Deterministic core
Start with a fake model adapter so request validation, response schemas, timeouts, failures, and redaction can be tested without network calls or usage charges.
API Camp · LLM companion tutorial
API + LLM Tutorial
A guided build brief for placing an LLM behind a small typed API while keeping the default tests deterministic, offline, and free of credentials.
Untrusted output
Treat model responses as untrusted data: validate structure, constrain size, escape rendered content, and keep generated actions behind explicit approval.
Optional live adapter
Add a provider only after the local contract passes. Load credentials from runtime secrets, set cost and timeout limits, and never put keys in source, URLs, logs, or browser storage.
Start here
A practical workflow
- Define typed request and response models
- Create a model-adapter protocol
- Implement a deterministic fake adapter
- Test validation, timeout, retry, and redaction with pytest
- Exercise the browser-to-API path with Playwright
- Optionally connect one live provider behind runtime configuration
Tutorial contents
What is included
- Runnable local Python API
- Deterministic fake-model adapter
- Optional server-side HTTP adapter
- 19 pytest and Playwright checks
- Security and privacy boundaries
- Portfolio exercises and CI workflow
Review the release