Is it possible to debug vercel AI SDK retry for structured output in langfuse?

[leo91000]

Describe your question

I'm currently using Langfuse with the Vercel AI SDK. It works well, but I want to debug structured output retries from the Vercel AI SDK's generateObject.

I know the AI SDK is retrying because it never succeeds on the first attempt. However, I'd like the model to succeed on the first try if possible (for latency reasons).

In Langfuse, it would be really helpful if I could see:

1. Raw provider requests being made
2. Retry attempts from the Vercel AI SDK

Is this something that's already possible?

Langfuse Cloud or Self-Hosted?

Langfuse Cloud

If Self-Hosted

No response

If Langfuse Cloud

cmipvwlwx004gad07z7dpdsij

SDK and integration versions

Langfuse packages

@langfuse/client: 4.4.9
@langfuse/otel: 4.4.9
@langfuse/tracing: 4.4.9

Vercel AI SDK

ai: 6.0.39
@ai-sdk/anthropic: ^3.0.15
@ai-sdk/google: ^3.0.10
@ai-sdk/openai: ^3.0.12
@ai-sdk/amazon-bedrock: ^3.0.73
@ai-sdk/mistral: ^3.0.9
@ai-sdk/xai: ^3.0.26
@ai-sdk/deepseek: ^1.0.0
@ai-sdk/provider: ^3.0.4
@ai-sdk/devtools: ^0.0.6 (devDependency)

OpenTelemetry

@opentelemetry/sdk-node: 0.208.0

Pre-Submission Checklist

• I have checked for existing issues/discussions and consulted Langfuse AI.

[inkeep-ai-assistant[bot]]

Based on the available documentation, Langfuse can capture detailed observability data from the Vercel AI SDK through OpenTelemetry integration(1). Since you're using @langfuse/otel: 4.4.9 and @opentelemetry/sdk-node: 0.208.0, you're already set up for this integration.

To see raw provider requests and retry attempts from the Vercel AI SDK's generateObject, you should be able to capture this data at the observation level in Langfuse. The OpenTelemetry integration maps various attributes to Langfuse observations(1):

For input/output visibility, Langfuse maps:

• langfuse.observation.input
• langfuse.observation.output
• gen_ai.prompt
• gen_ai.completion(1)

For model parameters and usage, it captures:

• langfuse.observation.model.parameters (as JSON string)
• gen_ai.request.*
• langfuse.observation.usage_details (as JSON string)
• gen_ai.usage.*(1)

Each retry attempt should create separate observations (spans) within your trace, allowing you to see:

1. Individual attempts with their inputs/outputs
2. Token usage per attempt
3. Timing information for each retry

If you're not seeing this level of detail currently, you may need to ensure that:

• The Vercel AI SDK's OpenTelemetry instrumentation is properly configured
• All spans from retry attempts are being exported to Langfuse

The documentation doesn't specifically mention Vercel AI SDK retry behavior, but the general OpenTelemetry integration should capture all spans emitted by any OTEL-instrumented library(1).

📚 Sources:

1. LLM Observability via OpenTelemetry

---

Have another question? Just tag @inkeep.

[jannikmaierhoefer]

Hi @leo91000, thanks for bringing this up!

As far as I am aware, in the current integration, structured output retries show up as multiple provider-call spans inside a single generateObject call.

The telemetry does not label a span as “retry #2” or attach a “retry reason = schema validation failed” field. You infer retries by counting doGenerate spans / timing.

Which version of the AI SDK are you using? Do you also have a trace where you know that a retry happened but there is only one generation shown? If yes, It would be great if you could share the link to the trace.