Lab 097: Interactive: Tokens, Sampling, and Context Windows¶
Free-first design
Complete the guided exercise with Copilot Free or the included simulation. Paid features are optional and explicitly labeled.
AI Credit impact
Low for guided work; medium/high for agentic or repeated runs. Prefer automatic model selection, plan before acting, keep context focused, and avoid unnecessary parallel sessions.
What You'll Learn¶
- Estimate tokens from text
- Explain temperature, top-k, and top-p
- Budget a context window
- Recognize truncation and context dilution
GH-300 alignment¶
Prompt and context foundations. This mapping follows the GH-300 curriculum snapshot dated 2026-08-07. The official outline may retain older product wording such as “Copilot Edits”; this lab teaches the current product experience and calls out that difference.
Concept Map¶
flowchart LR
I[Instructions] --> W[Context window]
U[User input] --> W
R[Retrieved context] --> W
T[Tool output] --> W
W --> O[Reserved output]
W --> M[Safety margin]
Hands-on Exercise¶
Step 1: Estimate tokens¶
Use the widget below. The estimate uses a transparent approximation, not a model-specific tokenizer. Compare short prose, code, JSON, and accented text.
Step 2: Explore sampling¶
Move temperature, top-k, and top-p. Predict whether output becomes more deterministic or diverse, then reveal the explanation.
Step 3: Budget the window¶
Allocate system instructions, user input, retrieved context, tool output, and response. Keep a safety margin and remove low-value context first.
Supporting kit¶
The site build generates stable PDF and ZIP downloads. The ZIP contains the lab guide and every supporting file declared for this lab.