Notes & Essays

Retrieve top-k facts, refuse below 0.6 without calling the LLM, otherwise answer from a grounded prompt with citations — plus model fallback when backoff on one flash model is not enough.

Embeddings turn text into 3072-number vectors of meaning. Cosine similarity scored two coffee sentences at 0.801 and an unrelated sentence at 0.526 — far from zero.

ChatGPT is an app. The API is for programs: a streaming TypeScript call with @google/genai, a backoff retry for 503 and 429, and an API key that stays out of git.

Seven facts embedded once, ranked by cosine similarity. Mazecare ranked first on a healthcare résumé question without the word health — FHIR and clinical notes were enough. KiriminAja at 0.627 is the warning. Rendang topped out at 0.550.

An LLM is a next-token predictor. Tokens set cost and limits, sampling and temperature change how varied the wording is, and a sourced answer is retrieval before generation.

Learning in public: Hermes Agent on Biznet GIO Cloud, DeepSeek API, Discord gateway live (checkpoint + /sethome) — stack, install path, and what part 2 will harden.