11. Hybrid Search

Hybrid search combines keyword search (exact matches) with semantic search (meaning matches) to deliver better results than either method alone. It’s the standard approach used by every production RAG system.

Semantic search understands meaning but can miss exact terms. Keyword search finds exact matches but misses context. Hybrid search takes both results, merges them, and returns the best overall matches. It’s not an either/or — it’s both.

You search for “React useState.” What do you want?

Semantic search finds: “Managing component state in React,” “useReducer vs useState,” “React hooks tutorial” — all about the topic but none containing the exact phrase.

Keyword search finds: “React useState documentation,” “React useState not updating,” “useState returns undefined” — all containing the exact phrase but potentially missing related content.

Hybrid search finds both sets and returns the best combined results. You get the exact documentation AND related tutorials.

Imagine a library with two specialist librarians:

Librarian A (Keyword): Matches exact words only. You say “ancient Rome” — she finds books with “ancient” and “Rome” in the title. She’s fast and precise but misses “The Roman Empire” and “Caesar’s legacy.”

Librarian B (Semantic): Understands meaning. You say “ancient Rome” — she finds books about Roman history, culture, architecture. She’s thorough but might bring you a book about Roman mythology when you wanted military history.

Together: They compare notes, merge their findings, and bring you the most relevant books from both approaches. That’s hybrid search.

flowchart TD
    Q["User Query:\n'tell me about React hooks'"] --> KW["🔍 Keyword Search\n(BM25 / Elasticsearch)"]
    Q --> VS["🔢 Semantic Search\n(Vector / Embedding)"]

    KW --> KW_RESULTS["Exact matches:\n'React Hooks documentation'\n'useEffect Hook guide'\n'Hooks API reference'"]
    VS --> VS_RESULTS["Meaning matches:\n'Managing side effects'\n'State in functional components'\n'React lifecycle patterns'"]

    KW_RESULTS --> MERGE["🔄 Merge & Normalize\n(combine + deduplicate + rescore)"]
    VS_RESULTS --> MERGE

    MERGE --> FINAL["Top 5 Results:\n1. 'React Hooks documentation' (KW)\n2. 'Managing side effects' (Semantic)\n3. 'useEffect Hook guide' (KW)\n4. 'State in functional components' (Semantic)\n5. 'Hooks API reference' (KW)"]

    style Q fill:#3b82f6,color:#fff
    style KW fill:#f59e0b,color:#fff
    style VS fill:#8b5cf6,color:#fff
    style MERGE fill:#22c55e,color:#fff

flowchart LR
    subgraph QUERY["Query Processing"]
        Q["User Question"] --> Q_EMBED["🔢 Embedding\nModel"]
        Q --> Q_KEYWORD["🔍 BM25\nTokenization"]
    end

    subgraph SEARCH["Dual Search"]
        Q_EMBED --> VEC["Vector Search\n(ANN index)"]
        Q_KEYWORD --> KW["Keyword Search\n(inverted index)"]
    end

    subgraph MERGE_RESULTS["Merge & Rerank"]
        VEC --> NORM["Normalize Scores\n(0-1 range)"]
        KW --> NORM
        NORM --> COMBINE["Weighted Combination\n0.5 × vector + 0.5 × keyword"]
        COMBINE --> TOP["Top-K Results"]
    end

    style QUERY fill:#3b82f6,color:#fff
    style SEARCH fill:#8b5cf6,color:#fff
    style MERGE_RESULTS fill:#22c55e,color:#fff

1. User submits a query — “How do I reset my password in the admin panel?”
2. Vector search — Embed the query, search ANN index for semantically similar chunks
3. Keyword search — Tokenize the query, search inverted index (BM25) for exact/partial matches
4. Normalize scores — Both methods return scores in different ranges. Normalize both to 0-1
5. Weighted combination — final_score = alpha × vector_score + (1 - alpha) × keyword_score (alpha is typically 0.5-0.7)
6. Return top-K — The highest combined scores become the final results

flowchart TD
    subgraph COMPARISON["Search Method Comparison"]
        SCENARIO["Query: 'Fix login bug in auth module'"]
        KW_ONLY["🔍 Keyword Only\n'login bug auth module'\n❌ Misses:\n'authentication error'\n'user cannot sign in'"]
        SEM_ONLY["🔢 Semantic Only\n'fix login issue'\n❌ Misses:\n'exact auth module docs'\n'bug report #142'"]
        HYBRID["🔄 Hybrid\n✅ Both exact docs\n✅ Related topics\n✅ Best coverage"]
    end

    SCENARIO --> KW_ONLY
    SCENARIO --> SEM_ONLY
    SCENARIO --> HYBRID

    style KW_ONLY fill:#f59e0b,color:#fff
    style SEM_ONLY fill:#8b5cf6,color:#fff
    style HYBRID fill:#22c55e,color:#fff

Q: What is hybrid search and why is it better than pure semantic search?

Hybrid search combines keyword search (exact word matching with BM25) and semantic search (vector similarity). It’s better because it catches both exact matches (product codes, names) AND meaning-based matches (synonyms, paraphrases) — providing better coverage than either alone.

Q: How do you combine results from keyword and semantic search?

Two common methods: (1) Score normalization — normalize both scores to 0-1, then combine: final = alpha × vector + (1-alpha) × keyword. (2) Reciprocal Rank Fusion (RRF) — assign each result a score based on its rank in each list: score = 1 / (k + rank). RRF is simpler and doesn’t need score normalization.

Q: Design a hybrid search system for a code documentation platform with 500K+ documents.

Architecture: (1) Indexing — index code docs with both BM25 (code-specific analyzer) and vector embeddings (code-aware model like Voyage-code). (2) Query routing — detect if query contains code terms (function names, keywords) → boost keyword weight. (3) Hybrid strategy — use RRF with k=60 for merging, alpha=0.6 vector bias for natural language queries. (4) Performance — run keyword and vector searches in parallel, cache frequent queries. (5) Evaluation — track NDCG@10 on a curated test set of code search queries.

Previous: 10 — Building a Complete RAG Pipeline →

Next: 12 — Re-ranking →