Relevance and BM25

Key points

• BM25 has been the default similarity since 5.0, replacing the classic TF-IDF implementation.
• Three inputs decide a score: term frequency, inverse document frequency, and field length.
• Term frequency saturates, so the tenth occurrence of a word adds far less than the second.
• A rare term is worth more, because idf rises as the number of documents holding it falls.
• The same hit in a short title outranks it in a long body, thanks to length normalisation.
• k1 defaults to 1.2 and b to 0.75, and the explain API shows every factor.

Example

GET /articles/_search
{
  "query": { "match": { "title": "elasticsearch tuning" } },
  "explain": true
}

GET /articles/_explain/7
{ "query": { "match": { "title": "elasticsearch tuning" } } }

# each term contributes: idf * (tf saturated by k1 and normalised by b)

Rare terms in short fields score highest, and repetition saturates - run explain before tuning boosts.