LanceDB

LanceDB is an open-source, embedded vector database — it runs inside your process like SQLite does, storing data on disk (or object storage like S3) in the Lance columnar format. No server to deploy, no connection strings: you point it at a directory, and you get vector search, full-text search, SQL-style filtering, and versioned datasets. A managed LanceDB Cloud exists for when you want the same API without operating storage yourself.

What makes LanceDB more than “Chroma with a different file format” is that Lance format itself is built for multimodal AI data at scale. It stores vectors alongside raw data — text, images, metadata — in one place, supports fast random access, and handles datasets far larger than RAM because indexes and data are memory-mapped from disk. That makes it a genuine option not just for prototypes but for serious workloads: recommendation corpora, image search, and training-data management, all without a database server in your architecture diagram.

In practice you’ll pip install lancedb, then db = lancedb.connect("./data"), create a table from a list of dicts or a Pandas/Arrow table with db.create_table("docs", data), and query with table.search(query_vector).where("category = 'api'").limit(5).to_list(). It can call your embedding function automatically via its embedding registry, and hybrid search combines vector and BM25 full-text scores in one query. TypeScript and Rust clients mirror the Python API.