SwarnDB
Vector database that thinks in graphs
Start new thread
Chirotpal

SwarnDB - Vector database that thinks in graphs

by•
SwarnDB is a Rust vector database that adds a virtual graph layer and 15+ built-in vector math operations on top of HNSW search. Query vectors, traverse the relationships SwarnDB computes for you, and run analytics like k-means or PCA, all in one engine.

Add a comment

Replies

Best
Chirotpal
Maker
📌
Hey PH folks, this is Chirotpal, and I created SwarnDB. SwarnDB is a vector database written in Rust that does not stop at nearest-neighbor search. It computes the relationships between your vectors automatically and exposes them as a graph you can traverse. So you get vector search, graph traversal, and 15+ vector math operations (k-means, PCA, SLERP interpolation, cone search, drift detection, maximal marginal relevance, ghost-vector detection, and more) in one engine, instead of stitching three databases together. What's in this release (v1.0.3): - 2,398 queries per second at 98% recall on DBpedia 1M (1536 dim) with default HNSW parameters on a 32-core box, 8 concurrent searcher threads. Sub-7ms p99. - File-based bulk ingestion: stage a `.npy` or flat `.f32` file on disk, point the server at it, and the server memory-maps it directly. Working memory stays bounded by the index being built, not by the input file size. Loading a million vectors no longer balloons your container. - Plain HNSW collections become queryable within seconds of the server opening its ports. - Multi-collection databases load collections in parallel at startup. - Transparent crash recovery: incremental delta replay if a snapshot exists, full write-ahead log replay otherwise, both happen automatically before traffic resumes. - Operational endpoints for Kubernetes orchestrators: /healthz, /readyz, /startupz, plus a global /recovery_status and per-collection /persistence_status. - Async and sync Python clients with identical method names and return types. - Bulk inserts produce checkpoints and a resume token so interrupted loads pick up from the last committed batch. - Multi-arch Docker image (linux/amd64 + linux/arm64) on Docker Hub. - Pure-Python wheel on PyPI for Linux x86_64, Linux ARM64, macOS Apple Silicon, and Windows x86_64. Try it in 30 seconds: - `pip install swarndb` - `docker run -d -p 8080:8080 -p 50051:50051 sarthiai/swarndb` Source: github.com/SarthiAI/SwarnDB Benchmarks: github.com/SarthiAI/SwarnDB/blob/main/docs/benchmarks.md License: Elastic License 2.0 (free for production use). I will be in the thread all day. Would love to hear: what are you building with vector search today, and where does the lack of a relationship/graph layer hurt? Also happy to dive into the benchmark methodology, the persistence model, or the SDK design if anyone wants the deep end.