Business ranking

Candidates and preferences

First establish mandatory constraints such as access, stock and category. Rank eligible candidates afterward. A popular product in the wrong category should not appear merely because of a large business signal. Every signal has a range and distribution: sales counts from zero to a million should not be blindly added to text scores.

Functions and normalization

function_score combines the original score and functions through boost_mode and score_mode. field_value_factor uses numeric fields; logarithmic modifiers compress large ranges but need valid inputs. Decay functions decrease contributions with distance from dates, numbers or locations. rank_feature is designed for prepared positive signals and needs its own mapping.

Scripts and a second stage

script_score is flexible but expensive across many candidates; scores must be valid non-negative numbers. Rescore applies expensive ranking to a top window on each shard, so a small window limits possible improvements. Check not only the first product but also rare queries and missing feature values.

{
  "query": {
    "function_score": {
      "query": {
        "match": {
          "name": "headphones"
        }
      },
      "field_value_factor": {
        "field": "rating",
        "modifier": "sqrt",
        "missing": 1
      },
      "boost_mode": "multiply"
    }
  }
}

Catalog reasoning example

With text scores 8 and 6 and sales 10 and 100000, naive multiplication overwhelmingly favors the second product. Logarithmic or saturation transformations compress the range, and bounded influence preserves strong text evidence. Do not tune from two numbers: inspect the sales distribution, missing history, new products and exact-model queries.

Relevance under constraints

Separate three responsibilities: must describes text relevance, filter describes eligibility, and must_not describes exclusions. Our example searches travel using name^3 and description, requires stock and price below 800, then excludes preorder. Keep multi_match in the scored must path rather than inside a local or outer filter.

Do not add a field sort when the goal is a relevance ranking. A boost placed on an unrelated clause or inside an aggregation does not give the actual text query the intended weighting. Inspect the complete matching set and the relative order of documents rather than hardcoding the numeric score of one hit.

{
  "query": {
    "bool": {
      "must": [
        {
          "multi_match": {
            "query": "travel",
            "fields": [
              "name^3",
              "description"
            ]
          }
        }
      ],
      "filter": [
        {
          "term": {
            "in_stock": true
          }
        },
        {
          "range": {
            "price": {
              "lt": 800
            }
          }
        }
      ],
      "must_not": [
        {
          "term": {
            "status": "preorder"
          }
        }
      ]
    }
  },
  "size": 2,
  "_source": [
    "name",
    "price",
    "status",
    "in_stock"
  ]
}

{
  "hits": {
    "total": {
      "value": 34,
      "relation": "eq"
    },
    "hits": [
      {
        "_id": "23",
        "_score": 4.8997154,
        "_source": {
          "name": "Dell Travel Speaker 23",
          "price": 100,
          "in_stock": true,
          "status": "discontinued"
        }
      },
      {
        "_id": "53",
        "_score": 4.8997154,
        "_source": {
          "name": "Dell Travel Speaker 53",
          "price": 100,
          "in_stock": true,
          "status": "discontinued"
        }
      }
    ]
  }
}