Practical Optimization of Elasticsearch Search Ranking

Elasticsearch (ES) is a popular open‑source full‑text search engine. While it enables rapid construction of a search platform, its generic configuration often yields unsatisfactory results for specific content domains. This article, authored by Cao Yi, a Tencent Application Development Engineer, shares practical experiences in optimizing ES search ranking.

The default ES ranking is based on a relevance score calculated from the query keywords and document content. Because ES is a generic engine, it cannot fully understand the semantics of the indexed data, especially for Chinese text, which requires plugins and preprocessing. Consequently, platform‑specific optimizations are essential.

1. Optimizing ES Query DSL

After building the index, the first step is to refine the Query DSL. The article discusses several techniques:

Using multi_match for quick full‑text queries, but recognizing its limitations.

Adding bool filters (e.g., tags, categories) to narrow results without affecting relevance scoring; filters are cached and improve query speed.

Understanding the difference between must (contributes to scoring) and filter (does not).

Ensuring that term queries are applied to keyword‑type fields rather than text fields.

2. Boosting Phrase Matching

To improve the weight of exact phrase matches, the article recommends using match_phrase , which requires all tokens to appear in the correct order. The slop parameter can relax the order constraint. Combining match_phrase with match inside a bool should clause yields higher scores for documents that preserve the phrase order.

3. Applying Boost

Boost can increase the weight of specific fields (e.g., title) or entire query clauses. The boosted score equals the default score multiplied by the boost factor. Recommendations include boosting high‑quality fields and giving higher weight to match_phrase than to plain match .

4. Using function_score for Static Scoring Factors

Beyond dynamic relevance, static factors such as document freshness, popularity, quality, and promotional weight can be incorporated via function_score . The five supported function types are:

script_score – custom script.

weight – constant multiplier.

random_score – random value.

field_value_factor – uses a field’s value (e.g., sqrt(1.2 * doc['likes'].value) ).

Decay functions ( linear , exp , gauss ) – smoothly decrease scores based on distance from an origin (e.g., time, location).

Examples illustrate how to configure field_value_factor and decay functions with parameters origin , scale , decay , and offset .

5. Final DSL Example

The article presents a comprehensive DSL that combines the above techniques, demonstrating a well‑tuned query that delivers satisfactory search results.

6. Optimizing the Relevance Algorithm (Similarity)

ES’s default similarity is BM25, a probabilistic model that superseded TF‑IDF. Tuning the two adjustable parameters, k1 (term‑frequency saturation) and b (field‑length normalization), can significantly affect scores. For collections with uneven document lengths, lowering b (e.g., to 0.2) reduces the impact of length on relevance.

7. Recommendations

Prioritize data quality and DSL tuning before adjusting similarity.

Avoid excessive plugins (synonyms, pinyin) early on.

Monitor user behavior (repeat queries, pagination) to gauge search satisfaction.

Use the _explain API to analyze bad cases and iteratively improve the ranking.

Conclusion

Building a professional search platform with Elasticsearch requires systematic search tuning, including DSL optimization, relevance algorithm adjustment, and static scoring factors. The practices described provide actionable guidance for achieving higher relevance and better user experience.