Vector Search for RAG

← All RAG Cookbooks

Vector Search Deep Dive

HNSW vs FLAT indexes, hybrid search, metadata filtering, and performance optimization.

HNSW vs FLAT Indexes

HNSW (Approximate)

• ~0.5ms search latency
• 95%+ recall accuracy
• Best for large datasets (>10K vectors)
• Higher memory overhead

FLAT (Exact)

• Linear scan - O(n) complexity
• 100% accuracy guaranteed
• Best for small datasets (<10K vectors)
• Lower memory overhead

# HNSW Index (recommended for production)
FT.CREATE idx:hnsw ON HASH PREFIX 1 "doc:"
  SCHEMA
    embedding VECTOR HNSW 6
      TYPE FLOAT32
      DIM 1536
      DISTANCE_METRIC COSINE
      M 16                  # Max connections per node
      EF_CONSTRUCTION 200   # Build-time accuracy

# FLAT Index (for small datasets or exact search)
FT.CREATE idx:flat ON HASH PREFIX 1 "doc:"
  SCHEMA
    embedding VECTOR FLAT 6
      TYPE FLOAT32
      DIM 1536
      DISTANCE_METRIC COSINE

HNSW Parameters

Hybrid Search: Vectors + Keywords

Combine vector similarity with keyword matching for more precise results:

# Create index with text and vector fields
FT.CREATE idx:hybrid ON HASH PREFIX 1 "doc:"
  SCHEMA
    content TAG
    category TAG
    embedding VECTOR HNSW 6
      TYPE FLOAT32
      DIM 1536
      DISTANCE_METRIC COSINE

# Hybrid query: vector search + category filter
FT.SEARCH idx:hybrid
  "(@category:{technical})=>[KNN 5 @embedding $vec AS score]"
  PARAMS 2 vec [embedding_bytes]
  DIALECT 2

# Hybrid query: vector search + keyword filter
FT.SEARCH idx:hybrid
  "(@content:{python programming})=>[KNN 5 @embedding $vec AS score]"
  PARAMS 2 vec [embedding_bytes]
  DIALECT 2

Metadata Filtering

# Index with multiple filter fields
FT.CREATE idx:docs ON HASH PREFIX 1 "doc:"
  SCHEMA
    content TAG
    category TAG
    author TAG
    date NUMERIC
    embedding VECTOR HNSW 6
      TYPE FLOAT32
      DIM 1536
      DISTANCE_METRIC COSINE

# Filter by category
FT.SEARCH idx:docs
  "(@category:{billing})=>[KNN 5 @embedding $vec AS score]"
  ...

# Filter by date range
FT.SEARCH idx:docs
  "(@date:[1704067200 1706745600])=>[KNN 5 @embedding $vec AS score]"
  ...

# Multiple filters
FT.SEARCH idx:docs
  "(@category:{support} @author:{john})=>[KNN 5 @embedding $vec AS score]"
  ...

Python Implementation

from redis.commands.search.query import Query

def hybrid_search(query_text, category=None, date_range=None, k=5):
    # Build filter
    filters = []
    if category:
        filters.append(f"@category:{{{category}}}")
    if date_range:
        filters.append(f"@date:[{date_range[0]} {date_range[1]}]")

    filter_str = " ".join(filters) if filters else "*"

    # Build query
    query_embedding = get_embedding(query_text)
    query_bytes = np.array(query_embedding, dtype=np.float32).tobytes()

    q = Query(f"({filter_str})=>[KNN {k} @embedding $vec AS score]")
    q.return_fields("content", "category", "score")
    q.dialect(2)

    return client.ft("idx:docs").search(q, {"vec": query_bytes})

← Semantic Caching Next: Cache Invalidation →