Ask your filing cabinet

The Knowledge node is in beta and available to every organization. Its behavior and API may change.

Extraction answers “what is on this document”. Knowledge answers “what do these documents say about X”, across the workflow’s current corpus. Completed runs made while a Knowledge node exists add their parsed documents to that corpus. Every answer comes back with the quotes it rests on, each resolved to a page and a region of the original PDF.

You ask

→

It answers

NodesCredits27 per pageYou get

The workflow

The node has no options. Its presence opts the workflow into knowledge ingestion when runs complete.

{
  "name": "Ask the supplier file",
  "description": "Index every document the workflow parses, then ask it questions",
  "nodes": [
    { "id": "parse_1", "type": "parse" },
    { "id": "knowledge_1", "type": "knowledge" }
  ],
  "edges": [{ "source": "parse_1", "target": "knowledge_1" }]
}

import os

from anyformat.sdk import Client, WorkflowDefinition
from anyformat.workflow.definition import Edge
from anyformat.workflow.nodes import KnowledgeNode, ParseNode

client = Client(api_key=os.environ["ANYFORMAT_API_KEY"])

workflow = client.create_workflow(
    WorkflowDefinition(
        name="Ask the supplier file",
        description="Index every document the workflow parses, then ask it questions",
        nodes=[
            ParseNode(id="parse_1", type="parse"),
            KnowledgeNode(id="knowledge_1", type="knowledge"),
        ],
        edges=[Edge(source="parse_1", target="knowledge_1")],
    )
)

The corpus is workflow-scoped, not version-scoped. It keeps the current parsed content for each live file across runs and versions. It is not a version history. If documents were processed before you add Knowledge, a member opens the workflow’s Knowledge tab and chooses Index existing documents. Use Retry indexing after a failure. Once a corpus exists it never needs a manual update: new or changed documents re-index in the background, and the existing corpus keeps serving in the meantime. These maintenance actions reuse processed documents and do not spend knowledge-index credits.

Ask it

Upload and run documents as you would with any workflow. Runs completed after Knowledge is present ingest normally. Then ask.

curl -X POST "https://api.anyformat.ai/v3/workflows/$WORKFLOW_ID/knowledge/ask" \
  -H "Authorization: Bearer $ANYFORMAT_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"question": "Which invoices are in this file, who issued each one, and what is the total of each?"}'

answer = workflow.ask("Which invoices are in this file, who issued each one, and what is the total of each?")

print(answer.answer)
for citation in answer.citations:
    print(citation.quote, "page", citation.page)

Response

The answer is markdown, written to be shown to a person. The citations are the point: each one carries the exact quote, the page, and a bbox in normalized page coordinates, so you can draw the box on the PDF instead of asking anyone to trust the answer. path shows which tools the agent used to get there, and steps_used how many it took.

{
  "answer": "The document `2026-08/mixed_partition.md` contains three invoices:\n\n*   **INV-2026-1000**, issued by **Acme Industries**, with a total of **$1,810.50**.\n*   **INV-2026-1001**, issued by **Globex Corporation**, with a total of **$932.45**.\n*   **INV-2026-1002**, issued by **Initech LLC**, with a total of **$5,648.15**.\n\nThe file also includes a shipping label and a boarding pass.",
  "citations": [
    {
      "path": "2026-08/mixed_partition.md",
      "quote": "Invoice #: INV-2026-1000",
      "file_id": "06a95794-11f6-77e1-8000-f49fa028d98e",
      "block_id": "p1_b1",
      "page": 1,
      "bbox": { "x0": 0.113, "x1": 0.325, "y0": 0.132, "y1": 0.21 }
    },
    {
      "path": "2026-08/mixed_partition.md",
      "quote": "**TOTAL: $1,810.50**",
      "file_id": "06a95794-11f6-77e1-8000-f49fa028d98e",
      "block_id": "p1_b9",
      "page": 1,
      "bbox": { "x0": 0.584, "x1": 0.756, "y0": 0.431, "y1": 0.483 }
    }
  ],
  "path": [
    { "tool": "kb_ls",   "args": { "folder": "." } },
    { "tool": "kb_read", "args": { "path": ["INDEX.md", "2026-08/mixed_partition.md"] } }
  ],
  "steps_used": 2,
  "thread_id": null
}

Follow-up questions

Mint a thread_id yourself, starting with kb-, and pass it to ask in the context of what came before. Reuse the same id to continue the conversation; omit it and every question stands alone.

{ "question": "Which of those is the largest?", "thread_id": "kb-invoices-review-1" }

When it goes wrong

409 KNOWLEDGE_NOT_ENABLED. The workflow has no Knowledge node. This is permanent until you add one: do not retry. 409 KNOWLEDGE_NOT_READY. No snapshot has ever been published for this workflow. Wait when the Knowledge tab shows Indexing documents. Otherwise, use Index existing documents or Retry indexing in that tab as appropriate. Once a snapshot exists, this error does not return: a stale corpus keeps serving while it re-indexes. The answer is right but you cannot prove it. Read the citations, not the prose. Each bbox is normalized to the page, so multiply by the rendered page size to draw it. An answer whose citations do not support it is the one to escalate. It answers about the wrong document. The corpus holds the current parsed content for every live file in the workflow. Give the question the scope you mean: name the supplier, the month or the file. The bill is bigger than you expected. Questions are metered on the text the agent actually reads, so a broad sweep over a large corpus costs more than a narrow question. Ask narrowly.

Examples

Read the page

Messy piles

Money and compliance

At scale

Ask your filing cabinet

The workflow

Ask it

Response

Follow-up questions

When it goes wrong

Next steps

Knowledge

Ask

​The workflow

​Ask it

​Response

​Follow-up questions

​When it goes wrong

​Next steps

Knowledge

Ask

The workflow

Ask it

Response

Follow-up questions

When it goes wrong

Next steps