AI Observability troubleshooting

Contents

This page covers troubleshooting for AI Observability. For setup, see the installation guides.

How much does AI Observability cost?

Your first 100,000 $ai_model events each month are free – i.e. if you never exceed this number, you can use AI Observability for free.

After this, we charge a small amount for each $ai_* event you send.

Go to the pricing page to use our calculator to get an estimate. You can also view an estimate on your billing page.

Why can't I see any of my LLM events?

There are a few reasons why you might not be seeing any LLM events:

  • There might be a delay in the events being processed. You can check the status of our ingestion on the PostHog status page.
  • If you are in a serverless environment make sure you are calling posthog.shutdown() before ending the request.

Why does my generation show no output?

The trace view shows No output when a generation event has no response content to render. When the event carries enough signal to explain the gap, a sentence appears under the box. Each sentence points at a different cause:

The notice saysWhat happenedWhat to do
The response hit its token limit.The provider stopped the response at the request's output limit, so the content was cut off.Raise max_tokens (or your provider's equivalent) on the request.
The provider blocked the response.The provider's safety or content filter stopped the response before any content was returned.Review the prompt and your provider's safety settings. The provider's own response includes the details.
The model may have spent its budget on reasoning.The model used the output budget on reasoning tokens and returned no text.Raise the output token limit, or lower the reasoning effort or thinking budget.
All of them may have been reasoning.The reasoning token count matches the output token count, so every billed token was likely reasoning.Raise the output token limit, or lower the reasoning effort or thinking budget.
None of them were text.The provider reported that none of the output tokens were text. The output may have been another modality, like audio or images.Check which output modalities the request asked for.
The response may have been cut short, or the SDK may not have captured it.The event has no stop reason, so PostHog can't tell whether the response was interrupted or the content never reached the event.Update the PostHog SDK to the latest version. If you capture generation events manually, include $ai_output_choices and $ai_stop_reason.
Only No output, with no sentence under itThe provider billed no tokens for this generation.Check the provider-side logs for this request.

These sentences are read from the event's $ai_output_tokens, $ai_reasoning_tokens, $ai_text_output_tokens, and $ai_stop_reason properties. Events captured by older SDK versions often lack $ai_stop_reason, so updating the SDK is the quickest way to get a specific explanation.

Solved community questions

Still have questions?

Was this page useful?