TopolearnUsers judge search by whether the best result is first, which takes deliberate relevance tuning.
Elasticsearch scores full-text matches with BM25 by default. For each query term in each field, the score grows with:
- term frequency: how often the term occurs in the field (with diminishing returns, controlled by
k1), - inverse document frequency: rarer terms across the index count more,
- shorter fields: a match in a short title beats one in a long body (length normalization, controlled by
b).
Scores are relative, only comparable within one query, so never apply fixed score thresholds across queries.
Tuning tools, roughly in the order to try them:
- field boosts in
multi_match("title^3"), andtype: "best_fields","most_fields"or"cross_fields"depending on how fields relate, - phrase proximity: add a
match_phrasewithslopinshouldto reward words appearing together, - business signals with
function_scoreorscript_score: blend popularity, recency (gaussdecay on a date) or rating into the score;rank_featurefields are an efficient alternative, - synonyms at search time.
Use "explain": true to see each document's score breakdown, and evaluate changes against a set of real queries with judged results (the _rank_eval API) rather than eyeballing one query.
For hybrid lexical and semantic search, scores from different methods aren't comparable; combine result lists with reciprocal rank fusion (RRF) instead.
Example
GET /articles/_search
{
"query": {
"function_score": {
"query": { "bool": {
"must": [ { "multi_match": { "query": "solar battery storage",
"fields": ["title^3", "body"], "type": "best_fields" } } ],
"should": [ { "match_phrase": { "title": { "query": "battery storage", "slop": 1 } } } ]
} },
"functions": [
{ "gauss": { "published": { "origin": "now", "scale": "30d", "decay": 0.5 } } },
{ "field_value_factor": { "field": "likes", "modifier": "log1p", "missing": 0 } }
],
"score_mode": "sum",
"boost_mode": "multiply"
}
},
"explain": false
}Remember
- BM25: term frequency, rarity (IDF) and field length drive scores
- Scores are relative to a query; don't threshold them
- Boost fields, reward phrases, then blend business signals
- Measure changes with judged queries, not single examples
Watch out for
- Tuning boosts based on one query and making ten others worse.