Structured Data Extraction Pipeline

Trigger condition

Steps

1. 1

   Classify each document by type: invoice, contract, report, email, other

   tool: llm · output: doc_type_map

2. 2

   Apply type-specific extraction schema: route invoice→invoice_extractor, contract→contract_extractor, etc.

   tool: llm · output: extracted_fields

3. 3

   Validate extracted fields: type check, required field presence, cross-field consistency

   tool: python_pydantic · output: validation_report

4. 4

   Flag low-confidence extractions (LLM confidence < 0.85) for human review queue

   tool: python_script · output: review_queue

5. 5

   Write validated records to destination schema (PostgreSQL/BigQuery/CSV)

   tool: db_client · output: loaded_records

6. 6

   Generate extraction report: success rate, field-level accuracy, review queue size

   tool: python_script · output: extraction_report

Expected output

Tools required