The node
client.workflow(...)) has no knowledge() method yet, so the example builds the graph as a WorkflowDefinition. The API needs an edge into the node like any other non-Parse node; the edge only satisfies the graph, because indexing always reads the whole run.
Ask a question
Once the first index has built, ask the corpus throughPOST /v3/workflows/{workflow_id}/knowledge/ask:
thread_id that starts with kb- to ask a follow-up in the context of earlier questions in that thread. Omit it and each question stands alone. The full request and response, with the error codes, is on Ask. An AI agent asks the same question through the ask_knowledge tool of the MCP server.
In Studio
Drag Knowledge from the palette and it appears as a band around the whole graph, not as a card, because its scope is the whole workflow. Its panel has nothing to configure; it explains what the node does and points you to the workflow’s Knowledge tab, where you browse the indexed documents and ask questions.
Options
The node has no fields.id and type are all it carries.
The schema the API accepts is generated from the same source: Knowledge node schema.
What it returns
Nothing in the run results. Knowledge emits nothing downstream and adds no section to the run. It writes to the workflow’s knowledge base instead: after each completed run, the latest parse of every live file is indexed into the corpus. The corpus is workflow-scoped, not version-scoped. It accumulates across every run and every workflow version, so a document parsed under an older version is still answerable. You read the corpus in two places:- The Knowledge tab in the app: browse the documents, search them, and ask questions.
POST /v3/workflows/{workflow_id}/knowledge/ask: an answer plus its citations, each resolved to afile_id, apageand abboxin the source PDF. Without a Knowledge node on the workflow the endpoint returns409 KNOWLEDGE_NOT_ENABLED; while the first index is still building it returns409 KNOWLEDGE_NOT_READY.
Connects to
Attach it at any depth of the pipeline. Which node feeds it makes no difference: indexing always reads the whole run. A workflow has at most one Knowledge node; the API rejects a second one.
Billing
- Indexing: 2 credits per page whose content is new or changed since the last run. Re-running the workflow over unchanged documents indexes nothing and costs nothing.
- Asking: 8 credits per 10,000 effective input tokens the answering agent reads. A narrow question over a small corpus costs a fraction of a broad sweep over a large one. A question the organization cannot pay for is refused with
402before it runs.
Examples
- Contract analysis: a contract workflow whose documents have lasting value. Add Knowledge and “which suppliers have a 90-day termination clause?” becomes answerable with citations.
- Invoice processing: the usual Parse to Extract shape. Add Knowledge after Extract to keep every invoice searchable.

