I Deleted Pinecone, Redis, and 400 Lines of Python. My RAG Pipeline Still Works.

Last November I had a Pinecone bill for $70. For a side project. That nobody was using yet. I sat there looking at my architecture and counted: Pinecone for vectors, Redis for caching, a FastAPI service hitting OpenAI's embedding endpoint, LangChain stitching it all together with 400-something lines of Python, and PostgreSQL — which had my actual data the whole time, just sitting there doing nothing interesting. Five moving parts. I couldn't even run the thing locally without spinning up four Docker containers and praying they'd all connect. When Pinecone had that 20-minute outage in October, my "AI-powered" app just... died. The documents were in Postgres. The user was talking to Postgres. But the retrieval step went through San Francisco and back, so tough luck. I spent a weekend ripping it all out. Here's what replaced it. Before and after Before: User query → FastAPI → OpenAI (embed) → Pinecone (search) → Redis (cache check) → PostgreSQL (get docs) → OpenAI (generate) → Response Af