How I Built a RAG Pipeline Without a Framework

Most RAG tutorials hand you LangChain and call it a day. I wanted to understand what actually happens when a user asks a question — embedding, retrieval, reranking, and generation — without hiding behind abstractions.

Building from scratch forced me to think about chunking strategy, metadata filtering, and when vector search alone isn't enough. The biggest lesson: your retrieval quality matters more than your LLM choice.

I'll walk through the architecture decisions, the bugs that only showed up in production, and what I'd do differently on the next iteration.