Skip to content

LearnRank & assemble

Relevance score

The number a search system attaches to each result to order it. Its scale and meaning depend entirely on the method that produced it.

In DocSlurpGlossary entry

On this page

Every search result carries a number used to sort it. What that number means depends on what produced it:

Producer Typical range What a higher number means
BM25 0 to unbounded More rare query words, more often, in a shorter passage
Cosine similarity −1 to 1 Embeddings point in a more similar direction
Cross-encoder Model-specific The model judges the pair more relevant
RRF Small fractions, such as 0.032 Ranked higher by more lists

None of these is a probability that the passage answers the question. Most only promise that a higher score ranked higher for this query. A BM25 score of 14.2 can be a strong match for one query and a weak one for another, and a cosine of 0.70 may be noise for one embedding model and a close paraphrase for another.

The gotcha

Thresholding on a raw score (“drop anything under 0.75”) quietly assumes the score means the same thing across queries and models. Usually it doesn’t. To threshold, calibrate first: measure where unrelated text scores for your model and corpus, and set the cut above that (see score normalization and relevance floor).

DocSlurp returns three separate fields. score is a position score, 1 / (1 + index), so results score 1, 0.5, 0.333 and so on, whatever their raw match strength (core/server/src/search/session-execution.ts). relevance, from 0 to 1 and present when it was measured, comes from the retrieval branches before fusion and means the same thing on every query; it drives the relevance floor (core/server/src/search/relevance.ts). confidence is a high, medium or low band from a weighted mix of relative score, rank and citation detail, not a calibrated probability (core/server/src/search/confidence.ts).

In DocSlurp

In DocSlurp

Search returns score as a position score, 1 / (1 + index), after fusion and reranking (core/server/src/search/session-execution.ts). A separate relevance field (0 to 1, comparable across queries) drives the relevance floor (core/server/src/search/relevance.ts), and confidence is a heuristic band, not a probability (confidence.ts).

Sources

Jump to