This repository contains code for extracting structured data from legal documents using Large Language Models (LLMs). We use Langchain to create prompts and Pydantic to ensure the extracted data conforms to our defined schema.

Based on https://github.com/Zipstack/structured-extraction

This code should run on Linux, macOS. Windows is untested, however work is on to solve this.

You'll need API keys for OpenAI and LLMWhisperer. Once you have the keys, add them to the .env.example file in the root of the project - then rename to .env

project_root/
├── extract.py
├── schemas/
│   ├── definitions/
│   │   └── nda.json
│   └── prompts/
│       └── nda.json
├── .env
├── requirements.txt
└── README.md

1. Clone this repository and navigate to the project directory.
2. Create and activate a Python virtual environment:

   python3 -m venv .venv
   source .venv/bin/activate

3. Install the required dependencies:

   pip install -r requirements.txt

4. Run the script:

   python extract.py <path_to_pdf_or_directory> <schema_file_name>

   For example:

   python extract.py ./example/nda.pdf nda.json

The script processes PDF files containing NDAs, extracts the text content, and uses a language model to parse the information into a structured format. It handles multi-party NDAs, extracting comprehensive information about all aspects of the agreement.

The script now uses customisable schema and prompt configurations, which are loaded from JSON files in the schemas/definitions/ and schemas/prompts/ directories, respectively.

For each processed NDA, the script generates a JSON file containing the extracted information. The structure of this information is defined in the schema file (schemas/definitions/nda.json).

You can customise both the data structure and the prompts used for extraction:

1. To modify the structure of the extracted data, edit the schemas/definitions/nda.json file.
2. To adjust the prompts used by the language model, edit the schemas/prompts/nda.json file.

You can create multiple schema and prompt configurations for different types of documents by adding new JSON files to these directories and specifying the file name when running the script.

• The accuracy of the extraction depends on the quality of the input PDFs and the capabilities of the language model.
• Very complex or non-standard NDAs may not be parsed correctly.
• Some nuanced legal concepts may not be fully captured, and the tool should not be considered a substitute for legal review.

Contributions to improve the script or extend its capabilities are welcome. Please submit a pull request or open an issue to discuss proposed changes.

MIT

This tool is for informational purposes only and should not be considered as legal advice. While it aims to extract a comprehensive set of information from NDAs, it may not capture all nuances or legal implications. Always consult with a qualified legal professional for interpretation, application, and advice regarding NDAs.