DocuMind AI is a secure, high-performance RAG (Retrieval-Augmented Generation) platform that allows users to upload PDFs and engage in contextual chat conversations. Designed with privacy first, it guarantees strict user data isolation for enterprise-grade safety, processing documents with blazing-fast LLaMA 3.3 inference.

Screenshots

Splash Screen	Login & Auth	Home & Library

Chat Interface	Document Citations	Settings

Features

• 🔒 Secure Data Isolation — ChromaDB collections are partitioned per user/document pair, ensuring zero cross-user data leaks.
• ⚡ Lightning-Fast Inference — Powered by LLaMA 3.3 70B on the Groq API for sub-second responses.
• 📈 Zero-Cost Embedded Search — Local Sentence-Transformers generate vector embeddings offline, removing external embedding API costs.
• 🔗 Citation Tracing — Every AI response includes clickable source page references for complete auditability.
• 📱 Cross-Platform Mobile App — A beautiful, responsive Material 3 experience built with Flutter and Riverpod.

Tech Stack

Frontend (Mobile)

• Framework: Flutter 3.41 / Dart 3.x
• State Management: Riverpod (AsyncNotifier) / Immutable State (@freezed)
• Networking: Dio with secure request/response interceptors
• Storage: flutter_secure_storage for session caching

Backend (API)

• Framework: Python / FastAPI
• ORM & Migrations: SQLAlchemy 2.0 (asyncpg) with Alembic
• Validation: Pydantic v2
• Database: SQLite (Development) / PostgreSQL 16 (Production)

AI / ML Infrastructure

• Orchestration: LangChain for managing structured RAG pipelines
• Embeddings: sentence-transformers running locally
• Vector Store: ChromaDB with segregated tenant collections
• LLM Engine: Groq API (LLaMA 3.3 70B)

Architecture Highlights

DocuMind AI is engineered to protect user privacy while maintaining consumer-level latency.

Document Vector Isolation
Instead of pushing all documents into a shared index, the FastAPI server dynamically spins up or mounts isolated ChromaDB collections named after unique user_id and document_id combinations. When a user deletes a file, the entire collection database directory is deleted, satisfying strict GDPR/data sovereignty standards.

Hybrid Local-Cloud RAG Pipeline
To optimize operation costs, vector embedding generation happens entirely on-premise/on-server using lightweight local Sentence-Transformers. Only the clean document chunks and the prompt context are transmitted to the Groq API, leveraging ultra-high-speed LLaMA 3.3 hardware to stream answers back via Server-Sent Events (SSE).