Hybrid Search

Combine vector similarity with full-text search for best-in-class retrieval. Get the semantic understanding of embeddings plus the precision of keyword matching.

What is Hybrid Search?

Hybrid search combines vector similarity (semantic meaning) with full-text search (keyword matching) to provide superior search results that understand both context and exact terms.

Vector Search Alone

✓ Understands semantic meaning
✓ Finds conceptually similar content
✗ May miss exact keyword matches
✗ Can return loosely related results

Full-Text Search Alone

✓ Precise keyword matching
✓ Fast for exact terms
✗ No semantic understanding
✗ Misses synonyms and context

Hybrid Search = Best of Both

Semantic understanding from vector embeddings
Precise keyword matching from full-text search
Weighted scoring to balance both approaches
Superior relevance and recall

Basic Hybrid Search

Setup Table

-- Create table with vector and text columns
CREATE TABLE knowledge_base (
  id SERIAL PRIMARY KEY,
  title TEXT,
  content TEXT,
  embedding vector(384),
  ts_vector tsvector
);

-- Create indexes
CREATE INDEX ON knowledge_base USING hnsw (embedding vector_l2_ops);
CREATE INDEX ON knowledge_base USING gin (ts_vector);

-- Insert data with embeddings and text index
INSERT INTO knowledge_base (title, content, embedding, ts_vector)
VALUES (
  'PostgreSQL Performance',
  'Optimize queries with indexes and EXPLAIN',
  embed_text('Optimize queries with indexes and EXPLAIN'),
  to_tsvector('Optimize queries with indexes and EXPLAIN')
);

Perform Hybrid Search

-- Hybrid search: 70% vector + 30% text
SELECT * FROM hybrid_search(
  'knowledge_base',      -- table name
  'content',             -- text column
  'embedding',           -- vector column
  'database optimization', -- query
  10,                    -- limit
  0.7,                   -- vector weight (70%)
  0.3                    -- text weight (30%)
);

-- Returns: id, title, content, vector_score, text_score, hybrid_score

Reranking for Precision

Cross-Encoder Reranking

Cross-encoders jointly encode the query and each candidate for higher precision scoring.

-- Rerank search results with cross-encoder
SELECT rerank_cross_encoder(
  'What is PostgreSQL?',                    -- query
  ARRAY[                                    -- candidate documents
    'PostgreSQL is a database',
    'MySQL is also a database',
    'Redis is a cache'
  ],
  'ms-marco-MiniLM-L-6-v2',                -- model
  3                                         -- top_n
);

-- Returns:
--  idx | score
-- -----+-------
--   0  | 0.945  (most relevant)
--   1  | 0.678
--   2  | 0.123  (least relevant)

MMR (Maximal Marginal Relevance)

MMR balances relevance with diversity to avoid redundant results.

-- MMR reranking for diverse results
SELECT * FROM mmr_rerank(
  'knowledge_base',           -- table
  'embedding',                -- vector column
  embed_text('database'),     -- query vector
  20,                         -- fetch top 20 candidates
  5,                          -- return top 5 diverse results
  0.7                         -- lambda (0.7 = 70% relevance, 30% diversity)
);

-- Returns: id, title, score, diversity

Best Practices

1. Tune Vector/Text Weights

Adjust the balance based on your use case:

0.9/0.1 - Heavily semantic (concepts matter more than exact terms)
0.7/0.3 - Balanced (default, works for most cases)
0.5/0.5 - Equal weight (both semantic and keywords important)
0.3/0.7 - Keyword-focused (exact terms critical)

2. Index Both Columns

-- Always index both for performance
CREATE INDEX ON docs USING hnsw (embedding vector_l2_ops);
CREATE INDEX ON docs USING gin (ts_vector);

3. Use Reranking for Top Results

Fetch 50-100 candidates with hybrid search, then rerank top 10-20 for best precision.