Relevance score (_score) in Elasticsearch - Deep Dive

Overview - Relevance score (_score)

What is it?

Relevance score, shown as _score in Elasticsearch, is a number that tells how well a document matches your search query. It helps rank search results so the most relevant ones appear first. The score is calculated based on how often the search terms appear and how important those terms are in the whole collection. This makes searching smarter and more useful.

Why it matters

Without relevance scores, search results would be random or just sorted by date or name, which might not help you find what you want quickly. Relevance scores solve the problem of ranking results by importance, saving time and improving user experience. This is crucial for search engines, online stores, and any system where finding the best match fast matters.

Where it fits

Before learning about relevance scores, you should understand basic Elasticsearch concepts like documents, fields, and queries. After this, you can explore advanced search features like boosting, custom scoring, and query tuning to improve search quality.

Mental Model

Core Idea

The relevance score (_score) measures how well a document matches your search query, guiding the order of search results.

Think of it like...

Imagine looking for a book in a library. The relevance score is like a librarian who knows which books mention your topic the most and puts those books on top of the pile for you.

┌───────────────┐
│ Search Query  │
└──────┬────────┘
       │
       ▼
┌─────────────────────────────┐
│ Documents in Elasticsearch   │
│ (each with text and fields)  │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Calculate _score for each    │
│ document based on query      │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Sort documents by _score     │
│ Highest score on top         │
└─────────────────────────────┘

Build-Up - 7 Steps

1

Foundation

What is _score in Elasticsearch

Concept: Introduce the basic idea of _score as a number showing match quality.

When you search in Elasticsearch, each document gets a number called _score. This number shows how well the document matches your search words. Higher means better match. Elasticsearch uses this to sort results so you see the best matches first.

Result

You understand that _score is a number assigned to each document after a search.

Understanding that _score exists helps you see why search results are ordered the way they are.

2

Foundation

How _score affects search results

Concept: Explain that _score controls the order of results shown to users.

After Elasticsearch calculates _score for each document, it sorts all documents by this score from highest to lowest. This means documents that match your query better appear first. Without _score, results would be unordered or random.

Result

Search results are ranked by relevance, improving user experience.

Knowing _score controls ranking helps you trust and interpret search results better.

3

Intermediate

Factors influencing the _score value

🤔Before reading on: do you think every matching word counts the same for _score, or do some words matter more? Commit to your answer.

Concept: Introduce how term frequency, inverse document frequency, and field length affect _score.

Elasticsearch uses a formula called TF-IDF (Term Frequency-Inverse Document Frequency) to calculate _score. It counts how often your search words appear in a document (term frequency). It also checks how rare those words are across all documents (inverse document frequency). Rare words add more to the score. Shorter fields with the term get a higher score too.

Result

Documents with many rare matching words and shorter matching fields get higher _score.

Understanding these factors explains why some documents rank higher even if they have fewer matching words.

4

Intermediate

How query types affect _score calculation

🤔Before reading on: do you think all query types calculate _score the same way, or do some ignore _score? Commit to your answer.

Concept: Explain that some queries calculate _score differently or not at all.

Most queries like match or multi_match calculate _score using TF-IDF or newer models. But some queries like filters or term queries can ignore _score and just include or exclude documents. This means filters speed up search by skipping scoring. Combining queries can mix scored and unscored parts.

Result

You learn that _score depends on query type and can be skipped for performance.

Knowing which queries affect _score helps you design faster and more relevant searches.

5

Intermediate

Role of boosting in modifying _score

🤔Before reading on: do you think boosting changes the _score or just the order of results? Commit to your answer.

Concept: Introduce boosting as a way to increase or decrease _score for certain terms or fields.

Boosting lets you tell Elasticsearch to give more importance to some words or fields by multiplying their _score contribution. For example, boosting the title field means matches there count more than matches in the description. You can boost queries or individual terms to fine-tune ranking.

Result

Boosted documents get higher _score and appear earlier in results.

Understanding boosting shows how to control search relevance beyond default scoring.

6

Advanced

Custom scoring with function_score query

🤔Before reading on: do you think you can change how _score is calculated using your own rules? Commit to your answer.

Concept: Explain how to customize _score by adding functions like decay, random, or script scoring.

Elasticsearch lets you modify _score using the function_score query. You can add functions that increase or decrease scores based on document fields, dates, or random values. For example, you can boost recent documents or those with higher ratings. This custom scoring combines with the original _score to produce final ranking.

Result

Search results reflect both text relevance and custom business rules.

Knowing how to customize _score empowers you to build smarter, context-aware search experiences.

7

Expert

How Elasticsearch scoring evolved with BM25

🤔Before reading on: do you think Elasticsearch always used the same scoring formula? Commit to your answer.

Concept: Describe the transition from TF-IDF to BM25 as the default scoring algorithm.

Originally, Elasticsearch used TF-IDF to calculate _score. Since version 5.0, it switched to BM25, a more advanced formula that better models term frequency saturation and document length normalization. BM25 improves relevance by avoiding overvaluing very frequent terms and better handling long documents. You can still configure which similarity algorithm to use per field.

Result

More accurate and user-friendly search rankings with BM25.

Understanding scoring evolution helps you appreciate modern search quality and how to tune it.

Under the Hood

When you run a search, Elasticsearch breaks down your query into terms and compares them to terms in each document's fields. It calculates a score for each document by combining how often terms appear (term frequency), how rare those terms are across all documents (inverse document frequency), and how long the field is (field length normalization). This calculation uses the BM25 algorithm by default, which balances these factors to produce a relevance score. The scores are then used to sort documents before returning results.

Why designed this way?

The scoring system was designed to mimic how humans judge relevance: terms that appear often in a document but are rare in the whole collection are more important. BM25 replaced TF-IDF to better handle cases where terms appear many times and to normalize for document length, improving search quality. This design balances accuracy and performance, making search fast and relevant.

┌───────────────┐
│ User Query    │
└──────┬────────┘
       │
       ▼
┌─────────────────────────────┐
│ Query Parsing & Analysis     │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Term Matching in Documents   │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Calculate TF (term freq)     │
│ Calculate IDF (inverse doc freq) │
│ Apply Field Length Norm      │
│ Use BM25 Formula             │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Compute _score per Document  │
└──────┬──────────────────────┘
       │
       ▼
┌─────────────────────────────┐
│ Sort Documents by _score     │
└─────────────────────────────┘

Myth Busters - 4 Common Misconceptions

Quick: Does a higher _score always mean the document is more useful? Commit yes or no.

Common Belief: A higher _score always means the document is the best and most useful result.

Reality: A higher _score means better text match, but not necessarily more useful. Other factors like freshness, popularity, or business rules might matter more.

Why it matters: Relying only on _score can lead to showing irrelevant or outdated results first, hurting user satisfaction.

Quick: Do filters affect _score? Commit yes or no.

Common Belief: Filters change the _score of documents by boosting or lowering it.

Reality: Filters do not affect _score; they only include or exclude documents without scoring.

Why it matters: Confusing filters with scoring can cause unexpected search behavior and performance issues.

Quick: Is _score always between 0 and 1? Commit yes or no.

Common Belief: _score is a normalized value between 0 and 1.

Reality: _score can be any positive number and is not normalized; it depends on the query and index.

Why it matters: Assuming normalization can cause wrong interpretations or incorrect custom scoring.

Quick: Does boosting a term always guarantee top ranking? Commit yes or no.

Common Belief: Boosting a term guarantees documents with that term appear first.

Reality: Boosting increases _score contribution but does not guarantee top ranking if other factors outweigh it.

Why it matters: Overusing boosting can lead to unexpected rankings and poor search quality.

Expert Zone

1

BM25 parameters like k1 and b can be tuned per field to adjust term frequency saturation and length normalization, affecting relevance subtly.

2

The _score is a floating-point number that can be influenced by index statistics that change as documents are added or removed, causing score shifts over time.

3

Combining multiple queries with different scoring behaviors (e.g., bool with must and filter) requires understanding how scores are combined or ignored to avoid surprises.

When NOT to use

Relying solely on _score is not ideal when business rules or user context matter more. In such cases, use function_score queries, custom scripts, or external ranking systems like learning-to-rank models.

Production Patterns

In production, _score is often combined with filters for performance, boosted by business priorities, and adjusted with decay functions for freshness. Monitoring score distribution helps detect index changes affecting relevance.

Connections

Information Retrieval

Relevance scoring in Elasticsearch builds on classic information retrieval models like TF-IDF and BM25.

Understanding these models from information retrieval theory helps grasp why _score behaves as it does and how to improve search quality.

Machine Learning Ranking

Custom scoring and boosting in Elasticsearch can be seen as simple ranking models, which relate to more advanced machine learning ranking techniques.

Knowing _score basics prepares you to integrate Elasticsearch with ML-based ranking for smarter search.

Human Decision Making

Relevance scoring mimics how humans judge importance by weighing rare and frequent terms differently.

Recognizing this connection helps appreciate the design of scoring algorithms as attempts to model human judgment.

Common Pitfalls

#1 Expecting _score to be consistent across different queries or index states.

Wrong approach:

Running the same query multiple times and assuming _score values will not change.

Correct approach:

Understand that _score depends on index statistics and can vary; use relative ranking rather than absolute score values.

Root cause:

Misunderstanding that _score is dynamic and depends on the current index content and query context.

#2 Using filters when you want to boost relevance, expecting _score to change.

Wrong approach:

Applying a filter query to boost documents with a certain field value.

Correct approach:

Use boosting or function_score queries to modify _score instead of filters.

Root cause:

Confusing filtering (which excludes or includes documents) with scoring (which ranks documents).

#3 Overusing boosting on multiple fields or terms without testing impact.

Wrong approach:

Adding high boosts to many fields hoping to improve relevance.

Correct approach:

Carefully test and tune boosts to avoid skewing results and maintain balanced relevance.

Root cause:

Lack of understanding of how boosting affects combined _score and ranking.

Key Takeaways

The _score in Elasticsearch measures how well a document matches a search query, guiding result ranking.

It is calculated using algorithms like BM25 that consider term frequency, rarity, and field length.

Not all queries affect _score; filters exclude documents without scoring to improve performance.

Boosting and custom scoring let you adjust _score to reflect business priorities or context.

Understanding _score's dynamic nature and limitations helps build better, more relevant search experiences.