Learnings — Index
What I actually did — one file per learning, each in the same format: what was encountered, how it was solved, and how it relates to the Reference book. A learning often starts as a negative (something too slow, too vague, broken) and ends with the fix and the theory behind it.
Every note links up to the concept it relates to; each concept links back down via a Related implementations line. Browse by topic with the tag pages, or scan the full list below (newest first).
All learnings
| Date | Learning | Tags | Reference |
|---|---|---|---|
| 2026-07-26 | Adding usage telemetry to the LLM adapter broke every call site silently | cost, infrastructure | The provider adapter layer |
| 2026-07-26 | One gateway URL, two credentials — chat and embeddings needed separate API keys | infrastructure, security | The provider adapter layer |
| 2026-07-25 | Consolidating the vector store into Postgres removed a second datastore | vector-db, rag | Vector-DB tradeoffs |
| 2026-07-25 | A dropped retry made V2 fail silently on truncated output | reliability | Graceful degradation |
| 2026-07-25 | Routing read/classify steps to a cheap model cut cost | cost, model-routing | Model routing |
| 2026-07-25 | Validating every phase into a schema made LLM output safe to build on | structured-outputs | Structured outputs |
New learnings are filed per ASSISTANT_PROMPT.md: send me the day's rough notes and I clean each one into this format, link it to the reference, and add a row here. More than one learning in a day just means more than one file — that's expected.