October 2026: 45 questions over the existing 104-chunk index, using CPU only.

Hit@5 88.89% · Recall@5 69.07% · nDCG@5 0.5677

Lexical retrieval only, without dense embeddings or section expansion. This does not reproduce the reported hybrid score. The calculator presets above remain illustrative.

All rankings and scores · Settings and checksums · Reproduction script

Failure analysis

Inspect the five weakest development cases, with missing pages, token matches, and clearly separated hypotheses.

Evidence and scope

The labels cover one 40-page MAX77751 datasheet: 33 development questions and 12 marked holdout in the original project. They were authored by the system builder. This is a transparent inspection tool, not an independent benchmark or a claim that these labels are exhaustive. The recorded baseline above is a separate local run against the existing index.

Once you inspect a holdout question here, do not treat it as unseen data for future tuning. A new evaluation needs fresh held-out documents and questions.

The underlying project combines EmbeddingGemma and BM25. Its engineering report describes table extraction fixes, boilerplate removal, and approaches that did not improve retrieval.

Label and scoring source snapshot: f83ff7abf531573227e5e1cf12c7a116bd69e447. Playground built October 12, 2026. No original datasheet pages or model weights are redistributed.