Hybrid Search: Combining Keyword and Vector Search

Why it exists

Keyword search is precise but misses conceptual context. Vector search understands concepts but can miss critical, exact terms. Hybrid search was created to solve this trade-off, combining the strengths of both to deliver more relevant results than either could alone.

The mental model

Think of hybrid search like two expert researchers working together. One is a meticulous librarian who finds every document with your exact keywords (full-text search). The other is a subject matter expert who understands the underlying topic and finds related material, even if the words don't match (vector search). Hybrid search acts as the manager who takes both sets of findings and intelligently synthesizes them into a single, superior ranked list.

How it works

A hybrid search system receives a single query but executes two searches in parallel against an index that contains both plain text and vector embeddings. First, a full-text search (using an algorithm like BM25) finds documents with exact keyword matches. Second, a vector similarity search (using an algorithm like HNSW) finds documents that are conceptually similar. The system then uses a fusion algorithm, typically Reciprocal Rank Fusion (RRF), to merge the two separate result lists. RRF creates a new, unified ranking based on the position of each item in both original lists, producing a single, more relevant result set.

When to use it

Use hybrid search when relevance is critical, especially for feeding context to LLMs in Retrieval-Augmented Generation (RAG) systems. It excels in scenarios where queries contain a mix of specific terms and general concepts, such as product searches with SKUs, queries with specialized jargon, or searches involving names and dates. It provides a robust retrieval safety net, catching both literal and semantic matches.

When not to use it

If your search problem is one-dimensional, hybrid search may be overkill. For simple log analysis where you only need exact keyword matches, standard full-text search is more efficient. Similarly, if you only care about conceptual similarity (like finding similar images) and specific terms are irrelevant, pure vector search is sufficient. Hybrid search adds complexity and computational cost that isn't always necessary.

One canonical example

A user searches an internal knowledge base for "troubleshooting error code 0x80070005 in the new billing system." A pure keyword search might miss a document titled "Fixing Access Denied Problems in the Finance Platform" which solves the issue but uses different phrasing. A pure vector search might find general articles about access issues but miss the critical document that mentions the exact error code. Hybrid search finds both, and the RRF algorithm likely promotes the document with both the exact error code and relevant concepts to the top of the results.