Prebuilt Schema

Prebuilt schemas are ready-to-use schema definitions for common document types. They define fields and instructions so you can start extraction without building a schema from scratch. Use them as-is or customize (add fields, modify instructions) for your use case.

Available Prebuilt Schemas

Each prebuilt schema includes field definitions (names, types, required flags) and instructions that guide the LLM for zero-shot extraction.

Using Prebuilt Schemas

Fetch and apply via API:

# List prebuilt schemas by category
curl -X GET "/api/extract/schemas/prebuilt?category=invoice"

# Use a prebuilt schema for extraction
POST /api/extract
{ "document_id": "...", "schema_id": "prebuilt:invoice" }

Use a prebuilt schema as-is for common document types, or copy and customize for your specific needs. Customization includes adding fields, modifying instructions, or changing field types.

Form Detection

DocLD can auto-detect document type and suggest a prebuilt schema:

1. Process document — Document is classified during pipeline processing, or on first request to the suggested-schema endpoint
2. Store — Classification is stored in document metadata
3. Recommend — GET /api/extract/suggested-schema?document_id=... returns the suggested schema (e.g., "Invoice", "Contract")
4. Apply — Use the recommended schema for extraction

Suggested schema accelerates setup when you have mixed document types. Use it for batch processing to route documents to the right schema.

Customization Flow

Instructions help the AI handle edge cases (e.g., "If tax is shown separately, extract as its own field"). Refine instructions based on extraction quality and ground truth feedback.

Best Practices

• Start with prebuilt — Use prebuilt schemas for common types before building custom
• Customize for edge cases — Add instructions for document variants (e.g., different invoice formats)
• Validate with ground truth — Use ground truth to measure accuracy of prebuilt vs custom schemas

Related Concepts

Prebuilt schemas are ready-made schemas for extraction. Zero-shot extraction works with prebuilt schemas without training. Ground truth measures extraction accuracy.