LanceDB

LanceDB is an open source vector database and multimodal lakehouse for AI teams, run as an embedded library or as an Enterprise cluster in a private cloud.

LanceDB keeps the data behind an AI model in one place, so curating a training set, engineering features and serving retrieval stop being separate copies in separate systems. Records of any modality sit in tables beside their embeddings. The tables are written in Lance, an open source file format. LanceDB Inc., based in San Francisco, develops both the database and the format. Search starts with approximate nearest-neighbour indexes from the IVF family, including IVF_HNSW_SQ, with quantization to shrink them. Full-text search sits beside vector search, and hybrid queries merge the two before a reranker from Cohere, Sentence Transformers or another provider reorders the results. Metadata filters can run before the vector search, and one table can hold several named vector columns, such as a text and an image embedding. The embedding function registry generates vectors during ingestion and again at query time. Tables keep versions with tags and can be restored to an earlier one, while Geneva, the Enterprise component, runs Python functions to compute new feature columns. The database suits machine learning teams that curate training data and run retrieval on the same tables, and developers who want a vector store inside their own application. LanceDB OSS is an embedded library under the Apache 2.0 license that a team runs itself. LanceDB Enterprise is the production edition, deployed as a private cloud or in the customer's own cloud account, with indexing managed by the platform. Enterprise is arranged directly with the LanceDB team rather than through a self-serve checkout. Typical work includes deduplicating a dataset before a training run, computing features at scale and answering retrieval queries for a RAG application or an agent. Embeddings come from the provider a team already uses, with integrations for OpenAI, Hugging Face, AWS Bedrock, Gemini and Ollama among others. A Parquet dataset can become a searchable table in one step. LanceDB fits teams who want the dataset that trains a model and the index that serves it to be the same table.