LLM App Patterns

Core Philosophy

1. Retrieval quality is a ceiling on generation quality — No prompt engineering compensates for bad RAG. Fix retrieval before tuning prompts.
2. Workflows beat agents for predictability — Use agents only when the execution path is genuinely unknown at design time. Everything else should be code.
3. Measure before optimizing — Add cost attribution and eval metrics first. Optimization without measurement is guessing.
4. Schema failures cascade — Unstructured LLM output is a reliability tax. Constrain output at the token level; don't parse free text.
5. Stream by default — Token streaming is the lowest-effort UX improvement for any LLM interface. Users read while the model generates.

Quick-Start: Common Problems

"My RAG system gives wrong answers"