Lesson 6.3: Hybrid Queries (Semantic + Structured)

The Power of Hybrid Search

Semantic search alone (imprecise):

-- Find similar documents
SELECT content FROM documents
ORDER BY embedding <=> query_embedding
LIMIT 10;

-- Problem: Returns results user doesn't have access to
-- Or from wrong departments, or outdated content

SQL filters alone (brittle):

-- Keyword search
SELECT content FROM documents
WHERE content ILIKE '%machine learning%'
    AND department = 'ai';

-- Problem: Misses synonyms, related concepts
-- "neural networks", "deep learning", "ML models" won't match

Hybrid (best of both):

SELECT
    content,
    title,
    department,
    embedding <=> query_embedding AS semantic_score,
    ts_rank(search_vector, query) AS keyword_score,
    -- Combined score
    (embedding <=> query_embedding) * 0.7 +
    (1 - ts_rank(search_vector, query)) * 0.3 AS combined_score
FROM documents
WHERE
    -- Structured filters (fast, precise)
    user_id = 123
    AND department IN ('ai', 'engineering')
    AND created_at > NOW() - INTERVAL '1 year'
    -- Semantic filter (flexible, recalls related content)
    AND embedding <=> query_embedding < 0.8
    -- Keyword filter (catches exact matches)
    AND search_vector @@ to_tsquery('machine | learning')
ORDER BY combined_score
LIMIT 20;

Pre-Filtering Strategy

Performance optimization:

-- Step 1: Filter with SQL (fast, uses B-tree indexes)
WITH filtered_docs AS (
    SELECT id, content, title
    FROM documents
    WHERE user_id = 123
        AND department = 'engineering'
        AND created_at > NOW() - INTERVAL '6 months'
    -- Reduces from 1M to 10k rows
)
-- Step 2: Vector search on subset (much faster)
SELECT
    d.content,
    e.embedding <=> query_embedding AS similarity
FROM filtered_docs d
JOIN embeddings e ON d.id = e.document_id
WHERE e.embedding <=> query_embedding < 0.7
ORDER BY similarity
LIMIT 10;

Key Takeaways

• Semantic search alone is imprecise - returns irrelevant or unauthorized results
• SQL filters alone are brittle - miss synonyms and related concepts
• Hybrid search combines semantic similarity with structured filters for best results
• Pre-filtering dramatically improves performance by reducing vector search scope
• Combined scoring allows balancing semantic and keyword relevance
• Full-text search (ts_rank) complements vector search for exact keyword matches
• Production systems should always combine semantic search with SQL filters