What semantic property does BM25 capture that dense embeddings tend to miss?

**BM25** is the classic keyword retrieval scorer. It looks at which exact words your query and the document have in common, weighted by how rare each word is across the whole corpus. A document that contains a rare word from your query scores high; a document that just contains common words like the and of scores low. Dense embeddings work differently. The embedding model converts your query and each document into a vector that captures meaning. Two pieces of text get similar vectors when they say similar things, even if they use different words. That is why dense retrieval beats BM25 on paraphrase queries. The catch is that the embedding model is trained to smooth over surface form. When your query is mostly a rare identifier, like a part number, a function name, or a customer ID, that rare token gets averaged into a vector that no longer points at exactly that string. BM25 keeps the literal token as the signal. So BM25 wins on exact rare-term lookups and dense wins on meaning. Real systems combine both.

**BM25** captures exact term and rare-token matching that dense embeddings smooth over. The scorer rewards documents containing the query's rare terms, weighted by inverse document frequency. A query containing a part number, an error code, or a proper noun pulls back documents that contain that exact token. Dense embeddings are trained to map paraphrases to nearby vectors. That is great when the user asks how do I cancel my subscription and the doc says steps to unsubscribe. It is bad when the user asks for ticket INC-48291 and the embedding model maps every five-character incident-style string into roughly the same neighborhood. The two retrievers fail on opposite queries, which is exactly why hybrid retrieval works. You run BM25 and a dense ANN search in parallel, then fuse the two ranked lists. Reciprocal Rank Fusion is the standard combiner because it works on ranks rather than raw scores, sidestepping the un-calibrated score problem between two different scoring systems. In modern RAG stacks (LlamaIndex, LangChain, Vespa, Weaviate, Pinecone hybrid) the default first-stage retriever is hybrid for this reason. Pure dense is the exception, used only when the corpus is short and paraphrastic.