GalaxDB

GalaxDB ships as a single binary that handles transactional rows, vector indexing, local embedding inference, blob storage, and versioned snapshots in one query engine. The surface API is SQL: your existing psycopg2 code connects unchanged, and semantic search runs inside a WHERE clause alongside standard SQL filters — no client-side result merging, no separate vector query, no second round-trip. An EMBEDDING MODEL column annotation in CREATE TABLE is all the configuration the inference sidecar needs; it handles queueing, back-pressure, and index updates on every INSERT without any external pipeline.

The differentiating feature is the time-travel and training-export combination. CREATE VERSION TAG pins the exact state of the database at a moment in time. AT VERSION queries replay that state months later. A single SQL command then exports that snapshot as a Lance dataset with zero-copy memory mapping into PyTorch — which means a training run is reproducible by definition, not by discipline. Teams that have spent engineering cycles rebuilding ‘what data did that model see?’ pipelines will recognize what that eliminates.

GalaxDB fits teams building a first or second AI application who want to avoid accumulating five separate services before the first user arrives. The vendor states the Cloud offering is in a coming-soon waitlist phase, so managed hosting is not yet available for production commitments. The self-hosted path is documented and the binary is downloadable, but v1.0-beta.1 status means teams accepting production risk should plan for schema or API changes before a stable release. Teams whose workloads require the operational maturity guarantees of Pinecone, managed RDS, or established S3 pipelines — SLA contracts, audit logs, enterprise support — will not find those here yet.

Benchmarks published on the vendor page show 0.990 recall@10 on HNSW over SIFT-1M at ef=200, 258K write TPS at 16 threads over 1M rows on NVMe, and 4.49 GB/s scan throughput using PAX blocks with zone-maps. The test suite covers 740 tests and 7 chaos scenarios. These numbers are vendor-reported and have not been independently verified at the time of this listing.