Knowledge bases

A knowledge base is the container your indexed documents live in. When you run a query, the search is limited to the base you pick: documents from one base are never mixed with another's.

What each knowledge base decides

Creating a base fixes two things:

• The embedding model. It turns text into vectors so it can be compared by meaning. The model also determines the dimension of those vectors.
• The provider credential used to call that model. It is yours: the tokens indexing consumes are billed to your account with the provider.

Creating a knowledge base

From Knowledge bases, the create button asks for a name, a provider credential and a model. The model list is filtered by the provider of the credential you pick, so you cannot build an invalid combination.

If you have no credential registered yet, the form says so and takes you to Settings → Provider credentials: nothing can be indexed without one.

Keeping original documents

While creating the base you can turn on keeping the original files. If you do, you also have to pick a storage credential (they are configured under Settings → Storage).

With the option on, every document you upload is stored as well as indexed, and a download action appears in its row. Without it, the platform extracts the text, indexes it and discards the file.

Two details that tend to surprise people:

• The option only affects future documents. Turning it on does not recover the originals of what you had already uploaded.
• Turning it off deletes nothing. Originals that were already stored stay there and can still be downloaded.

Changing a base's credential

In the base's header, the change-credential action lets you pick another one. The selector only offers credentials from providers that serve that base's embedding model, which does not change.

This is what you need when you rotate a key or move billing to another account.

When to create several bases

Splitting into separate bases is worth it when:

• The content belongs to areas that are never queried together (technical documentation and HR documents, say). Narrowing the base improves result quality.
• You want to try a different embedding model without touching what already works.
• You need to separate billing by using different provider credentials.

You do not need one base per document: a single base holding many related documents works better than many small ones.

Deleting a base

Deleting a base also removes all of its documents and all of their indexed passages. This cannot be undone, which is why it asks for confirmation.

Next steps

Carry on with Documents, which explains how to upload files and what each indexing state means.

Updated on Aug 04, 2026