Document Fetch

Retrieve documents already in the conversation or in storage.

Document Fetch retrieves one or more documents previously uploaded in the same conversation or saved to persistent storage, and enforces strict access control while doing it. It works from the documentId that Document Upload returned.

It fetches in batches — you pass a list of document descriptors and get a result per document, so one bad ID doesn't sink the rest.

How it works

1. Input resolution — uses $input (the previous skill's output) and $secret (vault secrets).

2. Processing — accepts an array of descriptors (documentId, source, storagePath?) and performs a batch fetch:

   • documentId + source='conversation' → fetches from the current conversation.

   • documentId + source='storage' → fetches from storage, at storagePath if given.

   • Documents from other conversations are not accessible.

3. Execution — blocking. All documents are retrieved within the one step.

4. Response:

   • Success — an array of results, each with content, metadata, source, and status.

   • Partial — some documents succeed and others fail (not found, unauthorized). The overall statusCode is 207.

   • Failure — a missing or invalid documents array returns 400.

Worked example: retrieving customer agreements for review

A company manages customer onboarding through a workflow. Along the way, agreements — service contracts, privacy consents — are uploaded and saved to persistent storage. Later a customer calls with a question, and the agent needs their specific signed agreement to answer accurately.

The problem. The agreements are stored, but finding and opening the right one by hand is slow. The agent needs an automated way to retrieve a specific document in response to a customer query.

The approach. Give the agent a Document Fetch skill configured against persistent storage. Pass the documentId for the agreement — resolved from the customer record, or carried forward from an earlier step — and the skill returns a signed URL the agent can work from.

Because the skill accepts an array, one call can retrieve the contract, the consent, and the ID proof together, then branch on each result's own statusCode.

Configuring which documents to fetch, and their source

Document Fetch data flow

Configuration reference

Field	Type	Required	Description
documents	array	✅	Documents to fetch. Each item: { documentId, source, storagePath? }.
name	string	—	Display label on the canvas.
description	string	—	Long-form help text.

Output

Field	Type	Always	Description
results	array	✅	One object per requested document.
statusCode	number	✅	Overall: 200 · 207 partial · 400/500.
error	string | null	—	Present if the overall execution failed.

Each entry in results:

Field	Type	Description
documentId	string	The ID that was requested.
url	string | null	Signed URL to download the document.
mimeType	string | null	MIME type, e.g. application/pdf.
fileName	string | null	Original filename.
source	string	conversation or storage.
statusCode	number	Status for this document — 200, 404, 403.
error	string | null	Error message for this document, if any.

Check each result's own statusCode, not just the overall one. A 207 means the call worked but some documents didn't — the failures are only visible per item.

Errors

Condition	Status	Notes
Empty or missing documents array	400	Invalid input; the skill refuses to run.
Document not found	404	Per item, inside results.
Unauthorized cross-conversation access	403	Isolation is enforced, not advisory.
Internal error	500	Unexpected exception.

Security

• Cross-conversation access is strictly forbidden. A document uploaded in one conversation cannot be fetched from another.

• $secret may be used for document ID lookup, not for content.

• Logs redact document content — only IDs, status codes, and error messages are recorded.

• All storage-based fetches are validated against documentId.

Testing in isolation

• Path — /skill-runtime/workflows/nodes/DocumentFetch/execute

• Method — POST

• Body:

{
  "nodeType": "DOCUMENT_FETCH",
  "config": { },
  "input": { }
}

---

To add this skill to an agent, see Adding a Skill to the Agent.