**BM25** and **dense embeddings** are two different ways to score how relevant a document is to a query, and they're good at different things.
BM25 is a classic keyword-matching algorithm. It scores documents based on how many of the query's exact terms appear, weighted by how rare those terms are across the corpus. It's great at finding documents that contain the literal words you searched for, especially acronyms, proper nouns, product names, and code identifiers.
Dense embeddings score documents by semantic similarity. The model embeds the query and each document into vectors, then measures cosine similarity. This captures meaning, so 'right to be forgotten' matches 'GDPR Article 17' even though zero words overlap.
Each method has blind spots. BM25 misses semantic paraphrases. Embeddings often miss rare terms and exact phrases. Hybrid retrieval runs both, then merges the rankings (usually via Reciprocal Rank Fusion), and consistently beats either alone.
**Hybrid retrieval** wins because BM25 and dense embeddings have **orthogonal failure modes**.
BM25 scores by exact-term overlap, weighted by inverse document frequency. It nails queries where the user types specific tokens that appear in the corpus, product names, acronyms, code identifiers, legal article numbers. Where it fails: paraphrases. A query for 'how to cancel my subscription' won't surface a doc titled 'membership termination'.
Dense embeddings invert this. The embedding model maps semantically similar text to nearby vectors, so paraphrases score high. Where embeddings fail: rare or out-of-vocabulary terms get smeared, and exact identifiers (`d_model`, `kCal`, `SKU-12345`) lose their distinctiveness in the geometric space.
Running both and fusing the results with **Reciprocal Rank Fusion** is the canonical fix:
$$\text{score}(d) = \sum_i \frac{1}{k + \text{rank}_i(d)}$$
with k=60 by convention. It's parameter-free, robust, and recovers the union of both methods' wins.
Production stacks (Pinecone hybrid mode, Weaviate hybrid, Elasticsearch ESRE) ship this as a primitive. Skipping the lexical leg is a common newcomer mistake.
Quick reference
TL;DR
BM25 catches exact-term hits embeddings miss; embeddings catch semantic paraphrases BM25 misses. Fusing the two ranks beats either alone on real corpora.
Imagine you're searching a library two different ways. One librarian (BM25) is great at finding books where your exact words appear, perfect when you ask for 'GDPR Article 17' or a specific product name. The other librarian (dense embeddings) is great at finding books that mean the same thing as your question, even if the words are different, perfect when you ask 'what about right to be forgotten?' and the book uses different language. Each librarian misses what the other catches. So you ask both, then combine their picks. That's hybrid retrieval. You get the precision of exact-term matching plus the recall of semantic similarity, with a fusion step to merge the two rankings into one.
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-lexical corpus (SKU lookup, exact-phrase search) BM25 alone is faster, cheaper, and sufficient; the dense leg adds latency and cost with little quality gain.
Pure-semantic chat over conversational docs BM25 contributes noise when users always paraphrase; dense-only retrieval is simpler to operate and tune.
Strict latency budget under 50ms p99 Running two retrieval legs plus fusion adds 10-30ms; one well-tuned method is faster and the quality delta is usually small.
Quick scan · 2–4 min
Everything important, quickly.
3 min: orthogonal failure modes + RRF formula and k=60 + production databases that ship hybrid + when to skip + how rerank composes on top.
Side-by-side
Concern
BM25
Dense embeddings
Hybrid (RRF)
Exact-term match
Strong
Weak
Strong
Paraphrase / synonymy
Weak
Strong
Strong
Out-of-vocab tokens
Strong (treated as rare)
Weak (smeared)
Strong
Index build cost
Cheap (inverted index)
Expensive (GPU embed)
Both costs
When it wins solo
SKU/identifier search
Conversational paraphrase
General-purpose RAG
Real-world examples
Real products, models, and research that use this idea.
1Pinecone hybrid retrieval mode fuses dense and sparse (BM25-style) scores in one query.
2Weaviate's hybrid search uses RRF by default and exposes the alpha parameter to bias toward dense or sparse.
3Elasticsearch ESRE ships RRF as a first-class operator combining classic BM25 with vector search results.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you tune fusion weights on your specific corpus?
Sign in or go Pro to see the approach for this follow-up.
QWhen does adding a reranker change the hybrid story?
Sign in or go Pro to see the approach for this follow-up.
QWhat changes when your corpus is multilingual?
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
Assuming embeddings always beat keyword search. On acronyms, proper nouns, and code identifiers, BM25 routinely outperforms dense retrieval.