Multi-User PDF Q&A

A Streamlit-based web app for uploading and querying multiple PDF research papers concurrently, using FAISS for vector search, Redis for storage, Gemini 1.5 Flash API for answers, and Kubernetes/Docker for deployment.

Features

• Upload PDFs and ask questions about their content.
• Supports multiple users with isolated sessions via unique user_id.
• FAISS enables fast similarity search for context retrieval.
• Redis stores text chunks, embeddings, and chat history.
• Gemini 1.5 Flash API generates concise answers.
• Deployed on Kubernetes with Docker for scalability.

Prerequisites

• Docker Desktop: Installed and running on Windows.
• Kubernetes: Enabled in Docker Desktop.
• PowerShell: For running commands.
• Gemini API Key: Obtain from Google AI Studio.
• Free disk space: ~25GB for Docker images and builds.

Setup Instructions

1. Set up the Gemini API key in k8s/deployment.yaml and app/gemini_api.py:

   - name: GEMINI_API_KEY
     value: "<your-real-key>"

2. Ensure Docker Desktop and Kubernetes are running.

Project Structure

multi_user_pdf_qa/
├── app/
│   ├── Dockerfile
│   ├── requirements.txt
│   ├── main.py
│   ├── utils/
│   │   ├── embeddings.py
│   │   ├── pdf_processor.py
│   │   ├── redis_client.py
│   │   ├── gemini_api.py
├── k8s/
│   ├── deployment.yaml
│   ├── service.yaml
│   ├── redis/
│   │   ├── deployment.yaml
│   │   ├── service.yaml
├── screenshots/
│   ├── streamlit_paper1.png
│   ├── streamlit_paper2.png
│   ├── streamlit_paper3.png
│   ├── streamlit_paper4.png
│   ├── streamlit_paper5.png
│   ├── docker_build.png
│   ├── kubectl_commands.png
│   ├── docker_containers.png
│   ├── docker_images.png
│   ├── docker_volumes.png
│   ├── docker_builds.png
│   ├── kubectl_docker_status.png
├── readme.markdown

How It Works

1. PDF Upload: Users upload PDFs via Streamlit, saved temporarily.
2. Text Extraction: PyPDF2 extracts text, split into chunks.
3. Embedding: sentence-transformers (all-MiniLM-L6-v2) generates embeddings.
4. Storage: Chunks, embeddings, and FAISS index stored in Redis.
5. Querying: User questions are embedded, matched via FAISS, and context sent to Gemini 1.5 Flash.
6. Response: Gemini generates concise answers, stored in Redis chat history.
7. Multi-User: Kubernetes/Docker ensures isolated sessions per user.

Running the Project

1. Build the Docker image:

   cd app
   docker build -t multi-user-pdf-qa:latest .

   See screenshots/docker_build.png.

2. Deploy to Kubernetes:

   cd ../k8s
   kubectl apply -f redis/deployment.yaml
   kubectl apply -f redis/service.yaml
   kubectl apply -f deployment.yaml
   kubectl apply -f service.yaml

   See screenshots/kubectl_commands.png.

3. Forward the Streamlit service port:

   kubectl port-forward svc/streamlit-service 8501:8501

   See screenshots/kubectl_commands.png.

4. Access the app at http://localhost:8501.

   • Upload a PDF (e.g., Generative Agents Interactive Simulacra of Human Behavior.pdf).
   • Ask: “What is this paper about in one line?”
   • See screenshots/streamlit_paper1.png to streamlit_paper5.png for 5 papers.

5. Verify system status:

   kubectl get pods
   kubectl get svc
   docker ps

   See screenshots/kubectl_docker_status.png.

6. Check Docker Desktop:

   • Containers: screenshots/docker_containers.png
   • Images: screenshots/docker_images.png
   • Volumes: screenshots/docker_volumes.png
   • Builds: screenshots/docker_builds.png

Dependencies

• streamlit==1.45.1: Web app interface.
• redis==3.2.0: Stores chunks, embeddings, and history.
• faiss-cpu==1.7.4: Vector similarity search.
• PyPDF2==3.0.1: PDF text extraction.
• sentence-transformers==2.2.2: Text embeddings.
• google-generativeai==0.8.5: Gemini 1.5 Flash API.
• torch==2.0.1: ML framework.
• transformers==4.28.1: Hugging Face models.
• huggingface_hub==0.16.4: Model access.
• numpy==1.23.1: Numerical computations.

Demo Results

Tested with 5 research papers. Screenshots are available in the screenshots/ folder:

Streamlit UI

• 
• 
• 
• 
• 

Docker & Kubernetes

• 
• 
• 
• 
• 
• 
•