extractfold

extractfold is the structured-data sibling of docfold.

• docfold: document -> representation, such as Markdown, HTML, text, or layout.
• extractfold: document + JSON Schema -> schema-conformant Python dict.

The core package has no required third-party dependencies. Engines that need model weights, cloud SDKs, or provider SDKs live behind optional extras and import those dependencies lazily.

Install

pip install extractfold
pip install "extractfold[provider_router]"
pip install "extractfold[llm_structured]"
pip install "extractfold[all]"

Quick Start

import asyncio

from extractfold.engines import LLMStructuredEngine

schema = {
    "type": "object",
    "required": ["invoice_id", "total"],
    "properties": {
        "invoice_id": {"type": "string"},
        "vendor": {"type": "string"},
        "total": {"type": "number"},
    },
}

async def main() -> None:
    result = await LLMStructuredEngine().extract("invoice.pdf", schema)
    print(result.data)

asyncio.run(main())

CLI

extractfold extract invoice.pdf --schema invoice --engine llm_structured
extractfold extract invoice.pdf --schema schema.json --out result.json
extractfold compare invoice.pdf --schema schema.json --engines llm_structured,lift
extractfold list-engines
extractfold benchmark ./dataset --engines llm_structured,lift --out results.json

Engine Comparison

How To Choose

Engine Examples

from extractfold.engines import (
    AzureDocIntEngine,
    DocfoldLLMEngine,
    FenicEngine,
    GoogleDocAIEngine,
    InstructorEngine,
    LiftEngine,
    LLMStructuredEngine,
    LlamaExtractEngine,
    NuExtractEngine,
    ProviderRouterEngine,
    TextractEngine,
)

async def provider_call(**kwargs):
    # Call an application-owned gateway and return JSON-compatible data.
    return {"invoice_id": "INV-001", "total": 125.5}

result = await ProviderRouterEngine(provider_call=provider_call).extract("invoice.txt", "invoice")
result = await LiftEngine().extract("invoice.pdf", "invoice")
result = await NuExtractEngine(backend="hf").extract("invoice.txt", "invoice")
result = await LLMStructuredEngine(provider="anthropic").extract("invoice.pdf", "invoice")
result = await LLMStructuredEngine(provider="openai").extract("invoice.pdf", "invoice")
result = await LLMStructuredEngine(provider="gemini").extract("invoice.pdf", "invoice")
result = await InstructorEngine().extract("invoice.pdf", "invoice")
result = await FenicEngine().extract("invoice.txt", "invoice")
result = await LlamaExtractEngine().extract("invoice.pdf", "invoice")
result = await AzureDocIntEngine().extract("invoice.pdf", "invoice")
result = await GoogleDocAIEngine().extract("invoice.pdf", "invoice")
result = await TextractEngine().extract("invoice.pdf", "invoice")
result = await DocfoldLLMEngine().extract("invoice.pdf", "invoice")

Evaluation

extractfold.evaluation scores prediction folders shaped like this:

dataset/
  gold/invoice.json
  predictions/llm_structured/invoice.json
  predictions/lift/invoice.json

Metrics include field accuracy, schema compliance, precision/recall/F1, hallucination rate, type correctness, normalized value matching, and nested array alignment.

Development

pip install -e ".[dev]"
ruff check src/ tests/
mypy src/
pytest tests/ -m "not integration"