BM25 (1994) › Legacy and Modern Relevance
History of Language AIAll chapters BM25 (1994) › Legacy and Modern RelevanceOne section · Sections in this chapter6 / 7 ↓
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
BM25 remains a baseline for neural ranking and a lexical retriever in some retrieval-augmented generation systems. Later methods often compare against it or combine its exact-term signal with learned semantic representations.
Foundation for Neural Ranking
Neural ranking models that emerged in the 2000s and 2010s learned non-linear relationships between queries and documents. Many still used BM25 as a first-stage retriever, a feature, or a baseline because its lexical signal complemented learned representations.
Some neural rankers include BM25 scores as features or re-rank candidates retrieved by BM25. Such systems combine exact lexical matches with semantic signals learned from data.
BM25 exposes term frequency, document length, and term rarity as separate factors. Neural rankers can instead learn how to combine lexical and semantic features from training data.
Retrieval-Augmented Generation
Retrieval-augmented generation (RAG) systems retrieve documents to supply context for generation. BM25 can serve as the retriever or as a first stage before neural re-ranking.
In a RAG pipeline, BM25 can retrieve candidate documents quickly from a large knowledge base. Neural retrievers offer semantic matching, while BM25 contributes an inexpensive exact-term signal. Production systems may use either method or combine them in hybrid retrieval.
BM25 scores rank candidate documents, but they are not calibrated probabilities. A RAG pipeline may use those rankings directly or combine them with neural retriever and re-ranker scores.
Benchmark and Evaluation Standard
BM25 is a common baseline for evaluating new ranking and retrieval methods. Its implementation and behavior are well understood, so a comparison can show whether a more complex method improves on a strong lexical ranker.
As a benchmark, BM25 provides a stable reference point for comparing retrieval methods. It can also produce a reasonable baseline without extensive tuning, provided that preprocessing and evaluation conditions are held constant.
Because BM25 is widely implemented, researchers can reproduce a familiar baseline across collections. Meaningful comparisons still require reporting preprocessing, parameter choices, and evaluation metrics.