Measured,
not guessed.

Notes on RAG, retrieval, and LLM evaluation — written the way I actually work, failures included.

        The log
        2 entries
      

Latest · Vector Search

Jul 2026·8 min read

Building a vector store on MySQL 9, and everything I refused to fake

MySQL 9 gives you a VECTOR column and a distance function. No index, no relevance scores, nothing else. Here's what I learned building a real LangChain and LlamaIndex vector store on top of that, without pretending it had features it doesn't.

Vector SearchLangChainOpen Source

Read the note →

Jul 20266 min read

RAGLLM Evaluation

I graded 50 RAG answers with a hand-built harness. Half the metrics were lying to me.

Faithfulness sat at 1.0 across every config I tried. That number told me nothing, and here's the one I actually trust instead.

Read the note →

Newsletter

New notes, straight to your inbox.

One email when I publish something new on RAG, retrieval, and evaluation.