When RAG Goes Wrong: Notes from Building a Naive System

Tue, 02 Jun 2026 00:00:00 +0000

I built a basic RAG system over Anthropic’s documentation last week to develop hands-on intuition about where these systems actually break. The system was deliberately simple — fixed-size character chunks, OpenAI’s text-embedding-3-large, Pinecone for storage, top-5 retrieval. No reranking, no hybrid search, no clever chunking. The point wasn’t to build something good; it was to see what fails when you don’t.

Here’s what I found.

Failure 1: Vocabulary similarity is not relevance

I asked the system “What’s the difference between system prompts and developer messages?” Developer messages aren’t standard Anthropic terminology, but the system returned chunks from the system prompt documentation anyway — because they shared vocabulary with the query, not because they addressed it.

AI on Oliver Hudson

When RAG Goes Wrong: Notes from Building a Naive System

Failure 1: Vocabulary similarity is not relevance