Lancedb

Python SDK: pip install lancedb. Connect: import lancedb; db = lancedb.connect("./lancedb_data") or lancedb.connect("s3://bucket/path"). Create table: tbl = db.create_table("docs", data=[{"vector": [0.1, 0.2, ...], "text": "...", "id": "1"}]). Open table: tbl = db.open_table("docs"). Add data: tbl.add([{"vector": [...], "text": "...", "id": "2"}]). Search: results = tbl.search([0.1, 0.2, ...]).limit(5).to_pandas(). Filtered search: tbl.search(query_vec).where("category = 'tech'").limit(5). Full-text search: tbl.create_fts_index("text"); tbl.search("query text", query_type="fts"). Hybrid: tbl.search(query_vec, query_type="hybrid"). Schema: uses PyArrow schemas or inferred from data. JS client: npm install @lancedb/lancedb.

No auth for local/embedded mode — access controlled by filesystem permissions. S3/GCS backend: uses AWS/GCP credentials (env vars or instance roles). LanceDB Cloud: API key auth — Authorization: Bearer {api_key}. HTTPS enforced for Cloud. No built-in multi-user RBAC for embedded mode — single-process access assumed. Concurrent writes: LanceDB uses optimistic concurrency; concurrent reads are always safe. Cloud isolation: separate databases per project.

Embedded mode: no network failures — pure local I/O. S3 backend: network latency on reads/writes; implement retry logic for transient S3 errors. Versioning: LanceDB tables are versioned (Lance format) — query older versions with tbl.checkout(version). Compaction: run tbl.compact_files() periodically for optimal read performance. Index building: IVF-PQ index creation is synchronous — may take seconds to minutes for large tables. Memory: large tables can exceed RAM — Lance uses memory-mapped I/O for efficient access. Table updates: tbl.update({"category": "new"}, where="id = '1'"). Delete: tbl.delete("id = '1'").

lancedb.github.io/lancedb covers Python/JavaScript API reference, embedding integrations, cloud storage guide, full-text search, and LanceDB Cloud. Getting started: pip install lancedb — first vector search in under 3 minutes. LanceDB Cloud: free tier. Self-hosted: just pip install + local filesystem. GitHub: lancedb/lancedb (6k+ stars). Community via Discord (active). Good documentation; LanceDB has one of the most accessible embedded vector DB experiences.

LanceDB is the embedded vector database for agent pipelines that do not want to run a server — import the Python library, open or create a table, add embeddings, and query by vector similarity. No daemon, no Docker, no network hop. Persistent storage uses the Lance columnar format (Arrow-compatible), stored to local disk or cloud object storage (S3/GCS). Supports full-text search, hybrid search (vector + FTS), and metadata filtering. For agents: ideal for local RAG pipelines, edge deployments, or serverless functions where spinning up a database server is impractical. LanceDB Cloud provides a hosted serverless option. Confidence is docs-derived.