Course contentsShow
← AI Engineering
Lesson 279 of 1,8867. Vector Databases - FundamentalsFree lesson

Hybrid Search: Keyword + Vector

Combining traditional keyword search (BM25) with vector search for better retrieval accuracy.

Hybrid Search: Keyword + Vector

What You'll Learn

You'll learn how combining traditional keyword search with vector similarity search creates a more powerful retrieval system than either approach alone.

The Core Idea

Hybrid search merges two complementary search methods:

  1. Keyword search (BM25) - Finds documents containing specific terms, great for exact matches, names, and rare words
  2. Vector search - Finds semantically similar documents, great for concepts and paraphrases

Think of it like asking for restaurant recommendations: keyword search finds places with "vegan pizza" in the menu, while vector search finds Italian restaurants with plant-based options. Together, they catch more relevant results than either alone.

Why Combine Them?

Each approach has strengths the other lacks:

  • Vector search understands "cheap flights" and "affordable airfare" are similar, but might miss exact product codes or specific names
  • Keyword search finds "iPhone 15 Pro" precisely but misses "latest Apple flagship phone"

Hybrid search runs both queries and combines their results, typically by:

  • Computing separate scores for each method
  • Normalizing scores to the same scale
  • Weighted combination (e.g., 70% vector + 30% keyword)
  • Reciprocal rank fusion for result merging

When to Use Hybrid Search

Use hybrid search when:

  • Users mix specific terms with conceptual queries
  • Exact matches matter (product IDs, names) but semantics also help
  • You want the safety net of both approaches

The metadata filtering you learned earlier (lesson 278) works with both search types, giving you three-way power: keywords + vectors + filters.

Key Takeaway: Hybrid search combines BM25 keyword matching with vector similarity to leverage both exact-term precision and semantic understanding, producing more robust retrieval than either method alone.