Quickstart

This guide will help you get started with Document AI.

Installation

Requirements

Python >= 3.10
An API key for your chosen LLM provider (OpenAI, Anthropic, or Gemini) — or a local Ollama server

Install with uv

uv pip install doc-intelligence

Or with pip:

pip install doc-intelligence

Environment Setup

Document AI supports multiple LLM providers. Set up your API key for your chosen provider:

# OpenAI
echo "OPENAI_API_KEY=your-api-key-here" > .env

# Or Anthropic
echo "ANTHROPIC_API_KEY=your-api-key-here" > .env

# Or Gemini
echo "GOOGLE_API_KEY=your-api-key-here" > .env

Basic Usage

Here's a simple example to extract structured data from a PDF document.

Create a PDFProcessor once with your pipeline configuration. The document and schema are always provided per call:

from dotenv import load_dotenv
from pydantic import BaseModel

from doc_intelligence import PDFExtractionMode, PDFProcessor

load_dotenv()

class License(BaseModel):
    license_name: str

processor = PDFProcessor(
    provider="openai",
    model="gpt-5",
    include_citations=True,
    extraction_mode=PDFExtractionMode.SINGLE_PASS,
)

result = processor.extract(
    "https://example-files.online-convert.com/document/pdf/example.pdf",
    License,
)

# Or from a local file:
# result = processor.extract("path/to/your/document.pdf", License)

# Different schema, same processor
# result = processor.extract("receipt.pdf", Receipt, page_numbers=[0])

print(f"Extracted data: {result.data}")
print(f"Metadata: {result.metadata}")

Sample Output

The extract method returns an ExtractionResult with two attributes:

.data: The extracted data as a Pydantic model instance
.metadata: Citation information for each field with values and bounding boxes

result.data
# License(license_name='Attribution-ShareAlike 3.0 Unported')

result.metadata
# {
#     'license_name': {
#         'value': 'Attribution-ShareAlike 3.0 Unported',
#         'citations': [{
#             'page': 0,
#             'bboxes': [{
#                 'x0': 0.201,
#                 'top': 0.859,
#                 'x1': 0.565,
#                 'bottom': 0.872
#             }]
#         }]
#     }
# }

Configuration Reference

PDFProcessor constructor — pipeline config, set once:

Parameter	Description	Default
`provider`	LLM provider (`"openai"`, `"anthropic"`, `"gemini"`, `"ollama"`)	required*
`model`	Model name for the provider	provider default
`include_citations`	Include citation bounding boxes in results	`True`
`extraction_mode`	`SINGLE_PASS` or `MULTI_PASS`	`SINGLE_PASS`
`llm_config`	Generation parameters (e.g. `{"temperature": 0.2}`)	`None`

* Or pass a pre-built llm= instance instead of provider.

processor.extract() — document-specific, vary per call:

Parameter	Description
`uri`	Path or URL of the PDF (required)
`response_format`	Pydantic model class for the extraction schema (required)
`page_numbers`	List of 0-indexed page numbers to process (default: all pages)