🧠 macRAG: Local RAG Stack on Mac

macRAG is a modular, secure, and locally-deployable Retrieval-Augmented Generation (RAG) framework that runs on your Mac with full Dockerized support. It allows developers to:

• Load & chunk documents
• Embed & index them into vector databases
• Retrieve relevant content for LLM-based Q&A
• Visualize and evaluate model performance with integrated benchmarking tools

---

📐 Architecture Overview

---

📂 Project Structure

• src/ - Core modules: loaders, chunkers, retriever, embeddings, reranker
• data/ - Sample datasets including hotpot_sample.json
• streamlit_app/ - Interactive UI powered by Streamlit
• notebooks/ - Embedding visualization and metric evaluation notebooks
• eval/ - Evaluation scripts and charts
• tests/ - Unit tests for each major component
• .env.example - Environment variable template
• docker-compose.yml - Local deployment using Docker
• setup.sh - End-to-end setup script

---

🚀 Getting Started

bash setup.sh

Then visit the UI at: http://localhost:8501

---

🔐 Security Considerations

This is a local demo stack for experimentation. In production, add:

• Authentication (e.g., Streamlit Authenticator)
• HTTPS/TLS with a reverse proxy
• Rate-limiting, API auditing, input sanitization
• Data encryption for local storage

---

📊 Evaluation Support

• Supports BLEU, Precision, Recall
• Metric charts via Matplotlib
• Evaluation Notebook included under notebooks/

---

🛠️ Requirements

• Python 3.9+
• Docker
• Ollama CLI or similar for local LLMs

---

🙌 Acknowledgments

macRAG leverages open-source technologies including:

• LangChain
• FAISS / Chroma
• Streamlit
• Ollama

Inspired by practical needs for local GenAI prototyping.

🧠 Secure RAG Architecture on Vertex AI

This diagram illustrates a secure retrieval-augmented generation (RAG) pipeline leveraging Vertex AI, with integrated components like:

• Security Layer: Authentication, encryption, and audit logging
• Vector Database: Storing semantically indexed embeddings
• Generative Model: Vertex AI + Custom Foundation Models
• Prompt Flow: Query understanding, context injection, and synthesis
• Enterprise Data: Structured and unstructured sources under access control

---

🧱 Platform-Based RAG Approach

This project is structured to enable modular replacement and experimentation:

Layer	Component	Replaceable With
Chunking	src/chunkers.py	Markdown/semantic chunking
Embeddings	src/embeddings.py	ELSER / OpenAI / Cohere
Vector Store	FAISS / Chroma	Weaviate / Pinecone / Qdrant
Reranker (opt)	src/rerankers.py	Relevance classifier / BM25
LLM Backend	Ollama + LangChain	Vertex AI / OpenAI / Claude
App Interface	Streamlit	Gradio / FastAPI
Evaluation	notebooks/metric_eval.ipynb	Custom BLEU / ROUGE pipelines

---