README.md

metadata

title: RAG Observability Platform
emoji: 🚀
colorFrom: blue
colorTo: indigo
sdk: docker
pinned: false
app_port: 7860

RAG Observability Platform 🚀

RAG Observability Platform - Project Summary

Project Overview

The RAG Observability Platform is a production-grade Retrieval-Augmented Generation (RAG) system that demonstrates advanced MLOps practices and hybrid cloud-local deployment strategies. It combines cutting-edge ML inference optimization (Apple Silicon GPU) with MLOps observability frameworks for enterprise-ready applications.

---

What This Project Does

Core Functionality

1. Local RAG Pipeline (Mac M4)

   • Ingests unstructured text documents
   • Chunks documents using recursive text splitting
   • Generates embeddings via sentence-transformers (optimized for Apple Silicon via MPS acceleration)
   • Stores embeddings in ChromaDB (local vector database)
   • Retrieves relevant context and generates answers using Llama 3.2 3B model via MLX

2. Cloud Deployment (Hugging Face Spaces)

   • Docker containerization for reproducible deployment
   • Automatic fallback to CPU-based inference when MLX unavailable
   • Streamlit web UI for interactive chat with documents
   • Graceful degradation: maintains functionality across platforms

3. Experiment Tracking (Dagshub + MLflow)

   • Logs all ingestion runs with parameters and metrics
   • Centralized experiment monitoring from local machine
   • Version control for code and data via Git + DVC
   • Remote MLflow server for team collaboration

Technical Highlights

• Cross-Platform Optimization: Native M4 GPU (via MLX) for local development; CPU fallback for cloud
• Infrastructure as Code: Docker + UV for reproducible environments
• Modern Python Stack: LangChain (LCEL), Pydantic, asyncio-ready
• MLOps Best Practices: Experiment tracking, dependency management, secrets handling

---

Key Highlight

1. GPU Optimization: Understand when to use specialized tools (MLX for Apple Silicon) vs. standard libraries (PyTorch)
2. Cross-Platform Development: Device abstraction, graceful fallbacks, testing on multiple architectures
3. Dependency Management: Using UV for faster resolution, managing optional dependencies (local vs. cloud groups)
4. MLOps Practices: Experiment tracking, versioning data + code, secrets management
5. Production Deployment: Docker best practices, environment variable injection, port mapping
6. Modern Python: Type hints, LangChain LCEL (functional composition), error handling
7. Troubleshooting: Resolved Python version mismatches, binary file handling in Git, device compatibility issues

---

Why This Project Stands Out

• Full Stack: From local GPU optimization to cloud deployment
• Senior-Level Considerations:
  • Device compatibility across platforms
  • Graceful degradation (MLX → Transformers fallback)
  • Secrets management without pushing .env
  • Experiment observability
• Modern Tooling: UV (faster than pip), MLX (Apple Silicon optimization), LangChain LCEL (declarative chains)
• Problem Solving: Resolved real-world issues (ONNX version compatibility, Docker base image mismatch, GPU device detection)

---

GitHub/Portfolio Presentation

Repository Structure (visible in your GitHub):

rag-observability-platform/
├── src/
│   ├── ingestion/      (document loading, chunking, embedding)
│   ├── retrieval/      (RAG chain with LCEL)
│   └── generation/     (MLX wrapper, device handling)
├── app/frontend/       (Streamlit UI)
├── Dockerfile          (Cloud deployment)
├── pyproject.toml      (UV dependency management)
└── README.md           (project documentation)

Git History (visible in commits):

• Clean, semantic commits showing progression
• Branching strategy: master → mvp → frontend/backend
• Demonstrates collaborative workflow understanding

---

1. "Why MLX instead of PyTorch?"

   • MLX is optimized for Apple Silicon; PyTorch CPU mode is 10x slower on M4

2. "How do you handle the MLX import error in Docker?"

   • Try-except with fallback to transformers; dynamic device selection

3. "Why use Dagshub for this portfolio project?"

   • Demonstrates understanding of MLOps practices; shows ability to connect local experiments to remote tracking

4. "What would you do at scale?"

   • Move to managed inference (HF Inference API), DVC for larger datasets, Kubernetes for orchestration

---