Null Input & Output when calling Azure OpenAI model

[haroldpijpelink]

Hi there everyone,

I have this issue where I can run promptflow and get the outputs I want, but tracing does not seem to work properly. Instead of getting a success or fail status, it says the following in the output json:
status: {
status_code: "Unset"
description: ""
}
Moreover, the Input and Output fields show as Null in the UI and I have no information on the amounts of tokens used.

I get this both when I use start_trace() in the tracing docs and when I try to access through the Prompty extension. I realize that this might be some issue with OpenTelemetry settings and not Promptflow itself, but if anyone could point me in the right direction that'd be great :)

[Flowhill]

Have you set

environment_variables:
  PF_DISABLE_TRACING: "false"

In your flow.dag.yaml?

I'm getting the same issue when I'm running it through the python file itself (via fastapi). Then there is no neat trace.
I do get proper traces when I click the run button in the flow.dag.yaml with the promptflow extension.

[Flowhill]

@haroldpijpelink
I was able to solve this for myself.
By reverting to promptflow-tracing==1.17.0 (and thus promptflow itself as well) I did get the proper tracing on flows started with flow-as-a-function.

e.g.

f = load_flow(<path/to/flow/folder>)
        result = f(
            <input-mapping>
        )

I don't use start_trace() as I use the flow-as-a-function.

This seems to break as soon as you update to promptflow(-tracing)>=1.17.1

[aniruddhaadak80]

From my point of view, the suspicious part is that execution works but the observability envelope is losing structured input, output, and status data. That usually means the model call path and the trace or export path are not serializing the same event boundary, especially when Azure wrappers or FastAPI execution are involved.

I would inspect the exact span attributes being emitted around the model invocation and confirm whether the request and response payloads are dropped before they ever reach Promptflow''s UI layer.