📄 CV RAG System - HR Assistant

A RAG (Retrieval-Augmented Generation) system designed to help HR professionals analyze CVs and answer questions about candidates efficiently.

🎯 Project Overview

This project was created to dive deeper into RAG (Retrieval-Augmented Generation) architecture and its practical applications in document analysis.

Initial Idea

The original concept was to build an HR assistant using:

• Streamlit for the user interface
• LangChain for orchestrating the RAG pipeline
• ChromaDB as the vector database
• Ollama for local LLM deployment
• PDF processing to extract and analyze CV content

The system would allow HR professionals to upload CV PDFs and ask natural language questions about candidates.

🛠️ Technical Journey

First Approach (Local Setup)

Streamlit + LangChain + ChromaDB + Ollama (Gemma 2B)

Challenges encountered:

• Hardware limitations with local LLM deployment
• Accuracy issues with Gemma 2B model
• Inconsistent responses to HR questions
• Performance bottlenecks in document processing

Alternative Solution (Cloud-based)

Streamlit + PyPDF2 + Groq + Qwen

Improvements achieved:

• Better accuracy with Groq's infrastructure
• Faster response times
• More reliable Q&A capabilities

Alternatives and Implementation

Ongoing

RAG Architecture Deep Dive

Through this project, I gained hands-on understanding of the RAG stack layers:

Level 0 - Deployment

• Groq: High-performance LLM inference
• AWS/Google Cloud: Scalable infrastructure

Level 1 - Evaluation

• LangSmith: Monitoring and debugging

Level 2 - LLMs

• Qwen: Advanced reasoning capabilities
• Gemma2B

Level 3 - Framework

• LangChain: RAG orchestration
• LlamaIndex: Document indexing

Level 4 - Vector Database

• Chroma: Open-source vector DB

Level 5 - Embedding

• Ollama: Local embedding models
• OpenAI: High-quality embeddings

📚 Key Learnings

RAG Pipeline Understanding

1. Document Ingestion: PDF processing and chunking strategies
2. Embedding Generation: Converting text to vector representations
3. Vector Storage: Efficient similarity search and retrieval
4. Query Processing: Natural language question understanding
5. Context Retrieval: Finding relevant document sections
6. Answer Generation: LLM-based response synthesis

Performance Optimization

• Hardware requirements for local vs cloud LLM deployment
• Model selection impact on accuracy and speed
• Chunking strategies for optimal retrieval
• Prompt engineering for HR-specific queries

Infrastructure Decisions

• Local vs Cloud: Trade-offs between privacy and performance
• Model size: Balance between capability and resource usage
• Vector database: Choosing the right storage solution