Part 1 of 6 · 1 min read
Overview
In this module you will
- Explain how scores become probabilities (softmax) and how temperature, top-k and top-p reshape them.
- Pick sensible starting settings for extraction, for chat and for creative work.
- Explain why "temperature 0" may still not give identical answers twice.
- Describe structured output and why it is more reliable than asking nicely.
An agent that calls tools must produce well-formed requests every time. A loop that works nine times in ten is a loop that fails in production. Sampling settings are the cheapest reliability lever you have, and understanding them lets you stop treating the model as a slot machine.