Extraction

Structured Data Extraction#

After a document has been OCR-processed and classified, the extraction stage pulls out structured data: key-value fields, tables, and dates. The extractor uses an LLM to analyze the raw text and return typed, confidence-scored fields.

This transforms unstructured document content into machine-readable data that can feed into databases, workflows, and downstream systems.

Running Extraction#

curl -X POST "http://localhost:8008/api/v1/docs/extract?document_id=DOC_ID"

Response:

{
  "fields": [
    {"key": "full_name", "value": "Ahmed Al Maktoum", "confidence": 0.97, "field_type": "text"},
    {"key": "id_number", "value": "784-1990-1234567-1", "confidence": 0.99, "field_type": "text"},
    {"key": "nationality", "value": "UAE", "confidence": 0.96, "field_type": "text"},
    {"key": "expiry_date", "value": "2028-03-15", "confidence": 0.94, "field_type": "date"},
    {"key": "date_of_birth", "value": "1990-07-22", "confidence": 0.93, "field_type": "date"}
  ],
  "tables": [],
  "dates": ["2028-03-15", "1990-07-22"],
  "summary": "Emirates ID card for Ahmed Al Maktoum, expiring March 2028"
}

Response Fields#

Field Object#

Extraction by Document Type#

The extractor adapts its behavior based on the document classification:

Identity Documents#

Financial Documents#

Government Documents#

Table Extraction#

For documents containing tabular data (invoices with line items, financial statements with rows), the extractor identifies and structures table content:

{
  "tables": [
    {
      "headers": ["Item", "Quantity", "Unit Price", "Total"],
      "rows": [
        ["Office supplies", "10", "25.00 AED", "250.00 AED"],
        ["Printer cartridge", "2", "150.00 AED", "300.00 AED"]
      ]
    }
  ]
}

Validation#

After extraction, the validation step checks for completeness and format correctness:

curl -X POST "http://localhost:8008/api/v1/docs/validate?document_id=DOC_ID"

{
  "is_valid": true,
  "completeness_score": 0.85,
  "issues": [
    {"field": "nationality", "issue": "Missing required field", "severity": "warning"},
    {"field": "expiry_date", "issue": "Date is in the past", "severity": "info"}
  ]
}

Validation Checks#

Chat RAG Ingest Pipeline#

Docs AI integrates with Anar Chat's knowledge base through an ingest pipeline. Extracted and validated documents can be pushed to Chat's RAG system, making document content searchable and retrievable via the AI assistant.

This enables workflows where a citizen uploads a document, Docs AI processes it, and Chat can immediately answer questions about the document's content.

Configuration#