{
  "lesson": "22-voice-agents-pipecat-livekit",
  "title": "Voice Agents: Pipecat and LiveKit",
  "questions": [
    {
      "stage": "pre",
      "question": "Which two flow directions does a Pipecat pipeline use?",
      "options": [
        "Read and write",
        "Hot and cold",
        "Inbound and outbound",
        "DOWNSTREAM (source to sink) and UPSTREAM (feedback, cancel, barge-in)"
      ],
      "correct": 3,
      "explanation": "Frames flow downstream source-to-sink and upstream for control and cancellation."
    },
    {
      "stage": "pre",
      "question": "What is the canonical Pipecat voice pipeline?",
      "options": [
        "VAD -> STT -> LLM -> TTS -> transport",
        "Audio -> JSON -> SQL -> response",
        "LLM -> embed -> retrieve -> answer",
        "TTS -> STT -> LLM -> VAD"
      ],
      "correct": 0,
      "explanation": "VAD detects voice activity, STT transcribes, LLM responds, TTS speaks, transport delivers."
    },
    {
      "stage": "check",
      "question": "Which two voice agent classes does LiveKit Agents ship?",
      "options": [
        "BatchAgent and StreamAgent",
        "TextAgent and SpeechAgent",
        "MultimodalAgent (direct audio) and VoicePipelineAgent (STT/LLM/TTS cascade)",
        "LocalAgent and CloudAgent"
      ],
      "correct": 2,
      "explanation": "MultimodalAgent uses direct audio (Realtime-style); VoicePipelineAgent uses STT->LLM->TTS for text-level control."
    },
    {
      "stage": "check",
      "question": "What is barge-in and how is it handled?",
      "options": [
        "An LLM cost spike; reduce tokens",
        "The user interrupts while the agent is speaking; UPSTREAM cancel frames stop TTS mid-utterance",
        "Captcha failure; retry",
        "A provider outage; switch regions"
      ],
      "correct": 1,
      "explanation": "Barge-in is user interruption; UPSTREAM cancellation is how Pipecat handles it cleanly."
    },
    {
      "stage": "check",
      "question": "What end-to-end latency does the lesson describe as premium?",
      "options": [
        "About 450-600 ms",
        "About 1500 ms",
        "About 50 ms",
        "About 5000 ms"
      ],
      "correct": 0,
      "explanation": "Premium stacks land around 450-600 ms; 800-1200 ms is common; over 1500 ms feels broken."
    },
    {
      "stage": "post",
      "question": "What goes wrong if STT confidence is ignored?",
      "options": [
        "Calls get cheaper",
        "Latency improves",
        "TTS gets faster",
        "Low-confidence transcripts feed the LLM as if gospel, producing wrong answers; gate on confidence or ask for confirmation"
      ],
      "correct": 3,
      "explanation": "Treating low-confidence STT as truth is a top voice-agent failure mode."
    },
    {
      "stage": "post",
      "question": "Why does the lesson recommend summing component latencies before shipping?",
      "options": [
        "Latency is required by WebRTC",
        "Every component adds 50-200 ms; the sum determines whether the experience feels broken",
        "Providers bill by latency",
        "It satisfies a compliance requirement"
      ],
      "correct": 1,
      "explanation": "Sum the chain (VAD + STT + LLM + TTS + transport) before shipping; targets are tight."
    }
  ]
}