Advanced Search Relevancy in Elasticsearch

To improve search relevance in Elasticsearch, tune scoring before you add complexity: weight the right fields, boost business signals with function_score, add search-time synonyms, then layer in vector search and Learning to Rank, measuring each change against real queries. This guide walks through each technique with working examples.

Search relevancy is the cornerstone of a great search experience, and most gains come from getting the basics right before reaching for machine learning.

How does Elasticsearch score results?

Elasticsearch uses the BM25 algorithm by default, but you can customize scoring to match your specific needs.

Custom Scoring with Function Score

{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query": "elasticsearch tutorial",
          "fields": ["title^2", "content"]
        }
      },
      "functions": [
        {
          "filter": { "term": { "featured": true } },
          "weight": 1.5
        },
        {
          "field_value_factor": {
            "field": "popularity_score",
            "factor": 1.2,
            "modifier": "log1p"
          }
        }
      ],
      "score_mode": "multiply",
      "boost_mode": "multiply"
    }
  }
}

How do you add synonyms?

Synonyms help capture user intent even when they use different terminology. Elastic recommends applying them at search time with the synonym_graph filter: it handles multi-word synonyms correctly, and you can change the rules without reindexing.

Creating a synonym set

Store the rules with the synonyms API, so you can update them later without touching the index:

PUT _synonyms/my-synonyms
{
  "synonyms_set": [
    { "id": "es", "synonyms": "elasticsearch, elastic search" },
    { "id": "db", "synonyms": "database, db, datastore" },
    { "id": "find", "synonyms": "search, find, lookup" }
  ]
}

Using it in a search analyzer

{
  "settings": {
    "analysis": {
      "filter": {
        "my_synonyms": {
          "type": "synonym_graph",
          "synonyms_set": "my-synonyms",
          "updateable": true
        }
      },
      "analyzer": {
        "my_search_analyzer": {
          "tokenizer": "standard",
          "filter": ["lowercase", "my_synonyms"]
        }
      }
    }
  },
  "mappings": {
    "properties": {
      "content": {
        "type": "text",
        "analyzer": "standard",
        "search_analyzer": "my_search_analyzer"
      }
    }
  }
}

Updating the synonym set through the API reloads the analyzers that use it automatically.

How do you expand and weight queries?

1. Multi-Match Queries

Use different match types for various scenarios:

{
  "query": {
    "multi_match": {
      "query": "machine learning",
      "fields": ["title^3", "tags^2", "content"],
      "type": "best_fields",
      "tie_breaker": 0.3
    }
  }
}

2. Boosting Recent Content

Give preference to newer content:

{
  "query": {
    "function_score": {
      "query": { "match_all": {} },
      "functions": [
        {
          "gauss": {
            "date": {
              "origin": "now",
              "scale": "30d",
              "decay": 0.5
            }
          }
        }
      ]
    }
  }
}

How do you personalize search results?

User Behavior Tracking

Track user interactions to improve relevancy:

Click-through rates
Dwell time
Conversion rates
User preferences

Implementing Personalized Search

The {{...}} placeholders below are filled in by a search template (or your application) with the user's query and profile:

{
  "query": {
    "function_score": {
      "query": {
        "bool": {
          "must": [
            { "match": { "content": "{{query}}" } }
          ],
          "should": [
            { "terms": { "category": ["{{user_interests}}"] } }
          ]
        }
      },
      "functions": [
        {
          "filter": { "terms": { "author": ["{{followed_authors}}"] } },
          "weight": 1.3
        }
      ]
    }
  }
}

How do you test relevancy changes?

Implement A/B testing to measure relevancy improvements:

Define metrics: Click-through rate, conversion rate, user satisfaction
Create variants: Different scoring algorithms or query structures
Measure results: Statistical significance and business impact
Iterate: Continuously improve based on results

What are the advanced techniques?

1. Learning to Rank

Learning to Rank (available since Elasticsearch 8.12, on certain subscription levels) uses a trained model to re-rank the top results of a normal query:

Collect judgment data from user interactions or human ratings
Define features as templated queries (for example, BM25 scores on different fields) plus custom signals
Train a gradient boosted tree model such as XGBoost's XGBRanker outside Elasticsearch, and upload it with Eland
Apply it as a rescorer on the top N hits of each shard

{
  "query": {
    "multi_match": {
      "fields": ["title", "content"],
      "query": "the quick brown fox"
    }
  },
  "rescore": {
    "learning_to_rank": {
      "model_id": "ltr-model",
      "params": { "query_text": "the quick brown fox" }
    },
    "window_size": 100
  }
}

2. Semantic Search

Leverage vector search for semantic understanding. For production workloads, use approximate kNN on a dense_vector field rather than a script_score query, which scores every matching document and does not scale to large indices (it is still useful for exact scoring over a small, pre-filtered subset):

{
  "knn": {
    "field": "content_vector",
    "k": 10,
    "num_candidates": 100,
    "query_vector_builder": {
      "text_embedding": {
        "model_id": "my-text-embedding-model",
        "model_text": "elasticsearch tutorial"
      }
    }
  }
}

Combining this with a BM25 query gives you hybrid search. How the two scores get combined matters more than most teams expect; we measured it in Hybrid search scoring, explained.

How do you monitor relevancy?

Track key metrics:

Query performance: Response times and throughput
User engagement: Click-through rates and session duration
Business metrics: Conversion rates and revenue impact
Search quality: Manual relevancy assessments

Best Practices

Start simple: Begin with basic relevancy tuning before implementing complex features
Use analytics: Leverage search analytics to understand user behavior
Test continuously: Regular A/B testing ensures ongoing improvement
Monitor performance: Balance relevancy improvements with query performance
Gather feedback: Collect user feedback to validate relevancy improvements

Conclusion

Advanced search relevancy requires a combination of technical expertise and user understanding. By implementing these techniques and continuously measuring their impact, you can create a search experience that truly serves your users' needs.

Sources

Elastic documentation, Search with synonyms
Elastic documentation, Synonym graph token filter
Elastic documentation, Learning To Rank
Elastic documentation, Search using Learning To Rank
Elastic documentation, kNN search

Advanced Search Relevancy Techniques with Elasticsearch