Hybrid search is a retrieval method that combines keyword ranking and semantic ranking into one result list, so exact terms and meaning-based matches both surface. In commercial real estate it catches a defined term like "Base Rent" that keyword search alone finds and a paraphrased clause that semantic search alone finds, then fuses the two rankings into a single ordered list.
How Does Hybrid Search Work?
Hybrid search works by running two retrievers in parallel, a keyword retriever and a semantic retriever, then merging their ranked lists into one. The keyword side, often BM25, scores exact and near-exact term matches. The semantic side scores by vector-embedding similarity. A fusion step combines both rankings so a passage strong on either signal rises to the top.
The most common merge is Reciprocal Rank Fusion, introduced by Cormack, Clarke, and Buttcher in their 2009 SIGIR paper "Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods." RRF scores each passage as the sum over both lists of 1 divided by (k plus its rank), with k commonly set to 60. It fuses on rank position, not raw scores, so incompatible keyword and similarity scales combine cleanly.
Component | Function |
Keyword retriever | BM25 ranks passages by exact term overlap |
Semantic retriever | Vector similarity ranks passages by meaning |
Fusion (RRF) | Merges both ranked lists into one order |
Constant k | Dampens the weight of top ranks; commonly 60 |
Final list | One ordered result blending both signals |
Why Hybrid Search Matters
Hybrid search matters because CRE documents mix exact defined terms with varied plain-language wording, and each retriever alone misses half the problem. Keyword search finds "Base Rent" but misses a clause that says "the fixed annual charge." Semantic search finds the paraphrase but can rank an exact suite number or defined term below a loose paraphrase. Fusing both recovers what either drops.
The lift is measurable. On the public WANDS retrieval benchmark, BM25 and pure vector search score near-identical NDCG around 0.698, while a tuned hybrid reaches 0.7497, a 7.4 percent gain over either alone, per DigitalApplied's 2026 reference. The working principle is direct: hybrid search keeps the exactness of keywords and the recall of meaning, so a controlling clause is rarely lost to either a vocabulary gap or a missing exact term.
Example
Hybrid search is clearest when one query needs both signals. An analyst searches a 90-page retail lease for the defined term "Base Rent," which appears once on page 41, while a plain-language clause on page 12 says "the fixed annual charge payable by Tenant." Keyword search ranks page 41 first; semantic search ranks page 12 first.
Passage | Keyword rank | Semantic rank | RRF score (k=60) |
Page 41 "Base Rent" | 1 | 4 | 1/61 + 1/64 = 0.0320 |
Page 12 "fixed annual charge" | 6 | 1 | 1/66 + 1/61 = 0.0316 |
Using RRF with k equal to 60, page 41 scores 1/61 plus 1/64, which is 0.01639 plus 0.01563, or 0.0320. Page 12 scores 1/66 plus 1/61, which is 0.01515 plus 0.01639, or 0.0316. Both passages land at the top of one fused list, so the analyst sees the defined term and the paraphrase together. Keyword search alone would have buried page 12; semantic search alone would have buried page 41.
Variations and Edge Cases
Hybrid search is a family of designs, and the merge strategy and weighting change results. Some systems weight one retriever heavier, others add a re-ranker after fusion. The variants below trade tuning control against simplicity.
Variant | Treatment |
Reciprocal Rank Fusion | Merges on rank; no score normalization needed |
Weighted score fusion | Blends normalized scores; needs scale calibration |
Re-ranked hybrid | A cross-encoder re-scores the fused top results |
Filtered hybrid | Metadata narrows the corpus before both retrievers run |
Sparse-dense | Pairs a sparse keyword vector with a dense semantic vector |
Hybrid Search vs Semantic Search
Hybrid search is often confused with semantic search, and the difference is how many signals rank the results. Semantic search ranks by one signal, vector-embedding similarity, so it retrieves by meaning alone. Hybrid search runs a keyword retriever and a semantic retriever together, then fuses both rankings, so exact terms and meaning both influence the order.
Semantic search can miss an exact defined term or a suite number that a paraphrase outranks. Hybrid search adds the keyword signal back, so "Base Rent" and "the fixed annual charge" both surface. In production CRE retrieval, hybrid search is the common default precisely because leases pair rigid defined terms with loose surrounding prose.
Frequently Asked Questions
What is hybrid search?Hybrid search is a retrieval method that combines keyword ranking and semantic ranking into one result list. It runs both retrievers in parallel and fuses their rankings, so exact defined terms and paraphrased clauses both surface, which single-signal search would otherwise miss.
How is hybrid search different from semantic search?Semantic search ranks results by vector similarity alone, retrieving by meaning. Hybrid search adds a keyword retriever and fuses both rankings, so exact terms and meaning both count. Hybrid catches a defined term like "Base Rent" that pure semantic search can rank below a loose paraphrase.
What is reciprocal rank fusion in hybrid search?Reciprocal rank fusion is a method that merges two ranked lists by scoring each passage as the sum of 1 divided by (k plus its rank) across both lists, with k commonly 60. It fuses on rank position rather than raw scores, so incompatible keyword and similarity scales combine cleanly.
Related Terms
Semantic Search
Vector Database
Vector Embedding
Retrieval-Augmented Generation
Document Extraction