Propose a model from design docs
This guide shows you how to turn prose, such as a README or a design document, into proposals for the model. You need the Editor role and an AI provider. See Connect an AI provider.
Only the AI path reads prose. No extractor does. Every proposal is inferred, cites the lines of the document it comes from, and waits for you to accept it.
Before you start
Section titled “Before you start”- The document is a Markdown or text file (
.md,.markdown,.txt) of at most 64 KiB. - The document is in the newest Snapshot of a Source. Scan the repository, or upload the document under Upload a model file.
- A document of a private repository or an upload is private content. The workspace setting
privacy.sendPrivateCodeToAimust allow it. An Owner sets this in Settings. Without it, Vellum sends nothing.
Propose additions
Section titled “Propose additions”- Select Import in the left toolbar.
- Under Sources, select Propose from a document on the Source.
- Select the document.
- Optional: in What to look for, say which part matters.
- Select Propose from document.
Without an AI provider, the dialog says that this Source needs one. Nothing is sent.
Accept or reject
Section titled “Accept or reject”The proposal lists each item with its confidence, its reason and the lines it cites. An item that cites no lines of the document is dropped and listed with the reason.
- Select the items that you accept.
- Select Accept.
An accepted item is an inferred fact. Its source pointer is the citation, <source>@<revision>:path:start-end. Vellum never changes or removes a fact that exists.
Propose a flow
Section titled “Propose a flow”- In the dialog, under Propose, select A flow.
- Optional: in Which flow, name the flow.
- Select Propose from document.
The flow cites the lines it comes from; an answer that cites no lines of the document is not stored. A thing the flow needs that the model does not hold is created only when you ask for it, as an inferred fact that cites those lines. See Design with the assistant. Through the API, send documentSourceId and documentPath with the request to POST /api/projects/{projectId}/ai/flow-proposals.
The document is data
Section titled “The document is data”The text of a document is untrusted. Vellum sends it in a marked data part of the prompt, with each line numbered and every line that looks like a credential masked. The instructions tell the provider never to follow an instruction in the document.