Knowledge Base

A knowledge base lets a workflow answer from your own documents. You store documents in a collection, and a Retrieve block searches them at run time and hands the most relevant passages to a downstream agent as grounding. This is retrieval-augmented generation (RAG).

Collections and embedding models

A collection is a named set of documents that share one embedding model. The embedding model turns text into vectors so passages can be matched by meaning rather than exact words. You choose the model when you create the collection:

Standard embeddings run on our own infrastructure, use no credits, and work on every plan. They suit smaller documents, up to about 200 passages each.
Premium embeddings use a higher-quality hosted model, handle large documents, and are available on the Pro and Business plans. They are metered as managed AI credits, the same as any other managed model.

The model is fixed for the life of the collection, so every document in it and every query against it use the same one.

Add documents

Open Knowledge base from the Studio sidebar, create a collection, then add documents to it. A document can be:

Text you paste directly (every plan)
A public URL, fetched and indexed for you (Pro and Business)
A PDF or Word file you upload, up to 10 MB (Pro and Business)

Free workspaces ground agents with pasted text and the web source. URL and file uploads are a Pro and Business feature.

Each document is split into passages, embedded, and stored. Indexing runs in the background; a document shows as pending, then indexing, then ready with the number of passages it produced. Indexing with premium embeddings is metered as managed AI credits. A document too large for standard embeddings is rejected with a prompt to use premium embeddings.

Every document is screened by AI-powered content safety before it is indexed. Content that reaches a high severity for hate, violence, sexual, or self-harm material is rejected rather than stored, so a workflow cannot be grounded in harmful text.

Click a document to inspect it: pasted text opens in a reader, a URL document opens its page in a new tab, and an uploaded file downloads through a short-lived secure link. Use the edit control on an expanded collection to rename it or change its description; the embedding model can only change while the collection is empty.

Retrieve block

Press the + handle on the block you want it to follow and choose Knowledge from the palette. In its configuration panel:

Choose the source: your knowledge base, the web, or both.
For the knowledge base, pick one or more collections to search.
Set how many passages to return and the minimum similarity a passage must reach to be included.
Write the query, usually a reference to an earlier block’s output, such as a chat message or trigger value.

Wire the Retrieve block into an agent block downstream. At run time the retrieved context is placed in front of the agent so its answer is grounded in your documents.

Deleting and erasure

Deleting a document removes it and its indexed passages. Deleting a collection removes every document in it, its passages, and any uploaded files. Erasure covers both the stored documents and the vectors derived from them.

Limits

Uploaded files are capped at 10 MB and must be a PDF or Word file.
A URL document must be a public address; addresses that resolve to private or internal networks are blocked.
URL and file uploads, and premium embeddings, are Pro and Business features. A Free workspace can paste text, create standard-embedding collections, and use the web source. See Billing & Limits.

Next: Agent Blocks or Workflow Builder