Two complementary approaches to information retrieval: sparse (lexical and keyword-based, such as BM25 and TF-IDF) and dense (neural and embedding-based, such as vector search). Modern systems combine both for best results.
Sparse retrieval represents documents as high-dimensional sparse vectors of term frequencies. It is fast, interpretable, and great at exact keyword matches. Dense retrieval represents documents as low-dimensional dense embeddings from a neural model. It captures semantic meaning and handles paraphrase, but misses rare-keyword precision. Hybrid retrieval combines both (often via Reciprocal Rank Fusion) and consistently outperforms either alone on real-world benchmarks like BEIR.
Building a hybrid search system that runs both BM25 and vector search in parallel, then fuses results, for 15-25% better recall than either alone on diverse query types.
The sparse/dense tradeoff is foundational to modern search design. Neither alone is sufficient for production-grade retrieval, and hybrid is the new default.
Need help implementing this in your business?
Get Started