Best Practice of using OpenInference Instrumentation?

[ruqianq]

Hi I am active user of OpenInference. Currently we have multiple projects are using Openinference to trace our AI agents and export trace to Phoenix for eval. Often time the traces are quite overwhelmed, is there a user guidance for best practice of using openinference and sending trace to phoenix? Thanks!

[axiomofjoy]

Hey @ruqianq, can you elaborate on what you mean by "overwhelmed"? Do you mean by the volume of traces?

[ruqianq]

Thank you for your response, and apologies for the vague question earlier.

Yes, the volume is quite large—especially when we’re running multiple agents and tools. We’re glad to see that all the spans are being logged in Phoenix, but at the same time, we’re feeling pretty lost and unsure how to educate our users to use the tool better.

If you have any recommendations or best practices for navigating or interpreting the logs more effectively, that would be incredibly helpful.

[axiomofjoy]

Thanks @ruqianq, I'll ask our dev rel team to lean in here.

[Jgilhuly]

Hi @ruqianq - great question here, the traces view can get pretty extensive.

While the auto-instrumentors are our recommended approach, it can be helpful to incorporate some amount of manual instrumentation to better group traces.

For multi-agent systems, I'd recommend grouping each agent invocation under a top level agent trace with the agent name as the span name. This lets you easily distinguish agent executions in traces, and lets you filter for specific agents in the main traces view.

Lastly, I highly recommend incorporating Sessions if they are applicable for your use case. These give you a higher level way to read and follow multi-turn conversations with agents.

If you can share which instrumentors you're using, we can also give more specific recommendations. Or feel free to grab time with our team to talk more specifically!