docs(pipecat): document instrument_user_aggregator for realtime user transcript

[Caroline di Vittorio (carolinedivittorio)]

Summary

A speech-to-speech (realtime) Pipecat model (for example, OpenAIRealtimeLLMService) has no separate speech-to-text stage, so the user's transcript is never emitted as an OTel span — it arrives through the user context aggregator's on_user_turn_message_added callback. Without wiring it up, the trace shows only the assistant side.

This mirrors what the LiveKit integration docs did for instrument_session, but for the Pipecat integration's instrument_user_aggregator.

Changes

• New section "When using Pipecat with a realtime model" in src/langsmith/trace-with-pipecat.mdx, placed after "Group a conversation into a thread" (matching the LiveKit doc's structure):
  • Explains that realtime models deliver the user transcript through the user context aggregator, not an OTel span.
  • Shows calling span_processor.instrument_user_aggregator(context_aggregator, conversation_id) right after building the aggregator, correlating by the id set with set_thread_id.
  • Notes that OpenAI Realtime additionally requires input-audio transcription (InputAudioTranscription) on the session, or the aggregator never fires (the model receives raw audio and produces no user-side text). Other realtime services that surface user text through the aggregator themselves (e.g. Gemini Live) need only the instrument_user_aggregator call.
  • Warns not to call it for the STT/LLM/TTS cascade, where the transcript is already captured from the speech-to-text step (would double-count the user's turns).
  • Pins the new API to langsmith[pipecat]>=0.10.6 (the release that introduced instrument_user_aggregator).
• Updated the intro paragraph to note realtime models need one extra call and link to the new section (same shape as the LiveKit doc).

Verification

• instrument_user_aggregator was confirmed to ship in published langsmith==0.10.6 (absent in 0.10.4 / 0.10.5), so the >=0.10.6 pin is accurate.
• The OpenAI Realtime session config (AudioInput / InputAudioTranscription) and LLMContextAggregatorPair(..., realtime_service_mode=True) in the example match the instrumented demo in feat(pipecat): capture realtime user transcript in openai + gemini demos voice-demo#15.
• Anchor #when-using-pipecat-with-a-realtime-model matches the generated heading slug.

[github-actions]

Thanks for opening a docs PR, Caroline di Vittorio (@carolinedivittorio)! When it's ready for review, please add the relevant reviewers:

• @katmayb or @fjmorris (LangSmith)

[github-actions]

Mintlify preview branch generated: preview-pipeca-1784309653-931857a

Site preview: https://langchain-5e9cc07a-preview-pipeca-1784309653-931857a.mintlify.site

Important

Preview links may take a few minutes to start working while the deployment finishes.

Changed documentation pages (preview deep links):

• src/langsmith/trace-with-pipecat.mdx