Extract structured data from patient intake forms with DSPy (v1)

We appreciate a star ⭐ at CocoIndex Github if this is helpful.

This example shows how to use DSPy with Gemini 2.5 Flash (vision model) to extract structured data from patient intake PDFs using CocoIndex v1.

• Pydantic Models (main.py) - Defines the data structure using Pydantic for type safety
• DSPy Module (main.py) - Defines the extraction signature and module using DSPy's ChainOfThought with vision support
• CocoIndex v1 App (main.py) - Wraps DSPy in a custom function, processes files incrementally, and writes results to JSON files

Key Features

• Native PDF Support: Converts PDFs to images and processes directly with vision models
• DSPy Vision Integration: Uses DSPy's Image type with ChainOfThought for visual document understanding
• Structured Outputs: Pydantic models ensure type-safe, validated extraction
• No Text Extraction Required: Directly processes PDF images without intermediate markdown conversion
• Incremental Processing: CocoIndex handles batching and caching automatically
• Portable outputs: JSON results stored locally under output_patients/

Run

1. Install dependencies

Install from the project's pyproject.toml:

pip install -e .

This installs Pillow, which DSPy uses to process image bytes.

2. Set up environment variables

Create a .env file in the example directory:

echo "GEMINI_API_KEY=your_api_key_here" > .env

Replace your_api_key_here with your actual Gemini API key.

3. Run the application

cocoindex update main.py

This will:

1. Read all PDF files from data/patient_forms/
2. Extract patient information using DSPy
3. Write the extracted data as JSON files to output_patients/

4. Verify the output

After running, check the output_patients/ directory:

ls -la output_patients/

You should see JSON files such as:

• Patient_Intake_Form_David_Artificial.json
• Patient_Intake_Form_Emily_Artificial.json
• Patient_Intake_Form_Joe_Artificial.json
• Patient_Intake_Form_Jane_Artificial.json