Δ is candidate minus baseline. Positive is better on this collection. Raw BM25F and RRF scores have different units; compare their ranks and relevance metrics instead.
On narrow screens, scroll tables horizontally. Keyboard users can focus a table region and use the arrow keys.
Download comparison JSON — full inputs, configuration options, all 22 rankings, grades and unrounded metrics; no upload or account.
Inspect a query
All 22 queries
Averages can hide individual regressions. Inspect the underlying ranks, including relevant documents missing from the top 10.
How to read this
- P@5 divides relevant hits by 5, including empty slots. R@10 and nDCG@10 use the top 10. MRR and MAP use the full retrieved ranking, up to 100 results. Gain is the relevance grade, with log₂(rank + 1) discount.
- Grades: 3 directly answers, 2 substantially relevant, 1 partly relevant, 0 not relevant under this collection’s judgments. Unlisted documents receive grade 0. These are one author’s labels.
- Configurations were fixed before evaluation. A difference can come from a single query; this page makes no significance or general-quality claim.
- Toy hybrid vectors capture spelling overlap, not meaning. They are hashed character trigrams, not semantic embeddings. No model, provider or internet retrieval is called.
- Queries run as written, without the main demo’s last-word prefix completion. BM25F scores are comparable only within the same query and configuration; RRF scores have different units.
- The JSON download records the full synthetic inputs and results for reproduction. It is not authenticated provenance. This page neither imports private data nor saves browser storage.