ZizkaDB + LiveKit Agents

PyPI: zizkadb-livekit 0.2.3 · depends on zizkadb-sdk>=0.5.0 + livekit-agents>=1.3.0

Log voice calls to ZizkaDB with the same Activity → Sessions / Events model as text agents.

• One session_id per call (use the LiveKit room name or job id)
• Transcript only — text copied from LiveKit's chat history and make_session_report() (no audio in ZizkaDB)
• Pipeline events — tool calls, handoffs, state changes, interruptions, token usage
• Causal lineage — a single-stranded parent_id chain for Why? in the dashboard
• Never on the audio path — writes are queued and drained in the background; a ZizkaDB outage degrades observability, never the call

Install

pip install "zizkadb-livekit>=0.2.3"

Start ZizkaDB (Docker quickstart):

curl -fsSL https://raw.githubusercontent.com/ZIZKA-AI-SL/ZizkaDB/main/scripts/quickstart-remote.sh | bash

Usage

import os
from livekit.agents import AgentServer, JobContext, AgentSession, Agent
from zizkadb import ZizkaDB
from zizkadb_livekit import ZizkaDBLiveKitObserver, pop_observer, register_observer

server = AgentServer()


async def on_session_end(ctx: JobContext) -> None:
    # on_session_end receives the same JobContext as the entrypoint, so the
    # observer is looked up per call. One worker serves many calls concurrently;
    # a module-level observer would let them overwrite each other's transcripts.
    observer = pop_observer(ctx)
    if observer is None:
        return
    try:
        await observer.ingest_session_report(ctx)
    finally:
        await observer.aclose()


@server.rtc_session(agent_name="support-voice", on_session_end=on_session_end)
async def entrypoint(ctx: JobContext):
    db = ZizkaDB(host=os.getenv("ZIZKADB_HOST", "http://localhost:8000"))
    observer = ZizkaDBLiveKitObserver(
        db,
        agent=os.getenv("ZIZKADB_AGENT", "support-voice"),
        session_id=ctx.room.name,
    )
    register_observer(ctx, observer)

    session = AgentSession(...)  # your STT / LLM / TTS
    # Attach before session.start() so no turn is missed.
    observer.attach(session, job_ctx=ctx)
    observer.queue_session_started(room=ctx.room.name, job_id=ctx.job.id)

    await session.start(agent=Agent(instructions="..."), room=ctx.room)
    await ctx.connect()

Open http://localhost:3001 → Activity → your agent → Sessions.

Verbosity

Voice pipelines emit many more events than text agents, so events are tiered. Each level logs its own tier and everything quieter; errors and session boundaries are always logged.

observer = ZizkaDBLiveKitObserver(db, agent="...", session_id="...", level="verbose")

or set ZIZKADB_EVENT_LEVEL globally.

Event types

Manual helpers: log_tool_call, log_tool_result, log_error.

Lifecycle

• attach(session, job_ctx=ctx) — subscribe to the live event stream, before session.start(). Never raises.
• queue_session_started(...) — non-blocking; use await log_session_started(...) if you need the event id.
• ingest_session_report(ctx) — back-fill the full transcript at call end.
• flush(timeout=None) — wait for queued writes (default flush_timeout, 10s).
• aclose() — flush, stop the writer, release the pooled connection. Idempotent; also usable as async with.

Environment

export ZIZKADB_HOST=http://localhost:8000   # self-host
export ZIZKADB_AGENT=support-voice
export ZIZKADB_EVENT_LEVEL=standard         # transcript | standard | verbose
# Cloud: export ZIZKADB_API_KEY=zizkadb_live_...

See examples/livekit-agent for a full sample.