HaystackDB

Minimal but performant Vector DB

Features

• Binary embeddings by default (soon int8 reranking)
• JSON filtering for queries
• Scalable, distributed architecture for use with multi replica deployments
• Durable (WAL), persistent data, mem mapped for fast access in the client

Benchmarks

On a MacBook with an M2, 1024 dimension, binary quantized.

FAISS is using a flat index, so brute force, but it's in memory. Haystack is storing the data on disk, and also brute forces.

TLDR is Haystack is ~10x faster despite being stored on disk.

100,000 Vectors
Haystack — 3.44ms
FAISS    — 29.67ms

500,000 Vectors
Haystack — 11.98ms
FAISS    - 146.50ms

1,000,000 Vectors
Haystack — 22.65ms
FAISS    — 293.91ms

Roadmap

• Quickstart Guide
• Quality benchmarks (this is in progress)
• Int8 reranking
• Better queries with more than simple equality (this is done now)
• Full text search
• Better insertion performance with batch B+Tree insertion (could probably be further improved, but good for now)
• Point in time backups/rollback
  • currently this is destructive (ie you cannot return forward after you go backwards), so a nondestructive version is next on the todo list.
• Cursor based pagination
• Schema migrations
• Vector Kmeans clustering with centroid similarity for improved search perf