Scripted Ranking Architecture

Elasticsearch 7.0 replaced the legacy function score mechanism with the script_score query. This module-based approach separates ranking logic from default relevance models like BM25, enabling developers to construct complex scoring pipelines. By combining mathematical transformations, geographic projections, and temporal weights, search outcomes can be precisely aligned with domain-specific requirements.

Core Arithmetic Operations

Direct field access uses the doc['field'].value syntax. Standard arithmetic operators execute natively within the Painless scripting environment.

{
  "script": {
    "source": "(doc['unit_cost'].value * 1.15) + doc['shipping_fee'].value"
  }
}

Saturation and Sigmoid Transformations

To cap the enfluence of extreme values, saturation() flattens score increments once a threshold is reached. Conversely, sigmoid() applies a smooth S-curve transition centered at a specified pivot, modulated by an exponent value.

{
  "query": {
    "script_score": {
      "query": { "match_all": {} },
      "script": {
        "source": "saturation(doc['engagement_score'].value, 85)"
      }
    }
  }
}

{
  "query": {
    "script_score": {
      "query": { "match_all": {} },
      "script": {
        "source": "sigmoid(doc['interaction_count'].value, 40, 0.7)"
      }
    }
  }
}

Decay Calculations

Spatial, numerical, and temporal fields support three decay profiles: Linear, Exponential, and Gaussian. Each profile requires an origin point, scale duration/width, optional offset, and a target decay ratio.

// Geographic decay example
{
  "script": {
    "source": "decayGeoGauss('42.3,-71.0', '50km', '10km', 0.8, doc['branch_coords'].value)",
    "lang": "painless"
  }
}

// Numerical decay example
{
  "script": {
    "source": "decayNumericExp(150, 25, 5, 0.6, doc['conversion_rate'].value)",
    "lang": "painless"
  }
}

// Temporal decay example
{
  "script": {
    "source": "decayDateLinear('2023-01-15T10:00:00Z', '3d', '0', 0.4, doc['publish_ts'].value)",
    "lang": "painless"
  }
}

Stochastic Injection and Statistical Modifiers

Rank diversification often requires injected randomness. Unseeded calls generate unique scores per request, while seeded variants guarantee deterministic outputs for pagination consistency. Logarithmic or power-based multipliers further refine field weighting.

{
  "script": {
    "source": "Math.random() * 0.95"
  }
}

{
  "script": {
    "source": "randomReproducible(doc['version_id'].value.toString(), 75)"
  }
}

{
  "script": {
    "source": "doc['rating'].value * Math.pow(params.weight_factor, 2)",
    "params": { "weight_factor": 3.5 }
  }
}

Hot-Reloading Synonym Dictionaries

Maintaining lexical mapping tables traditionally forced costly full reindexing cycles. Starting in 7.3, Elasticsearch supports live analyzer updates via the _reload_search_analyzers endpoint. By declaring a synonym filter as updateable: true and placing reference files within the node's configuration directory, modified rules activate immediately. Note that this operation only affects future tokenization; historical documents retain their original analytical state.

PUT /retail_catalog
{
  "settings": {
    "index": {
      "analysis": {
        "filter": {
          "dynamic_synonyms": {
            "type": "synonym_graph",
            "synonyms_path": "config/lexical_mappings.txt",
            "updateable": true
          }
        },
        "analyzer": {
          "custom_synonym_analyzer": {
            "tokenizer": "standard",
            "filter": ["lowercase", "dynamic_synonyms"]
          }
        }
      }
    }
  },
  "mappings": {
    "properties": {
      "product_category": {
        "type": "text",
        "search_analyzer": "custom_synonym_analyzer"
      }
    }
  }
}

POST /retail_catalog/_reload_search_analyzers

Case-Agnostic Exact Term Matching

Traditional term queries enforce strict binary equality, which often fails against inconsistently cased identifiers. The case_insensitive boolean now permits direct normalization suppression within leaf queries, eliminating the need for upstream data transformation or separate keyword subfields.

{
  "query": {
    "term": {
      "account_handle": {
        "value": "SysAdmin_Group",
        "case_insensitive": true
      }
    }
  }
}