Document AI field guide · reviewed August 23, 2026
Document AI, from readable files to reviewable structured data
Document AI is not just OCR. A usable workflow reads a supported document, maps it to a defined field schema, validates the returned shape, and routes uncertain or invalid results for review before downstream use.
Direct answer
Use raw OCR when text is the endpoint, a template parser when layouts are fixed, a schema-first API when business fields stay stable across layout variation, and a custom pipeline when you need to own the models, infrastructure, and exception logic. Whichever path you choose, test your own documents and define a human-review branch.
Core model
The four layers of a document AI workflow
Keeping these layers separate makes evaluations clearer. OCR quality, field extraction, type validation, and system integration answer different questions and should not be collapsed into one unsupported “accuracy” number.
| Layer | Question it answers | Operational output |
|---|---|---|
| 1. File intake and OCR | What text and layout can be read from this PDF or image? | Text, layout, or a visual representation for the extraction step. |
| 2. Extraction schema | Which business fields should the output contain? | Named, typed fields such as invoice_number, total_amount, and invoice_date. |
| 3. Validation and review | Does the returned shape pass type rules, and which reported scores fall below the schema threshold? | A completed or review branch plus validation details for the caller to inspect. |
| 4. Workflow mapping | Where should accepted fields go, and who resolves exceptions? | Your mapping into a database, accounting system, review queue, or another destination. |
Start here
Two practical guides
Implementation guide
Custom schema extraction
Follow one source-backed invoice example from canonical field definitions to the multipart request, response envelope, and review branch.
Open the worked guide →Decision guide
Build vs. buy document AI
Compare four tooling categories by output boundary, ownership burden, fit, and the question your evaluation must answer.
Use the decision matrix →Dokyumi’s current path
Schema-first extraction on one shared endpoint
- 1
Define fields
Create a schema in the dashboard. Field definitions use named keys, supported types, required flags, and optional descriptions or validation rules.
- 2
Select the schema
Send a supported file and the schema slug as multipart form fields to POST /api/v1/extract with a Bearer API key.
- 3
Inspect the envelope
Read data, confidence, validation, and meta. Confidence is model-reported and may omit fields; it is not a published accuracy benchmark.
- 4
Branch on status
completed means there are no type-validation errors or known reported scores below the schema threshold. review means inspect both validation.errors and validation.low_confidence_fields.
Apply the model
Use cases, field guides, and deeper reading
Invoice parser field guide
Fields, request shape, and review handling for invoice workflows.
Accounts payable
Map invoice fields into an AP review and posting process.
Real estate and leasing
Extract lease, property, tenant, date, and amount fields without claiming native property-system connectors.
Schema design article
Choose fields and types before automating the workflow.
Evaluation and review
Design representative tests and exception handling instead of relying on one universal score.
Security and data handling
Review the published controls, processor disclosures, and deletion-request limitations.
Evidence notes
Sources and limitations
Sources used
- Dokyumi API documentation — request fields, response envelope, file limits, status handling, rate limits, OCR modes, and webhook boundary
- Dokyumi pricing — current plan allowances, including 25 Free credits and page-weighted credit use
- Security and data handling — published storage, processor, training-use, and deletion-request disclosures
- Invoice parser guide — document-type field and workflow example using the current API envelope
Limitations
- Dokyumi has not published a held-out accuracy benchmark by document class; no universal extraction percentage is claimed here.
- Confidence values are model-reported signals, not calibrated probabilities or proof that a field is correct. The map may omit fields.
- Schema validation checks the returned shape and types. Required fields can still be null, and validation does not independently prove that every value matches the source document.
- JSON and upload-site webhooks are integration boundaries, not claims of native connectors to accounting, property-management, insurance, or other systems.
Published and last reviewed August 23, 2026. Product behavior can change; the linked API, pricing, and security pages are the controlling public references.