AIP Document Intelligence Overview

AIP Document Intelligence serves as Foundry's central hub for all enterprise document extraction workflows. It enables users to easily open media sets of documents, quickly run state-of-the-art extraction strategies, and evaluate the quality, speed, and token cost of those strategies. Once configured, you can deploy the extraction strategy as a Python transform to a batch pipeline with a single click.

An example PDF document from a media set previewed in AIP Document Intelligence.

AIP Document Intelligence combines advanced generative AI and traditional OCR technologies with an intuitive, walk-up interface designed for rapid testing and deployment.

An example of document extracting in Document Intelligence, using bounding boxes to highlight distinct areas of extraction.

• Unified Interface: Easily search for and select Foundry media sets to process.
• Quick Confirmation Loop: Quickly verify extraction results using Markdown output mapped interactively to bounding boxes on the original PDF document.

• Traditional Extractions: Leverage layout-blind and layout-aware methods:
  • Raw text: Extracts underlying metadata. Only available for electronically generated PDFs.
  • OCR: Traditional Optical Character Recognition that pulls text without preserving visual layout information.
  • Layout-aware OCR: Advanced OCR with bounding boxes that preserves structural layout.
• Generative AI (VLM) Extractions: Use vision language models (VLMs) with system prompts fine-tuned to extract clean Markdown. You can modify these default prompts using your own custom logic in the configuration interface.
• VLM Preprocessing: Combine layout-aware OCR preprocessing with generative VLMs to handle highly complex documents. Learn more in the document preprocessing guide.

• Execution Metrics: Review execution time, input/output token consumption, and general quality scores.
• AI-Assisted Evaluations: Use a "VLM as a judge" to score extraction success across specific structural elements (e.g., tables, code blocks, lists). Learn more about extraction evaluations.
• One-Click Deployment: Seamlessly deploy your finalized strategy into a Python transforms repository. Learn more about extraction strategy deployment.

AIP Document Intelligence uses an iterative, test-and-adjust workflow. Users import a media set, select a configuration, evaluate their strategy, and deploy once satisfied.

A diagram of the standard workflow when using AIP Document Intelligence.

Select an existing media set from your available Foundry files or upload a new media set directly from the application landing page.

Open the Configuration tab to choose your extraction method (traditional OCR vs. generative AI VLMs), enable document preprocessing if needed, and customize your model prompts. Select Save to store these settings.

The Configuration tab in AIP Document Intelligence

With your configuration saved, select Run in the top right corner of the Configuration tab to run the extraction process on the media set.

Navigate to the Extraction result tab to review the output format and layout mapping.

From the results tab, select Evaluate results to initiate an LLM-based evaluation of your output. Repeat Steps 3 through 5, modifying prompts and models, until you achieve your target extraction quality.

An example of an evaluation of document extraction results.

In the Extraction result tab, select Chunk next to an extraction result to see how the text splits. You can adjust chunking parameters interactively and preview the changes in real-time.

Open the Deployment tab, select your saved configuration, and click Create transform repository.

• Name your new Python transforms repository and choose its location.
• Enable optional chunking and embedding settings.
• Specify your output dataset and initiate the build. Refer to the repository's auto-generated README.md for deeper instructions.

A Python transform repository, preconfigured with an AIP Document Intelligence template.

• Integrated Playground: AIP Document Intelligence bridges the gap between raw PDF documents and structured, pipeline-ready text data.
• Flexible Hybrid Methods: Supports traditional OCR, layout-aware OCR, pure VLM extractions, and hybrid preprocessed VLM workflows.
• Objective Evaluations: Provides built-in token cost tracking, execution times, and automated VLM grading for structural elements.
• Code-Free to Code-First: Users can prototype, prompt-engineer, and test entirely within a visual UI, then export the exact logic into a production-grade Python transforms repository with one click.