How relevance is calculated

BM25 intuition

Rare terms generally distinguish documents better than common ones. Repetition helps with saturation, and field length affects normalization. A short title containing the query can therefore beat a long description. BM25 does not know margin, stock or user intent unless those signals are modeled separately.

Fields and alternatives

Boost changes a matching clause’s relative contribution; it cannot create a missing match. dis_max emphasizes the strongest branch, while tie_breaker adds a portion of others. A phrase boost in should can prefer close matches without making the phrase mandatory. Evaluate weights over a query set, not a single successful example.

Explaining and comparing scores

_explain breaks down one document’s score for a query. Scores depend on the query, index statistics, fields and shard distribution. A threshold such as _score>10 does not mean the same quality for all queries. dfs_query_then_fetch can refine distributed term statistics with an extra round trip; it does not replace relevance evaluation.

{
  "query": {
    "match": {
      "name": "wireless"
    }
  }
}

Catalog reasoning example

Suppose the correct headphones have a short matching name while a popular accessory repeats headphones in a long description. Raw frequency is insufficient. Separate fields, inspect analysis and add a justified name or phrase contribution. Explain both scores. If the desired product is absent from candidates, weight tuning cannot fix it: retrieval precedes ranking.

Prioritize title matches

In multi_match, the caret notation name^3 gives the name field extra weight relative to an unboosted description. The example searches travel in both fields. Boost changes the contribution of a field; it does not guarantee that every title match universally outranks every description match regardless of content and corpus statistics.

Do not multiply a displayed _score yourself or compare it to a hardcoded target. The engine computes the score from the query and index. In the trainer, a relevance exercise compares document-order relationships and the requested construction. Its task may require a particular boost value even when another value looks plausible.

{
  "query": {
    "multi_match": {
      "query": "travel",
      "fields": [
        "name^3",
        "description"
      ]
    }
  },
  "size": 2,
  "_source": [
    "name",
    "description"
  ]
}

{
  "hits": {
    "total": {
      "value": 178,
      "relation": "eq"
    },
    "hits": [
      {
        "_id": "3",
        "_score": 4.8997154,
        "_source": {
          "name": "Sony Travel Speaker 3",
          "description": "Precision equipment for professionals"
        }
      },
      {
        "_id": "13",
        "_score": 4.8997154,
        "_source": {
          "name": "Apple Travel Speaker 13",
          "description": "Premium wireless sound for travel"
        }
      }
    ]
  }
}