Undefined input/output for Spring AI traces when defining userId & sessionId

[marcjaner]

According to the Spring AI integration docs, userId and sessionId should be set up as follows:

public String processUserRequest(String userInput, String userId) {
    // Create a new span with user and session attributes
    Span span = tracer.spanBuilder("user-interaction")
            .setAttribute("langfuse.user.id", userId)
            .setAttribute("langfuse.session.id", UUID.randomUUID().toString())
            .startSpan();
 
    // Make sure to set the new span as active.
    try (Scope ignored = span.makeCurrent()) {
        // Your AI processing logic here
        // Any spans created within this scope will be children of the parent span
        String result = performAiOperation(userInput);
        return result;
    } finally {
        span.end();
    }
}
 
private String performAiOperation(String userInput) {
    // Your Spring AI code here
    // For example:
    List<Message> messages = new ArrayList<>();
    messages.add(new SystemMessage("You are a helpful assistant."));
    messages.add(new UserMessage(userInput));
 
    Prompt prompt = new Prompt(messages);
    ChatResponse response = chatClient.call(prompt);
    return response.getResult().getOutput().getContent();
}

This works and indeed the traces are received with a userId and sessionId as expected:

However, the fact that we're creating a parent trace and then the traces generated by Spring AI are children of the parent one, the input and output fields of the trace appear empty (check image above). In the Observations tab, the child observations contain both the input and output:

Would there be a way to escalate this input and output from the child observation to the parent trace?

I believe this behavior is already happening for traces generated via the AI SDK by Vercel:

with input and output being defined despite having multiple observations in the trace:

Thanks!

[inkeep-ai-assistant[bot]]

You're correct that when creating a parent span for userId and sessionId with Spring AI, the trace-level input and output fields appear empty. This is expected behavior based on how Langfuse handles trace input/output mapping (1).

According to the OpenTelemetry property mapping documentation, trace input and output are set from the root span's observation input and output (1). When you create a custom parent span without setting input/output on it, the trace-level fields remain empty even though child observations contain the actual LLM inputs and outputs.

To populate the trace-level input and output fields, you can explicitly set them using OpenTelemetry attributes on your parent span:

public String processUserRequest(String userInput, String userId) {
    Span span = tracer.spanBuilder("user-interaction")
            .setAttribute("langfuse.user.id", userId)
            .setAttribute("langfuse.session.id", UUID.randomUUID().toString())
            .startSpan();
 
    try (Scope ignored = span.makeCurrent()) {
        // Set trace-level input and output attributes
        span.setAttribute("langfuse.trace.input", userInput);
        
        String result = performAiOperation(userInput);
        
        // Set trace-level output
        span.setAttribute("langfuse.trace.output", result);
        
        return result;
    } finally {
        span.end();
    }
}

(2) (1)

The key attributes for trace-level data are:

• langfuse.trace.input: Sets the overall input for the entire trace (1)
• langfuse.trace.output: Sets the final output for the entire trace (1)

These langfuse.* namespace attributes take precedence over generic OpenTelemetry conventions and are recommended for manual instrumentation (1).

📚 Sources:

1. LLM Observability via OpenTelemetry
2. Integrating Langfuse with Spring AI

---

Have another question? Just tag @inkeep.