🤖 Kubeflow Documentation Assistant (GSoC 2025 PoC)

🧪 Proof of Concept for [Kubeflow GSoC Project 12 – “Empowering Kubeflow Documentation with LLMs”]

🔗 Live Demo: kubeflow.kacem-mathlouthi.tn

---

📌 Project Overview

This is a full-stack Retrieval-Augmented Generation (RAG) chatbot designed as a PoC for GSoC 2024 (Project 12) with Kubeflow. It combines GitHub-based documentation ingestion, vector search, and LLM-powered responses into an interactive real-time assistant.

---

🧠 How It Works

graph TD;
  A[User Question] --> B[Similarity Search];
  B --> C[Prompt + Docs];
  C --> D[LLM via GROQ API];
  D --> E[Chatbot Response];

Loading

1. 🧑 User submits a question
2. 🔍 Similar documents are retrieved using vector search
3. 🧾 Retrieved chunks are passed as context to an LLM
4. 🤖 Final response is generated and streamed back to the frontend

---

🧰 Tech Stack

🔙 Backend (FastAPI)

• FastAPI + WebSockets
• Azure OpenAI & GROQ for LLM & embeddings
• Weaviate (local or remote) for vector storage
• LangChain for text chunking & prompt formatting
• Docker & Poetry

🌐 Frontend (React)

• React + TypeScript + Vite
• TailwindCSS + ShadCN UI for design
• WebSocket-based real-time chat
• Docker + NGINX for serving
• Markdown rendering with syntax highlighting

---

📦 Repo Structure

kacemmathlouthi-kubeflow-demo/
├── backend/          # FastAPI, RAG logic, Weaviate, Gitingest integration
├── frontend/         # Chat UI, WebSocket client, Tailwind
├── docker-compose.yml
└── README.md         # ← You're here

---

🚀 Getting Started

🐳 Run with Docker Compose (recommended)

# From the root directory
docker-compose up --build

• Frontend available at: http://localhost
• Backend API at: http://localhost:8000

💻 Run locally

Backend

cd backend
poetry install
poetry run uvicorn app.main:app --reload

Frontend

cd frontend
npm install
npm run dev

---

🌍 Live Demo

The application is deployed via NGINX on Azure, and available here:

👉 kubeflow.kacem-mathlouthi.tn

---

✅ Implemented Features

• 🔁 Full RAG pipeline with LLM response generation
• 🔗 GitHub repo ingestion via Gitingest
• 🧠 Embedding + vector search with Weaviate
• 🌐 Real-time WebSocket chat
• ⚙️ Configurable LLM model, temperature, max tokens
• 🧾 Markdown support in replies
• 🚀 Dockerized & deployed to Azure

More info in frontend/src/components/implemented-features.tsx

---

🧪 Planned Features

• 🧠 Multi-turn memory and session tracking
• 📚 Ingest all Kubeflow repos, issues, articles, StackOverflow, etc.
• 🧬 RAG pipeline benchmarking & rerankers
• 🧠 Fine-tuning with Kubeflow Trainer
• ⚙️ Modular pipeline with Kubeflow Pipelines
• 📊 User feedback & prompt engineering

See frontend/src/components/upcoming-features.tsx

---