LLMs, GenAI, RAG, agents, evaluation · insight-1569
Keep "no results" and "the call failed" as different tool outcomes
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- ML engineer, Applied AI
- Industries
- aviation, retail
Summary: situation, action, reason, boundary
Situation
A lookup tool returns an empty list both when no record matches and when the backing service is down.
Action
Return a distinct error status with a retryable flag for failures, reserve the empty result for a successful search that found nothing, and test the agent's behaviour on both cases.
Reason
The model can only reason from what the tool says, so an outage reported as emptiness is read as a fact about the world.
Boundary
Search, lookup and list tools.
Symptom or risk
The agent concludes that the customer has no booking and tells them so, turning an outage into a confident wrong answer.
Exceptions
None for tools whose results drive a decision.
Trade-offs
Wrappers must propagate error details instead of swallowing them.
Pitfalls
Default values and caught exceptions that convert errors into empty data.