405nm
Intelligence
Models, agents and AI engineering.
6 Aug 20263 min read
What a retrieval eval harness actually needs
Retrieval quality is the whole ballgame in a RAG system, and most teams measure it with a vibe check.
11 Jun 20265 min read
Structured output is a schema problem, not a prompt problem
Every time I have failed to get reliable JSON out of a model, the fix was in the schema rather than in the instructions.