During our second-year GenAI lab module at Chalapathi Institute of Technology, Guntur, our team was tasked with building an end-to-end Retrieval-Augmented Generation (RAG) assistant for indexing and querying complex technical PDFs and university research papers. Most online YouTube tutorials make RAG look trivial with 5 lines of LangChain code, but when you build a real production pipeline on actual student queries, you run into real architectural bottlenecks.
We started with chunking strategies—naive fixed-size character chunking broke sentences in half, causing hallucinated context. Mentors guided us to implement recursive character splitting with 15% overlap and metadata tagging (page numbers, section headers). For storage and indexing, we benchmarked Pinecone vs self-hosted Qdrant on Docker. We picked Qdrant because of its native hybrid search (combining dense vector embeddings with sparse BM25 keyword matching) and zero API pricing overhead on local machines.
For embeddings, instead of burning OpenAI credits, we ran local sentence-transformers (all-MiniLM-L6-v2) and quantized models on our lab GPUs. The biggest lesson was evaluating retrieval quality: we implemented Ragas metrics to measure context precision and faithfulness. Seeing our model accurately answer obscure questions from 200-page operating systems documentation without hallucinating was the most satisfying moment of the semester. Building this from scratch gave us talking points that blew recruiters away in technical screening calls.
Edited