Redis Vector Search
Redis Vector Search extends Redis with vector similarity capabilities through the RediSearch module. Store embeddings alongside your existing Redis data, query with sub-millisecond latency, and leverage Redis's 1M+ ops/second throughput. Perfect for applications that need both traditional caching and vector search: recommendation engines, semantic search, real-time personalization. Key advantages: in-memory speed, horizontal scaling, familiar Redis API, and seamless integration with existing Redis infrastructure. Supports HNSW and FLAT indexing with cosine, L2, and IP distance metrics.
Overview
Redis Vector Search brings vector similarity to Redis—the world's most popular in-memory data store. Instead of deploying separate vector database, use Redis for both traditional caching (user sessions, API responses) and vector operations (semantic search, recommendations). Sub-millisecond query latency, 1M+ ops/second, and horizontal scaling via Redis Cluster. Use cases: real-time product recommendations (store user embeddings + product catalog), semantic search (cache query embeddings), session-based personalization (user context + vector search in same transaction).
Key Features
- **In-Memory Speed**: Sub-millisecond vector queries, 10-100× faster than disk-based vector DBs
- **Hybrid Queries**: Combine vector search with traditional filters (price range + semantic similarity)
- **HNSW Index**: Hierarchical Navigable Small World algorithm for fast approximate search
- **Redis Integration**: Use alongside existing Redis features (caching, pub/sub, streams)
- **Horizontal Scaling**: Redis Cluster distributes vectors across nodes
- **Multiple Metrics**: Cosine similarity, L2 distance, inner product
Business Integration
Redis Vector Search eliminates infrastructure complexity for real-time AI applications. E-commerce sites already using Redis for caching add semantic product search without new database. Gaming platforms using Redis for leaderboards add player matchmaking by playstyle similarity. Financial services using Redis for session management add fraud detection via transaction embedding similarity. The key advantage: consolidate infrastructure—one database for caching, key-value, and vectors reduces operational overhead and improves performance through data locality.
Implementation Example
Technical Specifications
- **Query Latency**: <1ms for HNSW queries (in-memory)
- **Throughput**: 100K-1M vector operations/second per node
- **Scale**: Billions of vectors across Redis Cluster
- **Dimensions**: Up to 32,768 dimensions supported
- **Index Types**: HNSW (fast approximate), FLAT (exact but slower)
- **Distance Metrics**: Cosine, L2, Inner Product
Best Practices
- Use HNSW for >10K vectors, FLAT for <10K where exact search needed
- Set appropriate M and EF_CONSTRUCTION for HNSW based on accuracy/speed tradeoff
- Use Redis Cluster for >10M vectors to distribute across nodes
- Combine vector search with filters for better relevance
- Monitor memory usage—vectors stored in RAM
- Use Redis persistence (RDB/AOF) to avoid losing vector data on restart