Hybrid Search

Definition

Hybrid search combines two or more retrieval signals — most commonly sparse (lexical/term-based) and dense (semantic/embedding-based) — to produce results that outperform either approach alone.

The intuition: sparse retrieval excels at exact keyword matching; dense retrieval excels at semantic understanding. Their failure modes are largely complementary.

Why Combine?

Retrieval Type	Strengths	Weaknesses
Sparse (BM25/SPLADE)	Exact terms, proper nouns, codes	Vocabulary mismatch, no synonyms
Dense (Bi-Encoder)	Semantics, paraphrases, intent	Weak on rare terms, slower
Hybrid	Best of both	More complex, harder to tune

Fusion Strategies

Reciprocal Rank Fusion (RRF)

Most common, parameter-free:

RRF_score(d) = Σ 1 / (k + rank_i(d))

where k=60 is a constant that smooths rank differences.

Advantages: No score normalization needed, robust, no training required.

Linear Score Combination

final_score = α × sparse_score + (1 − α) × dense_score

Requires score normalization (scores from different systems aren’t comparable).

Re-ranking

Retrieve N candidates from each system → merge → re-rank with Cross-Encoder.

Common Implementations

SPLADE + Bi-Encoder

• SPLADE for learned sparse retrieval (term expansion)
• Bi-Encoder (e.g., sentence-transformers) for dense semantic
• Fusion: RRF or learned combiner

BM25 + Dense

• BM25 for lexical baseline (no ML required)
• Dense encoder for semantic lift
• Popular in production (Elasticsearch, OpenSearch)

Vespa Hybrid

• Native ColBERT + BM25 hybrid in Vespa
• Jo Kristian Bergum’s work on ColBERT embedder

Implementation in Elasticsearch

# Two-phase: BM25 + ELSER (sparse) or BM25 + dense
GET /products/_search
{
  "query": {
    "bool": {
      "should": [
        {"match": {"text": "query"}},           # BM25
        {"text_expansion": {                     # ELSER sparse
          "ml.tokens": {"model_id": ".elser_model_1", "model_text": "query"}
        }}
      ]
    }
  }
}

Wormhole Vectors as Hybrid Bridge

Trey Grainger’s Wormhole Vectors concept extends hybrid search by identifying vectors that bridge multiple retrieval spaces (sparse, dense, behavioral). A document traverses from one space to another through these “wormhole” connections.

• See also: Wormhole Vectors Beyond Hybrid Search in OpenSearch — Dima Kan’s production implementation at Aiven

Related Concepts

• Dense Embeddings — the dense leg of hybrid search

• Sparse Embeddings — the sparse leg of hybrid search

• Sparse Vector Retrieval — one leg of hybrid

• Dense Vector Retrieval — the other leg

• SPLADE — preferred learned sparse component

• ELSER — Elastic’s sparse model for hybrid

• Bi-Encoder — preferred dense component

• ColBERT — alternative dense component with late interaction

• Wormhole Vectors — advanced multi-space traversal

• RAG — downstream use of hybrid retrieval

People

• Trey Grainger — Wormhole Vectors, multi-space traversal

• James Briggs — Hybrid SPLADE articles at Pinecone

• Jo Kristian Bergum — Vespa hybrid search, ColBERT embedder

• Stéphane Clinchant — SPLADE, sparse component

• Bayesian BM25 is Cool

• Bayesian BM25

Semantic Boosting

Semantic Boosting is a two-phase alternative to RRF/RSF fusion: run vector search first, inject the results as boost clauses into a final lexical query. The output comes entirely from the lexical engine, so faceting, highlighting, and pagination work natively without extra merging.

• See: Hybrid Search Blueprint Series Semantic Boosting — Erik Hatcher (MongoDB)
• See: Relative Score Fusion — score-normalization-based fusion, contrasted with RRF