Opportunity Relay White-label Next.js & AI delivery Book a 20-minute fit call
RCR Systems / Projects 04 · Document intelligence Español Open lab ↓

AI extraction · Exact evidence · Human approval

AI documents in.
Verified data out.

FieldDoc turns delivery notes and work orders into structured records without hiding the original. Every populated field must point back to literal source text.

Run a document
INPUT / DN-2041 DELIVERY
NOTE
Evidence stays attached
  1. 01 Captured
  2. 02 Extracted
  3. 03 Validated
  4. 04 Approved

Working project / not client data

The source remains the final authority.

Model output is only a proposal. The API validates a closed schema, clears any value whose evidence is not an exact substring, checks totals with deterministic code and keeps export behind a review step.

Inspect the architecture ↓

Try it · Scenario · Trace

Evidence inspector

Ready · no document leaves this page until you analyze it
  1. 1Choose or pasteUse fictional or de-identified document text.
  2. 2Analyze safelyConfirm one-time processing, then run extraction.
  3. 3Compare and correctOpen the evidence and fix any proposed value.
  4. 4Approve and exportHuman review unlocks the reviewed JSON.
Original / untrusted inputSource document
194 / 8,000 characters
Evidence preview 
DELIVERY NOTE DN-2041
Date: 2026-07-12
Supplier: Bay Parts LLC
Customer: Harbor Dental
Order: PO-88
Currency: USD

1 x Filter pack @ $120.00 = $120.00
Subtotal: $120.00
Tax: $8.40
Total: $128.40
Closed schema / proposed recordReview extracted fields
Verified0
Review0
Missing0
TypeNot analyzed

Choose a sample or paste a document, then analyze it. Results always remain proposals until you approve them.

Architecture / bounded by design

The model extracts.
The system decides.

  1. 01 / IntakeValidate before processing

    Same-origin requests, strict content type, 8,000-character input cap, rate limits and no public API key.

  2. 02 / ExtractionConstrain the output

    A low-cost model returns one closed schema. The document is explicitly treated as untrusted data.

  3. 03 / EvidenceReject unsupported values

    Every populated field needs an exact excerpt. A server-side post-validator clears anything that cannot be located.

  4. 04 / ControlKeep approval human

    Totals and duplicate checks are deterministic. Export unlocks only after an explicit source comparison.

Acceptance matrix

Designed for the document that does not cooperate.

Developer

Schema and hostile input

Unexpected properties, prompt injection, oversized bodies, false evidence, repeated requests and upstream failure.

With a manual

Repeatable operations

Load, compare, correct, approve and export while every decision remains visible in the same workspace.

Without a manual

Prepared first move

Three labelled samples, a single primary action, plain statuses and a recovery path that never destroys the original.

Need document work to stop at “review,” not at “magic”?

Build a document flow your team can verify.

Map my workflow ↗