Keyword search ranks documents by the query’s exact terms. Semantic search ranks them by how close their meaning sits to the question. Good retrieval uses both.
Every RAG system, enterprise search box and AI agent depends on finding the right passage first. This guide shows how BM25 scores a document, how embeddings and nearest-neighbour search work, the queries each one gets wrong, what the benchmarks found, and how hybrid search and re-ranking combine them.
Updated · 9 sources
Scores each document by how often the query’s words appear in it, how rare those words are across the collection, and how long the document is.
Turns the query and every passage into vectors with an embedding model and returns the passages whose vectors sit nearest.
Keyword search (lexical search, usually ranked with BM25) finds documents that contain the query’s words, weighting rare words and repeated matches. Semantic search embeds the query and the documents as vectors and finds the ones closest in meaning, even with no word in common. In keyword vs semantic search, keyword wins on exact codes, names and identifiers; semantic wins on paraphrase. Hybrid search runs both and fuses the rankings.
BM25 — “Best Matching 25” — is the default ranking function in both Elasticsearch and OpenSearch. It adds up a score for each query term from three ingredients.
f(t,D) × (k1 + 1)
score(t, D) = IDF(t) × ───────────────────────────────────────
f(t,D) + k1 × (1 − b + b × |D| / avgdl)
f(t,D) how often t appears in D
|D| length of D; avgdl average length in the collection
k1 saturation (Elasticsearch default 1.2)
b length normalisation (Elasticsearch default 0.75)Semantic search replaces word matching with geometry. An embedding model maps text with similar meaning to nearby points, and an approximate nearest-neighbour index finds the closest points fast.
Documents are split into passages sized for the embedding model.
Each passage becomes a vector of a few hundred to a few thousand numbers.
Structures such as HNSW let search scale logarithmically with the collection.
With the same model — vectors from different models are not comparable.
Cosine similarity or dot product picks the top-k passages.
Semantic search vs vector search. The terms are often used interchangeably. Strictly, vector search is the mechanism — nearest neighbours in a vector space — and semantic search is the goal of matching meaning, which vector search over text embeddings is the usual way to reach.
Dense retrievers beat BM25 decisively on the question-answering data they were trained for. Tested on unfamiliar domains, BM25 holds up — and in Anthropic’s tests, adding BM25 to embeddings cut retrieval failures further.
Error codes, part numbers, clause references, invoice numbers, rare surnames, new product names an embedding model never saw.
Paraphrase (“cancel my plan” → “terminating your subscription”), vague questions, synonyms, questions phrased unlike the document.
Negation (“contracts without a penalty clause”), counts and totals, questions whose answer is spread across many documents.
Hybrid search runs a keyword query and a vector query side by side and merges the two result lists. The merge is the hard part, because a BM25 score and a cosine similarity are on different scales.
for each document d:
score(d) = Σ over result lists 1 / (k + rank(d))
k = 60 (Elasticsearch default rank constant)
d ranked 1st by BM25, 9th by vectors:
1/61 + 1/69 = 0.0309
d ranked 4th by both:
1/64 + 1/64 = 0.0313 ← winsFirst-stage retrievers compare a query with millions of passages, so they must be cheap. A re-ranker reads the query and each candidate together and rescores only the shortlist — slower per passage, far more accurate.
Late interaction sits in between. ColBERT (2020) encodes query and document separately, then compares them token by token. Its authors report it runs two orders of magnitude faster than BERT-based re-rankers; BEIR found re-ranking and late-interaction models the strongest zero-shot, at a higher compute cost.
Look at a hundred real queries from your users before choosing. Their shape — codes, natural questions, or both — decides the design.
What do your users’ queries look like?
A fixed hybrid pipeline still runs the same search for every question. A Botminds agent chooses: by meaning, by exact words, or both — under the same filters — and moves to records when the question is a count.
Read every way an agent searches in agentic RAG, and see how indexes are built in the Academy lesson search indexes and vectorization.
In keyword vs semantic search, keyword search finds documents containing the query’s words, ranked by a function such as BM25. Semantic search finds documents close in meaning using embeddings, even with no shared words. Keyword is better for exact codes and names; semantic for paraphrased questions.
BM25 (Best Matching 25) is a ranking function for keyword search. It scores a document by how often each query term appears, with diminishing returns, how rare the term is across the collection, and how long the document is. It is the default in Elasticsearch and OpenSearch.
Not across the board. Trained dense retrievers beat BM25 on the question-answering data they were trained for, but the BEIR benchmark found BM25 a robust baseline on unfamiliar domains. Combining both usually beats either one.
Hybrid search runs keyword and semantic search on the same query and merges the two result lists — by normalising and weighting their scores, or by reciprocal rank fusion, which rewards documents that rank high in either list.
A second pass in which a slower, more accurate model — usually a cross-encoder that reads the query and passage together — rescores the top candidates from the first-stage search, so the best few go to the user or the LLM.
Vector search is the mechanism: finding nearest neighbours among vectors. Semantic search is the aim: matching meaning. Semantic search over text is usually done with vector search over embeddings.
Both weight terms by frequency in the document and rarity in the collection. BM25 adds saturation, so repeated terms give diminishing returns, and length normalisation, so long documents are not favoured just for containing more words.
Both. Agents search by meaning and by exact words under the same filters, choose per search or run both for one question, and narrow by label, document, page or sheet. See agentic RAG.
Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.