LangGraph instrumentation

parlot-instrumentation-langgraph registers a global LangChain callback handler via LangChain’s register_configure_hook / CallbackManager.configure injection so every invoke / ainvoke / astream emits Parlot Conversation Contract spans plus OTel GenAI operational spans.

Quick start

from parlot.instrumentation.langgraph import parlotize, close_session

parlotize("support-bot", version="0.1.0")

graph.invoke(
    {"messages": [("user", "Hello")]},
    config={"configurable": {"thread_id": "demo-1"}},
)

close_session("demo-1")  # optional; also flushed on process exit

Environment: PARLOT_ENDPOINT, PARLOT_API_KEY — see Environment Variables. Shared parlotize() options: Python SDK Reference. LangGraph-only kwargs (channel, modality) and close_session: Interactive API Reference.

For standalone (non-LiveKit) sessions, set channel= (e.g. "webchat") and optionally modality= so ingest does not assume voice:

parlotize("support-bot", version="0.1.0", channel="webchat", modality="text")

Generative AI content capture is shared across adapters — see Concepts → Generative AI content capture. Override for the process with parlotize(capture_genai_content=False).

See examples/langgraph/minimal-agent.

Span vocabulary

Session key is LangGraph thread_id (session.modality=text). If thread_id is omitted, Parlot mints one automatically (anon-…) for the process lifetime.

LiveKit + LangGraph

When a LiveKit voice agent uses livekit.plugins.langchain.LLMAdapter, call both:

from parlot.instrumentation.livekit import parlotize as parlotize_livekit
from parlot.instrumentation.langgraph import parlotize as parlotize_langgraph

parlotize_livekit("voice-bot", version="0.1.0")
parlotize_langgraph("voice-bot")  # adopts LiveKit TracerProvider when already set

• LiveKit owns parlot.session / parlot.turn (voice).
• LangGraph emits nested GenAI ops (chat, execute_tool, invoke_workflow) under the active OTel context (LiveKit’s remapped chat span).
• No duplicate contract spans.