Module 02 Activity

Scenario

Your prompts are f-strings and your configuration is in the code. Both need to change before this is testable.

What you build

Templates with declared inputs, a provider swap proven on an eval set, and configuration read from the environment with versions recorded.

Steps

1. Convert every prompt to a template with declared variables. Prove one fails at render time on a missing input.
2. Render the messages and inspect the roles. Decide deliberately where retrieved context should live and say why.
3. Swap the model object for a second provider without touching the chain, then re-run your eval set and report what changed.
4. Move model, temperature and limits into the environment with a fake model as the default.
5. Add prompt_version and policy_version to your config and to every log line.
6. Write down what temperature=0 does and does not guarantee for your tests.

Evidence to hand in

• The templates with declared inputs and one render-time failure.
• The rendered messages with the role decision explained.
• Eval results before and after the provider swap, per case.
• The environment-driven config with versions.
• Your written position on temperature and test assertions.

Review checklist

• No prompt is built by concatenation.
• The role of retrieved context is a stated decision.
• The provider swap was followed by a full eval re-run, per case.
• The default model is a fake, so misconfiguration fails safely.
• No secret is in any object that gets logged.

Mark complete