GalaxDB

AI-native database: SQL + vector search + training exports

Open Source

Developer Tools

Artificial Intelligence

GitHub

Visit WebsiteSee on Product HuntGithub ⧉Twitter ⧉

Hey Product Hunt! I'm Ibrahim, the founder behind GalaxDB. The problem Most AI teams run 5+ separate systems: PostgreSQL for transactions, Pinecone or Qdrant for vector search, Redis for caching, S3 for training data, and a feature store for serving. Each adds latency, cost, and operational complexity. Data gets duplicated across systems, embeddings go stale, and no single system tracks lineage end-to-end. What GalaxDB does One 7.9 MB binary that speaks PostgreSQL wire protocol. It replaces the entire stack: Full OLTP SQL with ACID Snapshot Isolation HNSW vector search with auto‑embedding, no external API, no separate vector DB Merkle‑DAG versioning with time‑travel queries and reproducible training snapshots Training‑data export in Lance format with zero‑copy PyTorch loading and built‑in dedup via MinHash LSH EU AI Act Article 13 compliant lineage, every training run is traceable to exact tag, filters, precision, and export hash Benchmarks (measured on AWS c6id.4xlarge, Intel Xeon 8375C, 32 GB RAM) 258,555 writes/sec with write p99 of 377 µs under sustained compaction 0.983 recall@10 on SIFT1M matching hnswlib, in safe Rust 4.49 GB/s column scans with 80% zone‑map block skip 7,390 wire‑protocol SELECT QPS - 2.3× PostgreSQL 16 Cold‑cache point read p50: 147 µs on a 50M‑row dataset exceeding RAM 7/7 chaos scenarios passed with zero data loss Try it in 60 seconds pip install galaxdb-client import galaxdb db = galaxdb.Database("./mydata") db.execute(""" CREATE TABLE docs ( id INT PRIMARY KEY, text TEXT EMBEDDING MODEL 'all-MiniLM-L6-v2' ) """) db.execute("INSERT INTO docs VALUES (1, 'machine learning is great')") results = db.execute( "SELECT * FROM docs WHERE SEMANTIC_MATCH(text, 'AI', 0.7)" ) Honest limitations (v1 is not everything yet) Single‑row wire INSERT is slow - 454 rows/sec. Batching restores performance: 20,267 rows/sec No secondary indexes yet - non‑PK lookups fall back to zone‑map pruning Vector index is single‑node - distributed ANN is planned for v2 macOS and Windows are dev platforms - production performance needs Linux 5.10+ with io_uring Links GitHub: https://github.com/zentrix-innov... Research paper with all benchmarks: https://doi.org/10.5281/zenodo.2... Docs: https://galaxdb.com/docs I would love your feedback — especially on the architecture, the benchmarks, and what you would want in v2. I will be here answering everything.