Improve a poor answer systematically
Find whether the cause is knowledge, instructions, model choice or routing before making broad changes.
Before you begin
Open Analytics or the relevant agent’s analytics/insights tab. Select a useful period and confirm which workspace or agent the report covers.
Step by step
- 1
Save a reproducible example
Record the question, expected answer, agent and time. Use a redacted test version if it contains customer information.
- 2
Inspect the evidence
Check citations and tool results. Determine whether the required fact existed, was current and was retrieved.
- 3
Fix the smallest cause
Correct the source for wrong facts, prompt for behavior, model for capability or workflow for routing. Avoid changing all four at once.
- 4
Run a small regression set
Retest the failing question, a neighbouring case and an unknown question. Confirm the fix did not introduce an overconfident guess.
- 5
Review after rollout
Inspect fresh conversations and refine the test set with recurring real questions.
Good to know
More prompt text is not always better. Conflicting instructions and duplicate policies can make behavior less predictable.
Check the result
Compare the same scope and period before drawing conclusions. Investigate real conversations and sources alongside aggregate metrics.
Questions & troubleshooting
A fluent answer is still wrong
Evaluate facts and justified next steps, not only tone or speed.
Rate this guide on this device. Ratings are not sent to Corthex.