A privacy-first, fully local Agentic AI legal assistant. Combines multimodal RAG (DeepSeek-R1 + Llama 3.2 Vision) via Telegram for zero-data-leakage contract analysis.

⚖️ Contract Buddy

An asynchronous, multimodal, and fully localized AI legal assistant deployed via Telegram. Designed with privacy and performance in mind, this project demonstrates advanced Agentic AI architecture, combining Retrieval-Augmented Generation (RAG) with local multimodal vision models to analyze, query, and summarize legal contracts.

🧠 Architecture Overview

Built to run entirely locally, ensuring zero data leakage for sensitive legal documents. The system dynamically manages VRAM by swapping between reasoning and translation models on the fly.

🚀 Core Features (Updated for Milestone 1)

• The Universal Translator Pipeline: Natively processes legal documents in English and 6 Indic Languages (Hindi, Marathi, Gujarati, Kannada, Tamil, Telugu).
• Smart OCR Extraction: Bypasses standard vision models in favor of EasyOCR, successfully extracting raw Devanagari and Latin scripts from physical document images and handwritten notes.
• Dynamic AI Routing: Uses a "Translate-Reason-Translate" workflow:
  • langdetect identifies the document language.
  • Sarvam-1 (an Indic-optimized LLM) translates regional text into English to prevent hallucination.
  • DeepSeek-R1 acts as the core legal brain, running risk analysis on the normalized English text.
  • Sarvam-1 translates the final legal analysis back into the user's selected UI language.
• Local RAG Pipeline: Utilizes sqlite-vec for ultra-fast, lightweight vector storage and sentence-transformers for embedding contract chunks.
• Stateful Memory: Implements a custom chat_id-keyed conversational memory buffer, enabling contextual follow-up questions and executive session summaries.
• Asynchronous Routing: Built with python-telegram-bot and asyncio.to_thread to prevent API timeouts during heavy local GPU inference.
• Output Sanitization: Engineered regex layers to cleanly parse and remove internal <think> tags from reasoning models.

🛠️ Tech S