TF-IDF and BM25 scoring in Elasticsearch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

What is the BM25 score output for a simple match query?

Given an Elasticsearch index with documents containing a field text, what is the expected BM25 score output for the query {"match": {"text": "quick brown"}} on the document {"text": "the quick brown fox"}?

Assume default BM25 parameters and that the document is the only one in the index.

Elasticsearch

{
  "query": {
    "match": {
      "text": "quick brown"
    }
  }
}

A A positive float score close to 1.0

B Zero, because the query terms are stop words

C A negative score, since BM25 can produce negative values

D An error because BM25 does not support match queries

💡 Hint

BM25 scores documents based on term frequency and inverse document frequency. The score is always positive.

BM25 scoring produces a positive relevance score based on how well the document matches the query terms. Since the document contains both "quick" and "brown", the score will be positive and typically close to 1.0 for a single document.

🧠 Conceptual

intermediate

Which factor does TF-IDF consider that BM25 improves upon?

TF-IDF and BM25 are both scoring algorithms used in Elasticsearch. Which aspect does BM25 improve compared to classic TF-IDF?

A BM25 uses only exact phrase matches instead of individual terms

B BM25 removes inverse document frequency to simplify scoring

C BM25 ignores term frequency to speed up scoring

D BM25 adds document length normalization to better handle longer documents

💡 Hint

Think about how document length affects relevance scores.

BM25 improves upon TF-IDF by adding document length normalization, which prevents longer documents from unfairly scoring higher just because they have more terms.

🔧 Debug

advanced

Why does this Elasticsearch query return zero scores with BM25?

Consider this Elasticsearch query using BM25 scoring:

{
  "query": {
    "bool": {
      "must": [
        {"match": {"content": "apple orange"}}
      ]
    }
  }
}

The index contains documents with the field content, but all returned documents have a score of 0. What is the most likely cause?

A The field <code>content</code> is mapped as <code>keyword</code> type, so BM25 scoring does not apply

B BM25 scoring always returns zero for boolean queries

C The query syntax is invalid and causes scoring to reset to zero

D The documents do not contain the terms "apple" or "orange"

💡 Hint

Check the field mapping type for content.

BM25 scoring applies only to text fields analyzed with full-text analyzers. If content is mapped as keyword, it is not analyzed and BM25 scoring does not work, resulting in zero scores.

📝 Syntax

advanced

Which query correctly disables BM25 and uses classic TF-IDF in Elasticsearch?

Elasticsearch uses BM25 by default. Which query snippet correctly disables BM25 and enables classic TF-IDF scoring for the body field?

A { "match": { "body": { "query": "search text" } }, "similarity": { "type": "classic" } }

B { "match": { "body": { "query": "search text", "similarity": "classic" } } }

C { "match": { "body": "search text" }, "similarity": "classic" }

D { "match": { "body": { "query": "search text" } }, "similarity": "classic" }

💡 Hint

Similarity settings are usually configured at the field mapping or query level inside the field object.

Option B correctly sets the similarity parameter inside the match query for the body field, enabling classic TF-IDF scoring. Other options either place similarity outside the field or use invalid syntax.

🚀 Application

expert

How to tune BM25 parameters to favor term frequency over document length?

You want to adjust BM25 scoring in Elasticsearch to give more importance to how often a term appears in a document, and less importance to the document length. Which parameter settings achieve this?

A Set both <code>b</code> and <code>k1</code> to 0

B Set <code>b</code> close to 1 and <code>k1</code> to a lower value like 0.5

C Set <code>b</code> close to 0 and <code>k1</code> to a higher value like 2.0

D Set <code>b</code> to 0.75 and <code>k1</code> to 1.2 (default values)

💡 Hint

Remember b controls length normalization and k1 controls term frequency saturation.

Setting b close to 0 disables length normalization, so document length has less effect. Increasing k1 increases the impact of term frequency, making repeated terms more influential.