Automation of Supplementary Agreements Drafting

A solution for automating the process of drafting supplementary agreements with NLP, developed as part of the TenderHack hackathon in Perm, Russia.

This system enables users to consult on legal matters, automatically generate documents, and edit them while ensuring compliance with Russian legal standards.

Key Features

• Legal Consultation: The system analyzes contracts and legislation to provide recommendations on the possibility of drafting a supplementary agreement.
• Request Validation: Before document generation, the system checks whether the request complies with legal norms, ensuring legal accuracy.
• Automated Agreement Generation: The system generates the text of the supplementary agreement based on the user’s request.
• Editing and Version Tracking: Users can manually edit the generated document, tracking changes in a manner similar to Google Docs.
• Text Rephrasing: The built-in AI model helps adapt document text based on user requests while maintaining legal accuracy.

Technologies

Web Interface:

• React + TypeScript: Provides flexibility and scalability for building complex interfaces. Strong typing simplifies code maintenance.
• Modular FSD Structure: Ensures scalability, isolates business logic from the interface, and reduces dependencies between modules.
• Telegram Bot on Aiogram: Acts as a frontend, enabling interaction with the system via chat.

React Web application's code is available here.

Backend:

• FastAPI: Built on Domain-Driven Design (DDD) and the C4 model, ensuring scalability and separation of business logic from technical details.
• RAG (Retrieval-Augmented Generation): Used for consultation services, integrating large language models (LLMs) and document search and ranking mechanisms.

AI Models:

• Gemma2 27B: Acts as a main LLM, creating responses based on context and performing text rephrasing.
• USER-BGE: Generates embeddings that allow efficient indexing and retrieval of contract and legal document information.
• Qwen2 7B: Works with long contexts to extract the essence of given documents.
• cross-encoder-russian-msmarco: Reranks search results to ensure the most relevant documents are passed to the main LLM.

Infrastructure:

• Distributed Architecture: The system operates on three servers with optimized language models through vLLM, minimizing latency and ensuring smooth operation.
• Data Storage: Utilizes S3 storage, MongoDB for document storage and data tracking and Qdrant for storing embeddings and efficiently retrieving relevant documents.
• Docker: Each component of the system is isolated to facilitate updates and ensure stable operation.