LanceDB

LanceDB (founded 2022, remote-first, backers include a disclosed $30M Series A) calls itself the AI-native multimodal lakehouse: an in-process database — the SQLite-shaped answer to vector search — built on the Lance columnar format, reading vectors, full-text indexes and blobs straight from S3-compatible object storage. Embedded means no server to babysit and no cluster bill while idle; the vendor claims 100,000+ queries per second, 100 billion+ rows in a single table, blob reads up to 68x faster than generic columns and storage cut around in half. LanceDB Cloud wraps the same library as serverless, usage-based infrastructure, and an Enterprise tier runs identical code against dedicated capability.

The judgement is economic. Your cost curve is object-storage requests plus compute time rather than always-on nodes, which is exactly why local-first apps, notebooks, edge deployments and RAG pipelines over existing buckets lean this way. The costs you do own are compaction and index rebuilds, and their payback lands in your engineering hours, not a vendor SLA. Residency is as good as your bucket's region and key management, but watch egress — retrieval beside the data is the point, retrieval across clouds is the bill. For extreme concurrent, low-latency serving, a hosted engine fronting the queries (Qdrant, Vespa) still wins; if retrieval should live beside the data, this is the library, with the caveat that its cloud console is younger than its rivals'.