⚡ Athena – Local/Cloud AI Research Assistant

Athena is a lightweight, full-stack AI assistant for querying academic PDFs using LLMs + Retrieval-Augmented Generation (RAG).
Optimized for local, cloud, and GPU-accelerated environments.

---

✨ Features

• PDF Parsing & Retrieval

  • Extracts text using PyMuPDF
  • Chunks + embeds text using MiniLM / BGE
  • Stores embeddings in FAISS for fast top-k retrieval

• Modular RAG Pipeline

  • Switchable models:
    • Mistral-7B (4-bit) → LlamaCpp
    • Phi-mini → Hugging Face
    • Mistral FP16 → Hugging Face
  • Controlled via environment variables

• Flexible Deployment

  • macOS (MPS acceleration)
  • AWS EC2 (GPU/CPU)
  • Azure App Services
  • Dockerized portability

• Performance Gains

  • Response time reduced from 5–10 min ➝ ~35s using quantized LlamaCpp

• User Experience Enhancements

  • Chat memory across sessions
  • Summarization toggle
  • Chunk-source referencing
  • Interactive HTML/CSS/JS frontend

---

🏗 Architecture

flowchart TD
    U[User] -->|Query| FE["Frontend (HTML/CSS/JS)"]
    FE -->|REST API| BE["FastAPI Backend"]
    BE -->|Extract Text| PDF["PyMuPDF Parser"]
    PDF -->|Chunks| EMB["Embeddings (MiniLM/BGE)"]
    EMB --> FAISS[(FAISS Index)]
    BE -->|Retrieve & Rerank| FAISS
    BE -->|Model Switch| LLM["LLMs (LlamaCpp / HuggingFace)"]
    LLM --> BE
    BE -->|Response with Sources| FE
    subgraph Cloud
      S3[(AWS S3 Storage)]
    end
    PDF --> S3

Loading

---

🛠 Tech Stack

Layer	Tools & Frameworks
Core AI	LlamaCpp, Hugging Face Transformers, FAISS
Text Processing	PyMuPDF, MiniLM, BGE
Backend	FastAPI
Frontend	HTML, CSS, JavaScript
Cloud	AWS S3, AWS EC2, Azure App Services
Deployment	Docker, macOS MPS

---

⚙ Setup & Installation

1️⃣ Clone Repository

git clone https://github.com/joze-Lee/Athena.git
cd athena

2️⃣ Create Virtual Environment

python3 -m venv venv
source venv/bin/activate   # Linux/Mac
venv\Scripts\activate      # Windows

3️⃣ Install Dependencies

Copy code
pip install -r requirements.txt

4️⃣ Configure Environment

Create a .env file:

MODEL_BACKEND=llamacpp    # options: llamacpp | huggingface
EMBEDDING_MODEL=MiniLM    # options: MiniLM | BGE
AWS_ACCESS_KEY=your-key
AWS_SECRET_KEY=your-secret
S3_BUCKET=athena-pdfs

▶️ Running Athena

uvicorn app.main:app --reload
Open in browser → http://localhost:8000

Docker

docker build -t athena .
docker run -p 8000:8000 athena

AWS EC2 → GPU/CPU instances + S3 integration

Azure App Service → Deploy FastAPI container

macOS MPS → Enable with LLAMA_CPP_MPS=1

💡 Usage

- Upload a PDF → Stored in S3 (or locally)
- Query → Athena retrieves relevant chunks via FAISS
- LLM → Generates answers with sources
- User → Can enable summarization, switch models, view references

📊 Benchmarks

Model	Backend	Precision	Avg Response Time
Mistral-7B (4-bit)	LlamaCpp	High	~35s
Mistral FP16	Hugging Face	Higher	~1.5–2 min
Phi-mini	Hugging Face	Moderate	~50s

🧩 Roadmap

 LangChain orchestration
 Multi-modal support (PDF + images)
 Domain-specific embedding fine-tuning
 WebSocket streaming responses

📜 License

 MIT License – see LICENSE for details.

🤝 Contributing

 Fork the repo
 Create a feature branch   
 Submit a PR 🚀