I built four GenAI projects before this one: a PDF chatbot, a tool-calling exam-prep agent, a manual ReAct agent built from scratch with LangGraph, and a multi-agent research assistant. All four worked. But "it works" was never something I could actually prove — I just read the output and decided it looked reasonable.
Project 5 was about fixing that. I built a standalone evaluation pipeline using RAGAS to score my RAG system on four metrics: Faithfulness, Answer Relevancy, Context Precision, and Context Recall. Getting it running taught me more about production LLM systems than any of the previous four projects — mostly because of what went wrong along the way.
What I built
A pipeline that: fetches ArXiv paper abstracts on a topic, chunks them, uses an LLM (Groq's llama-3.3-70b-versatile) to auto-generate question/ground-truth-answer pairs from those chunks, builds a FAISS index over the same chunks using sentence-transformers/all-MiniLM-L6-v2, runs each generated question through a simple retrieve-then-answer RAG pipeline, and finally scores the results with RAGAS.
Tech stack: Python, LangChain, FAISS, HuggingFace embeddings, Groq, the arxiv Python library, RAGAS.
