LANCEDB

LanceDB embedded Rust+Python vector DB Lance format.

Définition

LanceDB features : (1) Embedded mode : in-process Python ou Rust, no server requise, file-based storage local filesystem ou S3/GCS/Azure Blob, ~1 line setup db = lancedb.connect('s3://bucket/lance'). (2) Lance format : columnar format developed LanceDB (alternative Parquet, optimized random access vectors + scalar fields), open-source format Apache-2.0, integration Pandas + Polars + DuckDB + PyArrow natively. (3) Index : IVF-PQ default, HNSW emerging, scalar BTREE indexes. (4) Multi-modal : native support text + images + audio + video embeddings, multi-vector per row. (5) Versioning : Time Travel queries (version-aware, query data as of specific version), append + delete + update transactions ACID. (6) Distance : L2, Cosine, Dot Product. (7) Pre-filtering : SQL WHERE clauses filter before vector search, integrated query optimization. SDKs : Python, TypeScript, Rust, Java emerging. Cloud : LanceDB Cloud managed offering 2024 beta. Customers : startups + indie developers focus + Roblox emerging adopter.

Origine

LanceDB fondee 2022 par Chang She (ex-Pandas core developer, Anaconda) + Lei Xu ; Seed $11M 2023 ; ~15000+ GitHub stars 2024 ; Lance format adopted by Voyage AI + others.

Exemple en contexte

Indie developer building RAG application for personal photos : LanceDB embedded mode, stocke photos local filesystem Lance format, embed via OpenCLIP vision model, multi-modal search 'find photos with mountains', integration Pandas DataFrame operations preserved.

Termes liés

• Chroma — alternative embedded.