If your agent is giving inaccurate, incomplete, or off-topic answers, this guide walks you through the most effective ways to fix it.
Before changing anything, identify what kind of problem you're seeing:
| Symptom | Likely cause |
|---|---|
| Agent says "I don't know" for things it should know | Knowledge source missing or not trained |
| Agent gives outdated information | Source content is stale — needs updating and retraining |
| Agent goes off-topic | System prompt too permissive or missing scope limits |
| Agent gives inconsistent answers | Conflicting content across multiple sources |
| Answers are too long or too short | Temperature or max tokens need adjustment |
| Wrong tone or personality | System prompt needs a clearer persona definition |
Most quality issues trace back to the knowledge base, not the model.
Check coverage. Go to Knowledge Management and look at what sources are trained. If users are asking about something not covered, add a source for it and retrain.
Check training status. A source that failed to train is invisible to the agent. Re-upload it and trigger a new training run.
Remove duplicate or conflicting sources. If you have two documents that say different things about the same topic, consolidate them into one authoritative version.
Add Q&A pairs for stubborn issues. If the agent keeps getting a specific question wrong despite having relevant content, add an explicit Q&A pair in plain text or upload a excel sheet containing the Q&A pairs. Direct Q&A pairs take priority over scraped content.
The system prompt controls how the agent uses its knowledge. Small changes can have a big impact.
Add explicit scope boundaries:
"Only answer questions about [your product]. If a user asks about anything outside this scope, politely tell them that you cannot answer the question."
Specify citation behaviour:
"When answering, reference the relevant section of our help center where applicable."
Control response length:
"Keep answers concise — no more than 3 short paragraphs unless the question requires a step-by-step process."
Improve tone consistency:
"Always use first person plural ('we', 'our'). Never use technical jargon unless the user has used it first."
Temperature is the most impactful setting for response consistency:
Changes to model settings take effect immediately — no retraining needed.
Adding or editing knowledge sources does not automatically update the agent. You must retrain it.
After updating any source:
If you've made several changes, batch them all before retraining to avoid multiple cycles.
Start building intelligent AI agents to engage with customers.
