Context-AI 🚀

Context-AI is a production-grade, full-stack document analysis platform. Built with a monolithic Next.js and FastAPI architecture, it enables users to seamlessly upload documents, process them into AI-readable vector embeddings, and frictionlessly query that knowledge base in real-time.

🛠 Tech Stack Overview

• Frontend: Next.js 14 (App Router), React, TailwindCSS, Framer Motion, Lucide Icons
• Backend: FastAPI, SQLAlchemy (SQLite), Python 3.12, Uvicorn
• AI / Modeling: Mistral AI (Large Language Model), HuggingFace Sentence Transformers (Embeddings)
• Vector Database: ChromaDB
• Authentication: JWT Bearer Tokens, Bcrypt Hashing, Custom built 2-Step SMTP Email OTP Flow

---

🏗 Architecture & Flow

1. Security & Authentication

Context-AI uses a highly secure modern authentication flow:

• Upon signup, an account is flagged as unverified.
• A raw Python SMTP Client interfaces with a Gmail App Password to dispatch a custom HTML-styled 6-digit OTP mapping.
• The React application seamlessly transforms the UI to catch the OTP, hitting /api/v1/auth/verify-otp.
• We use strict middleware to emit HTTP 403 Forbidden errors if a verified user's lifecycle token doesn't map to a logged-in state.
• Passwords are salted and hashed using Passlib Bcrypt before striking the SQLite database.

2. The Vector Pipeline (RAG)

Retrieval-Augmented Generation (RAG) is the core engine of Context-AI:

1. Upload Phase: Users upload heavily formatted PDFs through the Next.js FormData engine. FastAPI catches the document into temporary storage.
2. Parsing & Chunking: PyPDFLoader strips the text. A Recursive Character Text Splitter breaks it into tiny overlapping semantic chunks (to prevent context loss between pages).
3. Embedding Generaton: We call sentence-transformers/all-MiniLM-L6-v2 down from HuggingFace to convert English text chunks into high-dimensional numerical vectors.
4. Vector Storage: The resulting vectors are natively cached into the persistent ChromaDB envelope located strictly inside the isolated backend directory.

3. The Query Engine

When users type a question into the beautiful dark-mode chat interface:

• FastAPI embeds their raw text question into a temporary vector.
• We run a mathematical distance search inside ChromaDB using Maximum Marginal Relevance (MMR) algorithms to fetch the most semantically relevant PDF chunks.
• We prepend these chunks into a dynamic Prompt Template and ship it exclusively to the Mistral AI language model.
• Mistral formulates an incredibly accurate response strictly constrained by the prompt boundary: "Use ONLY the provided context...".

---

🏃‍♂️ How to Run

Because this is a consolidated structure, startup is extremely seamless.

1. Create a root .env file containing your MISTRAL_API_KEY and SMTP configs.
2. Launch the orchestrator script:

# This automatically spins up Uvicorn (FastAPI) on port :8000 and Next.js on port :3000
./start_app.sh