Ship documents people can verify and trust.

Models read and propose, your rules validate, people approve exceptions and signatures seal the final file. Citations, confidence scores, human approvals, permissions and a full audit trail are built into the same flow.

Citations: Cited to page, or declared absent

Turn on citations in Extract and every field returns its source file, page and bounding box, or its timecode in a recording. Missing values come back as null instead of a guess, and manual edits drop the model citation so the record stays clear on who wrote what.

invoice-4471.pdfSelect a field to see it in the document.
Page 1 of 1
Invoice INV-4471, Northwind Supply Co.

Northwind Supply Co.

Subtotal $13,518.52. Tax (8%) $1,081.48.
1

Extracted fields

TypeScript
const run = await client.extract({
  file: { url: "https://example.com/invoice-4471.pdf" },
  schema: invoiceSchema,
  citations: true, // off by default; ask for the evidence
});

// Every field is cited, or declared absent. Nothing in between.
for (const [field, cited] of Object.entries(run.output?.citations ?? {})) {
  for (const c of cited) {
    if (c.notFound) console.log(field, "not in the document");
    else console.log(field, c.page, c.bbox, c.text, c.confidence);
  }
}

Score: A confidence score on every output

CloudRaker gives every AI output a confidence score from 0 to 100%, so you know when to trust a result, verify it or step in.

  1. Watch quality

    Track scores across runs to see where the model is sure and where it is not.

  2. Set your thresholds

    Pick the score a result needs to pass on its own, per field or per workflow.

  3. Route the rest to people

    Send low-confidence results to extra validation or to human review.

  vendor      Northwind Supply Co.   99%
  due_date    2026-11-14             98%
  total       $14,600.00             99%
  ship_to     Dock 4, Portland       62%
  below 80%   ship_to sent to review

Approvals: Pause and resume for human review

Pause high-stakes steps in code so a reviewer can sign off, with up to seven days to respond. Use your own UI via the API, or turn on the hosted review screen in CloudRaker Run (beta).

  1. Pause on your gate

    Whenever an action requires sign-off or a human task, the run pauses in a waiting state until a reviewer acts.

  2. Show the source beside the draft

    Display extracted values, optional confidence scores and source citations side by side in your own UI or in CloudRaker Run (beta).

  3. Approve or reject to resume

    Submit the approval or rejection through the API to record the decision and resume the run from where it stopped.

Permissions: Every person and agent limited to what they are allowed

Give each teammate and each agent only the access its job needs. Reviewers approve, agents call the tools you grant, and every action is recorded against who took it.

  1. People

    Teammates see and act on what their role allows, from running workflows to approving exceptions.

  2. Agents

    Each agent gets the tools, files and channels you grant it, and nothing else.

  3. On record

    Every call, edit and approval carries the person or agent behind it in the audit trail.

Audit trail: Steps, inputs, outputs, decisions and cost

Query any run to inspect its inputs, step outputs and citations, or filter past runs by type, status and your own metadata. Runs expire automatically by default, or you can delete them at once or pin them for long-term retention.

  1. Every step

    Inputs, step outputs, citations and scores, in order.

  2. Every decision

    Approvals and edits, with the person or agent behind them.

  3. Every credit

    Cost per step and per run, on the same record.

$ curl https://api.cloudraker.com/v1/runs/plr_01KY...
  extract     completed   4 fields cited
  approval    approved    ada@example.com
  fill        completed   w9-filled.pdf
  credits     18

Open weights, open dataset and public benchmarks.

We publish rakedoc-nano-2 weights under AGPL-3.0, our synthetic document dataset under CC BY 4.0 and our submitted ParseBench run (80.1 overall; LlamaParse Agentic Plus leads overall at 86.7). Where a competitor leads or a figure is internal (such as rakeaudio-asr ~5% WER), we state it plainly.

Test citations and scores on one file.

Try cited extraction with confidence scores on your $15 of free monthly usage. Looking for SOC 2 Type 2, HIPAA or data residency?