Writing
Newest first. Every post states what was measured, when, and at which commit.
Rerank before you upgrade the model
Most confidently wrong RAG answers are retrieval failures. Log each chunk's rank before and after reranking, measure hit@k, and fix the right stage first.
Failure modesWhat your assistant does when the vector store is down
Every retrieval system has a bad day. The ones that survive it decided in advance what degraded means, and never let a failed lookup sound confident.
Retrieval
Your RAG isn't hallucinating. Your retriever is missing.
Most "hallucination" bugs are retrieval failures wearing a costume. Here is how to tell the two apart before you reach for a bigger model or a longer prompt.
Cost & latency
Write the latency budget before you write the cache
Caching is the first thing teams reach for and the last thing they should. Instrument every stage, give each one a budget, then let the numbers pick the cache.
Evals
Fifty labelled failures beat a thousand synthetic questions
A small eval set built from real failures catches regressions that a generated benchmark never will. Here is how to build one, run it in CI and keep it honest.
Retrieval
Chunking is a retrieval decision, not a preprocessing step
Fixed-size chunking is a default, not a choice. Here is how to pick boundaries from your own failed queries, and the harness that tells you whether it worked.
One essay, when there is something measured to report.
Retrieval, evals, cost and latency — written from systems in front of real users, with the date every figure was measured.
No spam. One click to unsubscribe.