ClickHouse Vector Search

Overview

As of version 25.6.2.5, ClickHouse supports vector similarity indexing using the HNSW algorithm, enabling high-performance approximate nearest-neighbor (ANN) searches within the analytical database.

Supported ANN Index Types

HNSW (Hierarchical Navigable Small World)

• Current state-of-the-art method for approximate vector search
• Based on hierarchical navigable small world graphs algorithm
• Implemented via the USearch library
• Standard for fast vector retrieval

Annoy (Previously Supported)

• Based on the ANNOY library in C++ from Spotify
• Used in Spotify's music recommendations
• Currently experimental in ClickHouse

Search Approaches

ClickHouse offers two main approaches for vector search:

1. Brute-force search using distance functions

   • Highly accurate but slower
   • No index required
   • Best for smaller datasets

2. Indexing vector fields

   • Faster but less precise
   • Uses ANN indexes (HNSW)
   • Suitable for large-scale deployments

QBit Data Type

ClickHouse introduced a new data type QBit where no early decisions are needed and you can adjust precision and speed trade-off directly at query time. This allows for flexible optimization based on specific query requirements.

Memory & Performance Considerations

• The vector similarity index must be fully loaded from disk into main memory to perform searches
• Vector index is constructed fully in memory and then saved to disk
• Provides fast retrieval for billion-scale vector datasets

Use Cases

• Real-time analytics with vector search
• Recommendation systems
• Semantic search in data warehouses
• Combining analytical queries with similarity search
• Large-scale embedding storage and retrieval

Architecture

ClickHouse's vector search is built into its MergeTree table engine family, allowing seamless integration with existing analytical workloads and SQL queries.

Pricing

Free and open-source (Apache 2.0 license). Commercial cloud offerings available through ClickHouse Cloud.