I read a dozen different explanations of BM25 recently — from academic papers to “explain it like I’m 10” blog posts. Every single one converged on the same three ideas. That’s because BM25 really is that simple.

BM25 is a ranking algorithm. You type a query, it scores every document in the corpus, and the highest score wins. It’s been doing this since 1994. Google, Elasticsearch, Lucene, OpenSearch — they all use it or a close variant as their default scorer. Even RAG systems that rely on dense vector retrieval almost always include BM25 as part of a hybrid pipeline.

The “25” is just a version number. Like iPhone 25. Stephen Robertson and Karen Spärck Jones tried 24 versions before settling on this one.

The Three Rules

BM25 asks three questions about every document:

1. Does it have the words you searched for?

If a document mentions “dinosaur” 10 times, it’s probably more about dinosaurs than one that mentions it once. That’s term frequency — the most basic signal in search.

But BM25 is clever about it. Going from 0 to 1 mention is huge. Going from 1 to 2 is still significant. Going from 50 to 51 is almost nothing. The benefit levels off.

This is called TF saturation, and it’s what stops spammers from winning by repeating a keyword 500 times. BM25 caps the bonus. It’s like a teacher giving you one gold star for a good answer — not 100 gold stars for screaming it over and over.

2. Are the matching words rare?

Some words are in almost every document: “the”, “and”, “is”. Finding those tells you nothing. But if a document contains “triceratops” or “metabolomics” or “Part 11” — those are rare. A match on a rare word is worth massively more than a match on a common one.

This is Inverse Document Frequency (IDF). The analogy everyone lands on: if everyone in a building is wearing a blue shirt, that doesn’t help you find anyone. But if one person has a purple hat, that’s your clue. Rare words are the purple hat.

3. Is the document the right length?

A 10-page pamphlet that mentions “tree frog” three times is clearly about tree frogs. A 1,000-page encyclopedia that mentions “tree frog” three times? Probably not.

BM25 gives a fairness boost to shorter documents. A long document naturally contains more words, so it’ll match more terms just by being big — that’s not a signal of relevance. Document length normalization adjusts for this.

The Formula (In English)

Score = How Rare the Word Is × (Capped Word Count) / (Word Count + Page Length Penalty)

For every word in your query, BM25 calculates this score for every document. Add up the scores for all query terms. Highest total wins.

The actual formula:

Score(D, Q) = Σ IDF(qi) × [f(qi, D) × (k1 + 1)] / [f(qi, D) + k1 × (1 - b + b × |D|/avgdl)]

Two tuning knobs:

  • k1 (usually 1.2–2.0): How fast repetition stops helping. Higher = more credit for repeated terms.
  • b (usually 0.75): How much to punish long documents. 0 = ignore length entirely. 1 = full punishment.

Why It’s Not Dead

Every few years someone declares BM25 obsolete. Word2vec, BERT embeddings, dense retrieval, LLM-powered search — each was supposed to replace it. None did.

The reason is straightforward: BM25 is fast, interpretable, and extremely good at exact term matching. It doesn’t need GPU inference. It doesn’t hallucinate. It doesn’t require fine-tuning. It just counts words intelligently.

Vector search is great at finding semantically similar content — “electronic records shall maintain a secure chronological history” matches “audit trail requirements” because the meaning overlaps. But vector search struggles with exact terms: system names, document IDs, acronyms, regulatory section numbers. BM25 eats those for breakfast.

That’s why the most effective RAG retrieval architecture is hybrid:

Query → BM25 (exact terms) + Vector Search (meaning) → Combine → Rerank → LLM

BM25 catches the precise terminology. Vector search catches the paraphrases. Reranking merges them. The LLM gets the best of both worlds.

The Bottom Line

BM25 is three counting rules and two knobs. Rare words matter more. Repeated mentions help but level off. Short documents get a fairness boost. That’s it.

It’s been the backbone of information retrieval for over 30 years. If you’re building search, RAG, or any system that needs to find relevant documents — you need to understand BM25. Not because it’s fancy, but because it works.