Parse

Parse a document with one provider using the TypeScript SDK.

Use Unfold for hosted jobs or DirectUnfold for direct provider calls. Both expose the same parse(input, options) shape.

const result = await client.parse("./invoice.pdf", {
  provider: "mistral-ocr",
  outputs: ["markdown", "pages", "tables"],
  pageFields: ["markdown", "tables"],
  pages: [1, 2, 3],
  timeoutMs: 120_000,
})

Inputs

The SDK accepts:

• A file path or public HTTP(S) URL as a string
• URL, File, or Blob
• ArrayBuffer or an array-buffer view
• ReadableStream<Uint8Array>
• Explicit file, URL, byte, or named stream input objects

Use a named stream when uploading from object storage so Unfold receives the original filename and content type without buffering the stream:

const result = await client.parse({
  data: object.body,
  kind: "stream",
  mimeType: object.httpMetadata?.contentType,
  name: "report.pdf",
})

Options

Option	Type	Purpose
provider	string	Provider ID to use
outputs	ParseOutput[]	Requested normalized outputs; defaults to markdown
pageFields	ParsePageField[]	Fields retained on each requested page
pages	number[]	One-based page selection
providerOptions	object	Namespaced provider-native settings
includeRaw	boolean	Include the provider response
timeoutMs	number	Abort the operation after this duration
signal	AbortSignal	Cancel from your application

pageFields requires outputs: ["pages"] (or another output list containing pages). Unfold always retains each page’s pageNumber and warnings.

Hosted calls also accept idempotencyKey. When omitted, the client generates a UUID for the request.

Hosted results expose the resource IDs created by the convenience call:

await client.documents.delete(result.resources.documentId)

Deleting the document also removes its retained job and execution results.

Failed hosted requests throw UnfoldError. Its requestId identifies the server request when you need support or are correlating application logs.

Unfold rejects outputs and page fields the selected provider does not advertise before starting the provider request. Use client.providers.list() to discover the hosted capabilities at runtime.