Hannoy

Overview

Hannoy is a graph-based successor to Arroy (Meilisearch's tree-based ANN library), designed to combine the best aspects of graph-based approximate nearest neighbor search with KV-backed storage via LMDB. It represents the next evolution in Meilisearch's vector search technology.

Design Goals

Hannoy aims to combine:

• Graph-based ANN: Superior search performance of graph algorithms
• LMDB Storage: Production-ready persistent key-value storage
• Production Features: The reliability and update capabilities of Arroy

Evolution from Arroy

Arroy (Previous Generation)

• Tree-based LSH (Locality Sensitive Hashing) approach
• Built on LMDB for persistent storage
• Incremental updates without full rebuilds
• Memory-mapped for multi-process sharing

Hannoy (Current Generation)

• Graph-based algorithms (similar to HNSW)
• Maintains LMDB backend
• Enhanced search performance (approximately 10x faster)
• Retains production-ready features from Arroy

Key Features

• Graph-Based Search: Uses graph algorithms for more efficient nearest neighbor search
• LMDB Backend: Persistent, memory-mapped storage for reliability and multi-process access
• Incremental Updates: Support for updating indexes without expensive rebuilds
• Production-Ready: Designed for use in production environments like Meilisearch
• Memory Efficiency: Optimized for applications with millions of high-dimensional vectors

Performance Improvements

Blog posts suggest that moving from tree-based (Arroy) to graph-based (Hannoy) algorithms can provide up to 10x speedup in vector search performance while maintaining the same storage and update characteristics.

Use in Meilisearch

Hannoy is being developed as part of Meilisearch's vector search capabilities to:

• Handle millions of documents with high-dimensional embeddings (768-1536 dimensions)
• Provide fast incremental updates as content changes
• Maintain low memory footprint
• Support multi-modal search (text and images)

Technical Architecture

• Storage: LMDB (Lightning Memory-Mapped Database)
• Algorithm: Graph-based ANN (similar to HNSW)
• Language: Rust
• Integration: Built into Meilisearch's vector search stack

Distance Metrics

Likely supports similar metrics to Arroy:

• Euclidean distance
• Manhattan distance
• Cosine distance
• Dot (Inner) Product distance

Resources

• Blog: https://blog.kerollmops.com/from-trees-to-graphs-speeding-up-vector-search-10x-with-hannoy
• Related: Arroy at https://github.com/meilisearch/arroy
• Developer: Developed by the Meilisearch team

Significance

Hannoy represents the evolution of LMDB-backed vector search from tree-based to graph-based algorithms, maintaining production features while achieving significant performance improvements. It demonstrates how modern vector databases can combine cutting-edge algorithms with robust, persistent storage systems.