LanceDB iconLanceDB

Open-source multimodal database on the Lance columnar format for vector, full-text and SQL search plus AI training data

11k+ GitHub Stars
$30M Series A (2025)
Apache 2.0 License

Overview

LanceDB is an open-source database built on Lance, a columnar file format designed for AI data such as embeddings, images, video and text. It can run embedded inside your application (like SQLite, with no server) or as a managed service, and it stores data directly on local disk or object storage like S3. A single table supports vector search, full-text search, hybrid search and SQL filtering, and the same data can be used to curate and load training datasets.

That training-data angle is what sets it apart: LanceDB pitches itself as a multimodal lakehouse rather than just a vector database. It raised a $30M Series A in June 2025 led by Theory Ventures, and lists Netflix, Runway, Midjourney and Character.AI among its users.

The Verdict

Who Should Use LanceDB?

Best For

  • Multimodal data (images, video, audio, text) at scale
  • Embedded vector search in local or edge apps
  • Teams that want storage on S3 rather than in RAM
  • ML teams sharing data between search and training

Not Ideal For

  • Teams that want a classic client-server database only
  • Workloads needing the most mature managed service
  • Simple prototypes where pgvector is already available

What's Great

  • Embedded mode: vector search with no server to run
  • Object-storage architecture keeps costs low at scale
  • Vector, full-text and SQL filtering on one table
  • Handles very large multimodal datasets
  • Netflix reported large latency gains after migrating

Watch Out For

  • Smaller ecosystem than Pinecone or Milvus
  • Lakehouse model has a learning curve
  • Managed and enterprise pricing is not fully public
  • Embedded mode needs care with concurrent writers

Pricing

View all features & details

Key Features

  • Lance columnar format with fast random access
  • Vector, full-text and hybrid search
  • SQL filtering and versioned tables
  • Embedded or server deployment
  • Storage on local disk, S3, GCS or Azure
  • Training data loading and curation

Integrations

  • Python, TypeScript and Rust SDKs
  • LangChain, LlamaIndex
  • Pandas, Polars, Arrow, DuckDB
  • Ray and Spark

How It Compares

Feature LanceDB Chroma Milvus turbopuffer
Embedded mode Yes Yes Milvus Lite No
Object storage native Yes Cloud only Partial Yes
Multimodal and training data Core focus No No No
Open source Yes Yes Yes No
Full-text search Yes Yes Yes Yes

Is this your tool? Add a badge