Microsoft AI Foundry tracing - prompts (input & output)

Question

Microsoft AI Foundry tracing - prompts (input & output)

Patricia Torok (SK) 0

Hi, I'm testing Microsoft Foundry. I want to use the tracing feature and observe my application, which utilizes LangGraph v1.

I tested the tracing functionality using the following documentation: https://learn.microsoft.com/en-us/azure/ai-foundry/how-to/develop/trace-agents-sdk?view=foundry-classic#integrations . However, the prompts seem to be off. There are blank prompt fields that are not populated.

Question 1: Is this normal behavior?

Question 2: What are those Http calls?

I'm using Python 3.12 with the following libraries:

langchain-azure-ai==1.0.3

langchain-core==1.1.0

langchain-openai==1.1.0

langgraph==1.0.4
langchain-openai==1.1.0
azure-identity==1.25.1

1 answer

Your answer

Answer 1

SRILAKSHMI C 10,805 Microsoft External Staff Moderator

Hello Patricia Torok (SK),

Welcome to Microsoft Q&A and Thank you for reaching out.

I understand that you're encountering some issues with the tracing feature in Microsoft Foundry while using LangGraph v1.

Let’s break down your questions:

Question 1: Is this normal behavior?

Blank prompt fields not being populated can be a sign of a few things. If the tracing feature is set up correctly, you should typically see prompts being populated. Here are a few things you might want to check:

Ensure that you've linked your Application Insights correctly in the project’s Monitor settings.
Verify that the credentials and environment variables (like APPLICATION_INSIGHTS_CONNECTION_STRING and AZURE_OPENAI_ENDPOINT) are set up properly.
Check if you are using both server-side and client-side tracing. If you're only using server-side tracing, it may not capture all the necessary details, leading to incomplete spans.

Question 2: What are those HTTP calls?

The HTTP calls in the tracing context usually refer to the interactions between your application, the Azure AI services, and the tracing infrastructure (like Application Insights). For details on those calls, you can:

Enable logging for your Azure OpenAI API Management to view the requests being sent and received.
Use the tracing capabilities to monitor the spans and identify the specific HTTP calls that are being made during your application’s processing.

Suggested Steps to Troubleshoot:

Review your Tracing Setup: Go through the setup steps in the relevant documentation and make sure everything is configured correctly.
Check Logs: Look at the logs in Application Insights to see if there are any errors or missed data that might explain the blank prompts.
Testing Configuration: Run some test queries to see if the tracing behavior changes and provides the expected prompts.
Documentation Reference: Use the tracing integration guide to make sure you have implemented everything as expected.

Please refer this,

I Hope this helps. Do let me know if you have any further queries.

If this answers your query, please do click Accept Answer and Yes for was this answer helpful.

Thank you!

Patricia Torok (SK) 0 Reputation points

2025-12-03T10:47:10.6433333+00:00
Hi @SRILAKSHMI C , thank you for the response.

In addition to your answer for Question 1, I want to add a few more details:

I have linked my Application Insights correctly to the project's Monitoring settings.

The environment variables are set up correctly.

The logs seem to be correct, and the code does what it should.

The metadata of the trace is as follows (see below). I believe that the assistant and tool calls don't contain the 'content' part, which is why the fields in the 'Input field' view are blank. I did everything based on the Microsoft guide attached in my main thread.

Question 1: Is my understanding correct?

Question 2: Is this the expected behavior?

Thank you!
SRILAKSHMI C 10,805 Reputation points Microsoft External Staff Moderator

2025-12-05T15:02:34.53+00:00
Hi Patricia Torok (SK),,

Thank you for the additional details,

Based on the trace metadata you shared, your understanding is correct, and the behavior you are seeing is expected with the current version of Microsoft Foundry + LangGraph integration.

Answer to Question 1 — Yes, your understanding is correct

The reason the Input and Output prompt fields appear blank in the Foundry tracing UI is because:

LangGraph + LangChain’s tool-calling messages do not place user/assistant “content” in the same field Foundry expects for display.

In the trace metadata you shared:

The assistant step contains

{ "type": "tool_call", "name": "...", "arguments": {...} }

→ This message has no content field, so Foundry has nothing to display in the prompt box.

The tool step contains

{ "type": "tool_call_response", "result": "..." }

→ Again, no content field.

LangGraph “frames” the conversation as structured function-call messages, not text messages.

The tracing UI currently shows only:

role: user → content (works)

role: assistant → content (works only when assistant returns text)

It does not interpret function-call style messages as text prompts

So you are correct: The UI is blank because those messages do not contain content in the expected field.

Answer to Question 2 — Yes, this is expected behavior today

This is expected and known behavior in the current Foundry trace viewer:

The Foundry UI primarily visualizes text prompts.

LangGraph’s structured “tool_call” and “tool_call_response” messages do not contain text content, only function-call metadata.

The UI does not yet “hydrate” these messages into a readable, text-based prompt.

At the moment,

Tool call messages appear in the metadata but not in the prompt view.

Only raw text messages render in the “Input” and “Output” fields.

So yes, what you see is currently expected and not a misconfiguration on your side.

What those HTTP calls are

The HTTP spans you see are automatically generated traces from the Azure Identity library when LangGraph acquires tokens:

GET /metadata/identity/oauth2/token

Managed identity → Azure OpenAI

These are normal authentication calls between:

Your app

Azure Instance Metadata Service (IMDS)

Azure OpenAI service

These calls are not part of your prompt data they are infrastructure-level spans captured automatically by the tracer.

Recommended Steps,

To view clean, text-based prompts in tracing:

Return text alongside tool calls, for example:

return { "messages": [ {"role": "assistant", "content": "Calling tool…"}, {"role": "assistant", "tool_call": {...}} ] }

Or enable client-side tracing (langchain-azure-ai >= 1.0.5) which captures more granular spans.

Could you please take a moment to retake the survey on the above response? Your feedback is greatly appreciated.

Thank you!

Share via

Microsoft AI Foundry tracing - prompts (input & output)

1 answer

Your answer