Relevance: BM25, Boosting and function_score

How scores are computed and tuned

Elasticsearch scores text matches with BM25 by default: a term counts more when it is frequent in the document (with saturation, controlled by k1, default 1.2), rare across the index (inverse document frequency) and found in a shorter field (length normalisation, controlled by b, default 0.75). Scores of clauses are combined by the bool query. Tune relevance with field boosts (title^3), should clauses that reward signals, boosting queries that demote (negative boost) without excluding, and function_score, which multiplies or adds business signals such as popularity (field_value_factor), recency (decay functions like gauss) or fixed weights for matching filters. "explain": true shows how a score was built.

Text relevance plus popularity and freshness

Kibana Dev Tools console syntax; send the same requests with curl or a client library.

GET /products/_search
{
  "query": {
    "function_score": {
      "query": { "multi_match": { "query": "rain jacket", "fields": [ "name^3", "description" ] } },
      "functions": [
        {
          "field_value_factor": {
            "field": "sales_count",
            "modifier": "log1p",
            "missing": 0
          }
        },
        {
          "gauss": {
            "created_at": { "origin": "now", "scale": "30d", "decay": 0.5 }
          }
        },
        { "filter": { "term": { "featured": true } }, "weight": 2 }
      ],
      "score_mode": "sum",
      "boost_mode": "multiply"
    }
  }
}

Measure relevance changes

Collect a set of real queries with judged good results and compare before and after each tuning change (the ranking evaluation API, _rank_eval, can help). Tuning by anecdote often fixes one query and breaks ten.

Quick check: Under BM25, which term contributes more to the score, all else equal?

• A term that appears in every document
• A term that is rare across the index
• A term in a much longer field
• A term only in a filter clause