Backend Development › NoSQL & Other Data Stores

Relevance Scoring (BM25)

Ranking search results by how well they match.

Also known as: relevance scoring, BM25, search ranking

Relevance scoring decides the order of search results, not just which documents match. When many documents contain the query terms, ranking by how well each matches is what makes search useful — the best result comes first.

The classic algorithm is BM25 (a refinement of TF-IDF), and it balances three intuitions:

• Term frequency (TF) — a document that contains the query term more often is more relevant — but with diminishing returns (the tenth mention adds little).
• Inverse document frequency (IDF) — a rarer term is more informative. Matching “photosynthesis” counts far more than matching “the”.
• Document length — a match in a shorter document counts more than the same match in a long one, so long documents don’t win by sheer size.

score = IDF × term-frequency-saturation × length-normalisation
higher score → higher in results

Search engines apply this per field (title matches weigh more than body), combine fields, and let you tune boosts. Relevance is relative to the index’s statistics, which is why result order can differ after reindexing.

The classic mistakes:

• Assuming matching is ranking. Which documents match and how they’re ordered are separate concerns. A search that returns matches in arbitrary order is poor UX; ranking matters.
• Ignoring field weighting. A title or tag match is usually more meaningful than a body match. Weight fields to reflect that.
• Chasing one relevance score. Human relevance is fuzzy; measure with real queries and user behaviour, not by staring at scores. Tune against examples.
• Forgetting that scores are index-relative. Adding documents changes IDF, so scores and order shift over time. Don’t hard-code expectations of exact scores.
• Using relevance for filtering. If a result must be included (an exact ID match), handle that with filters/boosts rather than relying on it ranking high naturally.
• Overlooking synonyms and analysers. What matches is decided by the analysers; relevance can’t fix a query that never matched.

How to use it: rely on the engine’s BM25 as a strong default; boost important fields; combine with exact-match and fuzzy behaviour; and validate with real search queries and click data. Relevance is what turns a list of containing documents into a ranked, useful result set — the payoff of the inverted index.