Prompt, retrieval, and tool-use experiments where the hard part is incomplete business context, not model cleverness.
Why this note exists.
Real operational inputs are incomplete, inconsistent, and full of unstated assumptions. Model tests are useful only when they include that mess instead of polished demo examples.
Observations
- Missing context often causes more failure than model weakness.
- Retrieval needs boundaries and source quality checks.
- Tool-use tests should include unclear, partial, and contradictory examples.
What we learned
- Build evaluation cases from real workflow examples.
- Test failure behavior and escalation, not only successful answers.
- Use messy inputs to define what the workflow must capture earlier.
Useful artifact
- Workflow map or decision note.
- Test case, approval log, or delivery handoff.
- Reusable pattern for a future service path.