Skip to main content
A LanceDB table becomes nested folders, one level per group-by column. Each row is a Markdown card plus an optional blob file, such as its image.

Setup

A local directory needs no setup: LanceDB stores the table there, with no server. Object storage and LanceDB Cloud need credentials.
uri also takes s3://, gs:// and az:// URIs. Pass the store’s credentials or endpoint in storage_options.
  1. In the LanceDB Cloud UI, open the project and copy its database URI (it starts with db://) and an API key.
  2. Note the project’s region; region defaults to us-east-1.
  3. LanceDB Enterprise: the LanceDB team provides the URI, the API key and the endpoint that goes in host_override.
  4. In the config, set uri: ${LANCEDB_URI}, api_key: ${LANCEDB_API_KEY} and region: ${LANCEDB_REGION}.

Connect

workspace.yaml

Layout

  • Without table, every table is a top-level folder.
  • A card lists the row’s columns, without the vector and blob columns.
  • The search CLI ranks rows by meaning and returns their <id>.md paths, search_limit (10) by default. The table needs an embedding function on a source column.

Sample runs

The mount is read-only. One folder listing reads at most max_rows (1,000) rows, so keep group_by to low-cardinality columns.

SDK guides

Python

Every config field, the VFS.md and write support.

TypeScript

Every config field, the VFS.md and write support.