Answers are wrong, and retrieval metrics are poor. What do you do first?
Why B is the answer
Evaluate and fix retrieval
A model cannot use what it never receives, so poor retrieval metrics mean retrieval comes first.
In detail
- Poor retrieval metrics point at retrieval.
- Fix that first.
- Retest after each change.
Use case
An illustrative scenario, not a real incident.
Wrong answers line up with poor retrieval scores, so the team fixes retrieval before it touches the prompt.
See it in different ways
flowchart TD A["GenAI evaluation"] --> B["Retrieval: did it find the right passages?"] A --> C["Model: accuracy and relevance"] A --> D["Agent: task completion"] A --> E["Users: feedback"]
stateDiagram-v2 [*] --> InspectLogs InspectLogs --> Retrieval: wrong context InspectLogs --> Prompt: right context, wrong answer Retrieval --> Evaluate Prompt --> Evaluate Evaluate --> [*]
flowchart TD
Q{"Are retrieval metrics poor?"}
Q -->|"Yes"| A0["Fix retrieval"]
Q -->|"No"| A1["Look at the prompt"]Key point: Fix retrieval first.
Why the other options are wrong
- A. Rewrite the user interface
- The interface is not the cause.
- C. Fine-tune the model
- Fine-tuning does not fix retrieval.
- D. Raise the temperature
- Temperature does not add facts.