Debug guide
How to improve your conversation flow agent’s performance
Conversation flow is a powerful and flexible tool, which means that there’s a lot of action items one can take when the agent’s performance is not meeting your expectation. This guide is designed to help you identify the root cause of the issue, and provide actionable steps to improve the agent’s responses and transitions.
Step 1: Idenfify the issue
When agent is not responding as expected, there can be a few reasons:
- agent is not following instructions within a node
- agent transition is not working as expected
- actual conversation does not match the flow graph of agent (e.g. user does not follow your steps)
Step 2: Fix the issue
Please note that these issues are not mutually exclusive, and you can try multiple solutions to fix the issue.
Issue: Agent is not following instructions within a node
- If the instructions are long / contain logical decisions, you can try to break down the node into multiple nodes.
- If the instructions are consise but the agent cannot comprehend it, you can try using a more capable LLM model for this node.
- If you want some specific style on the response, you can add conversation finetune examples. Read more at Finetune Examples.
Issue: Agent transition is not working as expected
If you expect the agent to transition to a specific node, but it doesn’t, you can try to:
- check your transition conditions: make sure it matches all the conditions you expect it to trigger. Feel free to prompt engineer it or break it down to multiple conditions.
- add transition finetune examples: you can give model some examples so that it can learn your expectation. Read more at Finetune Examples.
If you are missing out on some scenarios that caused the transition to fail, you can try to:
- add more nodes to cover the scenarios, sometimes adding a few global nodes can help take care of corner cases of the call. Read more at Global Node.
- make your transition condition more general.
Issue: Actual conversation does not match the flow graph of agent
When user does not follow the strict flow defined, you can try to:
- add & mark a few key steps as global nodes, so that user can skip / jump between nodes. This is especially useful in inbound support use cases where there’s no strict call structure. Read more at Global Node.
- make your node instructions more general and let the model fill in the details.
Was this page helpful?