This page is the index for the themes introduced on the home page topic map. Each theme has its own detail page with original notes, vocabulary, and pointers into long-form essays. Themes are not silos—measurement informs retrieval design; prompts interact with cost and safety; operations is where they all meet in production.
Choose a theme
- Measurement & risk — contracts, rubrics, drift, and benchmarks that match your product’s risk tier. Pairs with the evaluation experiment and full essay.
- Retrieval systems — chunking, freshness, routing, and retrieval failure modes. Links to RAG experiment hub and RAG long read.
- Prompts as systems — versioning, schemas, regression tests. See prompt experiment and interface design essay.
- Cost & latency — token economics, caching, perceived speed, and how they feed into evaluation priorities.
- Safety & policy — refusals, citations, transparency, and measurable guardrails.
- Operations — logging, replay, shadow traffic, and alerts across the LLM stack.
Structured learning paths
If you prefer a curated order instead of browsing by theme, use the reading paths hub (Foundations, Data-heavy, Release managers) or the latest experiments overview that ties the three flagship essays together.
How themes compound
Measurement without operations produces dashboards nobody acts on; retrieval without safety policy ships plausible wrong answers in regulated domains; prompts without cost awareness blow budgets before evaluation catches regressions. Use this map to gap-check your roadmap: each theme names a class of failure that shows up in postmortems when ignored.
Essays vs. hubs
Theme and experiment pages are shorter orientations with vocabulary and cross-links. Long-form essays (evaluation, RAG, prompts) carry the full argument. Jump between them freely; links are reciprocal.
If you only read one long piece first
Choose by bottleneck: evaluation if releases feel blind; RAG if answers misquote or ignore documents; prompts if behavior shifts whenever someone edits a string. Then return to this map to fill gaps in cost, safety, and operations.