Hello @User I’m currently using the Vapi → Langfuse integration for observability and had a question regarding the granularity of the trace data being sent. From the traces I inspected (including the example trace in the documentation), it appears the integration logs LLM generation observations with start and end timestamps, but the conversation messages themselves (user/assistant turns) do not have associated timestamps. Because of this, it’s difficult to reconstruct the actual turn-by-turn timeline of a call. For example, while generation spans include timestamps, the conversation messages inside the trace input/output do not appear to carry per-turn timing information. This makes it challenging to analyze: when each user turn occurred the latency between user speech and model response full call timelines for debugging or analytics Is the current Langfuse integration intentionally limited to logging LLM generations rather than full conversational events? Are there plans to extend the integration to include richer call timeline telemetry (for example user turns, STT events, TTS playback events, etc.)? We’re trying to construct a full call timeline in Langfuse rather than only model generation telemetry, so any guidance on the recommended approach would be appreciated. Thanks.

Currently, the Vapi → Langfuse integration logs LLM generation traces with timestamps, but does not natively include per-turn (user/assistant) timestamps. As a workaround, you can enrich your traces by attaching custom metadata for each turn:

json
"assistant.observabilityPlan.metadata": {
  "user_turn_timestamp": "2026-03-16T08:37:50Z",
  "assistant_turn_timestamp": "2026-03-16T08:37:52Z"
}

This allows you to manually track turn timing within Langfuse until native support is available. For more details, see the [official documentation](https://docs.vapi.ai/providers/observability/langfuse). Source: - [Langfuse Integration with Vapi Documentation](https://docs.vapi.ai/providers/observability/langfuse)

Hey, You're correct - the current Langfuse integration only includes timestamps on LLM generation spans, not on individual conversation messages. This is a known limitation of the current implementation. We've escalated this to the engineering team. We'll follow up once we have more clarity on whether fuller call timeline telemetry (user turns, STT/TTS events, per-turn latency) is on the roadmap. Best, Vapi Support