Document Processing Flow

Last verified: 2026-07-26 against dev at f0388abb.

This page explains how documents move from upload to extracted, reviewable output.

Flow

1. User creates an assessment.
2. User uploads one or more documents.
3. Backend validates files and stores document metadata/content.
4. Documents are grouped into batches.
5. A batch run selects a processor.
6. The processor identifies an extractor/schema pair.
7. Backend queues extraction work for the relevant extractor app.
8. The extractor processes the document and returns structured data.
9. Backend stores extracted items and fields.
10. Analyst reviews, corrects, validates, and exports output.

Schemas, Extractors, and Processors

• A schema defines the target structured fields.
• An extractor implements how to read a document and produce fields.
• A processor combines one schema and one extractor.

This separation lets the product run different extraction strategies against the same or similar document sets.

Smart Chunking

For large ADI documents, smart chunking can split layout content into smaller page groups before LLM extraction. The router maps schema objects to likely page groups, then extraction runs against those groups and merges results.

See Smart chunking.

Review Loop

Corrections and validation let analysts improve final structured outputs without treating the raw extraction as automatically authoritative.