Best Matching 25. A sparse, term frequency based retrieval scoring function (the 25th iteration in the Okapi BM family). It scores a document by summing per-term contributions weighted by term frequency, inverse document frequency, and document length normalization. It excels at exact-token matching (rare names, IDs, error codes) where dense embeddings smooth over the signal. Hybrid retrieval pairs BM25 with dense embeddings and fuses results via Reciprocal Rank Fusion (RRF).
**BM25** stands for Best Matching 25. The 25 is the version number; it's the 25th iteration of the Okapi BM family of ranking functions from the 1990s. Despite being thirty years old, it's still the strongest sparse-retrieval baseline and a core component of modern production search.
What BM25 actually computes. For a query and a document, BM25 sums per-term contributions. For each query term, the contribution depends on three factors.
- Term frequency (TF): how often the term appears in the document. More occurrences mean a higher score, with diminishing returns.
- Inverse document frequency (IDF): how rare the term is across the whole corpus. Rare terms (a specific name, a product code) are worth more than common terms (the, and).
- Document length normalization: longer documents are penalized so they don't beat short documents just by accumulating more matches.
The data structure is an inverted index: a map from each term to the list of documents containing it. No embeddings, no neural network, no GPU. Just counting and a small amount of math per term.
BM25 excels at exact-token matching: rare names, product IDs, error codes, function names, anything where the literal string is the signal. Dense embeddings smooth over these because they map related concepts to nearby vectors, which can dilute the exact-token signal.
The modern pattern: hybrid retrieval. Run BM25 and a dense embedder side by side, then fuse the two ranked lists with Reciprocal Rank Fusion (RRF). The hybrid usually beats either alone because BM25 catches the exact matches dense misses, and dense catches the semantic matches BM25 misses.
**BM25** stands for Best Matching 25, the 25th iteration in the Okapi BM family of sparse retrieval scoring functions developed at City University London in the 1990s. It scores a document against a query by summing per-term contributions over the query's terms.
The scoring formula has three factors per term:
$$\text{score}(D, q) = \sum_{t \in q} \text{IDF}(t) \cdot \frac{f(t, D) \cdot (k_1 + 1)}{f(t, D) + k_1 \cdot (1 - b + b \cdot \frac{|D|}{\text{avgdl}})}$$
The components.
- `IDF(t)` is inverse document frequency, rewarding rare terms. Computed as `log((N - n_t + 0.5) / (n_t + 0.5))` where N is corpus size and `n_t` is the number of documents containing the term.
- `f(t, D)` is the raw term frequency in the document.
- `k_1` (typically 1.2-2.0) controls how fast term-frequency saturates.
- `b` (typically 0.75) controls how much document length normalizes the score; `|D|` is the document length and `avgdl` is the corpus average.
The data structure is an inverted index. Implementations like Elasticsearch, OpenSearch, Lucene, Tantivy, and the newer pure-Rust LanceDB-BM25 all use this same scoring family.
Where BM25 shines: exact-token matching. Rare entity names, product SKUs, error codes, function identifiers. Anything where the literal string is the signal. These are dense-embedding blind spots because contextual embedders smooth over surface form to capture meaning. A query for 'CVE-2024-3094' finds nothing semantically related; it needs the exact token, and BM25 surfaces it instantly.
The production pattern is hybrid retrieval. Run BM25 and a dense embedder in parallel, retrieve top-100 from each, fuse with Reciprocal Rank Fusion (RRF). The RRF score is `\sum_i 1/(k + rank_i)` with `k=60`. The hybrid typically improves recall@k by 5-15% over best single retriever and is the default RAG architecture in 2026.
Quick reference
TL;DR
BM25 (Best Matching 25) is a sparse term-frequency retrieval function that scores documents by TF, IDF, and length normalization; it pairs with dense embeddings in hybrid retrieval.
Picture an old-school librarian who scores every book by counting how often your search words appear inside it, downweights words that appear in almost every book (like the, and, of), and accounts for how long each book is so a giant encyclopedia doesn't beat a short article by sheer volume. That counting system is BM25. It's been around since the 1990s and is shockingly hard to beat for exact-match cases like rare names, product codes, and error strings. Modern systems usually run BM25 and a smart semantic search side by side and combine their answers.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Pure semantic retrieval over abstract or paraphrased queries where the user's wording differs from the corpus. BM25 requires lexical overlap. Queries like 'how do I stop my model from making things up' won't match documents using 'reduce LLM hallucinations'; dense embeddings handle this gracefully.
Cross-language retrieval where the query and document are in different languages. BM25 is monolingual at its core. Cross-language retrieval needs multilingual embeddings (BGE-M3, Cohere multilingual) or a translation step.
Pure-image or pure-code retrieval where token-level matching is too noisy or ill-defined. Modality-specific encoders (CLIP for images, voyage-code-3 for code) outperform BM25 substantially on these surfaces; BM25 is text token centric.
Quick scan · 2–4 min
Everything important, quickly.
6 min: spell out the acronym, walk through the three scoring factors, explain where BM25 outperforms dense embeddings, describe the hybrid + RRF pattern, and place BM25 in the 2026 production stack.
Real-world examples
Real products, models, and research that use this idea.
1Elasticsearch and OpenSearch implement BM25 as the default scoring function for full-text search queries.
2Anthropic's contextual retrieval architecture for Claude pairs BM25 with dense embeddings and a contextual chunk prefix to boost both retrievers.
3Apache Lucene (the engine underneath Elasticsearch, Solr, and many others) ships BM25 as its default scoring since version 6.0.
4Modern hybrid retrieval stacks at Pinecone, Weaviate, and Qdrant all support BM25 (or BM25-equivalents like SPLADE) alongside dense vectors with built-in RRF fusion.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhat's the difference between BM25 and TF-IDF, and why does it matter?
Sign in or go Pro to see the approach for this follow-up.
QHow does Reciprocal Rank Fusion work, and why is it the standard fusion algorithm for BM25 + dense?
Sign in or go Pro to see the approach for this follow-up.
QWhen would SPLADE or ColBERT replace BM25 in a hybrid stack?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating BM25 as obsolete because dense embeddings exist. It's still strong at exact-token matching and the workhorse of hybrid retrieval, where BM25 + dense + RRF beats either alone.