Understanding Elasticsearch Scoring: Lucene Scoring Functions, Query Boosting, and Function Score Queries

Elasticsearch relies on Lucene's practical scoring function to rank documents. The score is calculated as score(q,d) = queryNorm(q) · coord(q,d) · Σ ( tf(t in d) · idf(t)² · t.getBoost() · norm(t,d) ) (t in q) , where each component reflects term frequency, inverse document frequency, field boosts, and field-length normalization.

Term frequency ( tf(t in d) = √frequency ) gives higher weight to terms that appear more often in a document. Inverse document frequency ( idf(t) = 1 + log ( numDocs / (docFreq + 1)) ) reduces the weight of common terms across the corpus. Field-length norm ( norm(d) = 1 / √numTerms ) favors shorter fields, such as titles.

Query normalization ( queryNorm(q) ) attempts to make scores comparable across different queries, while coordination ( coord(q,d) ) rewards documents that contain a higher proportion of the query terms.

Beyond the default TF/IDF scoring, Elasticsearch offers query-time boosting, where a boost parameter can increase the importance of specific query clauses. The constant_score query can assign a fixed score to matching documents, ignoring TF/IDF.

The function_score query allows custom scoring functions to be combined with the original relevance score. Common functions include:

weight : applies a simple, unnormalized boost (e.g., weight: 2 ).

field_value_factor : modifies the score based on a numeric field. Example: { "query": { "function_score": { "query": { "match": { "name": "游泳馆" } }, "field_value_factor": { "field": "comment_score", "modifier": "log1p", "factor": 0.1 }, "boost_mode": "sum" } } }

random_score : adds a random component, optionally seeded for reproducible results.

Decay functions (linear, exp, gauss): score based on distance from an ideal value. Example using a Gaussian decay on location: { "query": { "function_score": { "query": { "match": { "name": "游泳馆" } }, "gauss": { "location": { "origin": { "lat": 31.227817, "lon": 121.358775 }, "offset": "5km", "scale": "10km" } }, "boost_mode": "sum" } } }

script_score : runs a custom script to compute the score. Simple Groovy script to boost swimming venues: return doc['category'].value == '游泳' ? 1.5 : 1.0

Multiple functions can be combined in a single function_score query. For example, to recommend venues within 5 km that have parking, high user ratings, and a random tie‑breaker: { "query": { "function_score": { "filter": { "geo_distance": { "distance": "5km", "location": { "lat": $lat, "lon": $lng } } }, "functions": [ { "filter": { "term": { "features": "停车位" } }, "weight": 2 }, { "field_value_factor": { "field": "comment_score", "factor": 1.5 } }, { "random_score": { "seed": "$id" } } ], "score_mode": "sum", "boost_mode": "multiply" } } }

The article concludes that while simple query combinations often yield satisfactory results, achieving optimal relevance may require iterative tuning of boosts, custom scripts, or decay functions to address specific business needs such as time‑based or distance‑based scoring.