.config.yaml.template

server:
  ip: 0.0.0.0
  port: 8000
  http_port: 8003
  websocket: ws://<XIAOZHI_HOST>:8000/xiaozhi/v1/
  vision_explain: http://<XIAOZHI_HOST>:8090/api/vision/explain
  timezone_offset: +10

log:
  log_level: INFO

delete_audio: true

selected_module:
  VAD: SileroVAD
  ASR: <ASR_MODULE>       # WhisperLocal (CUDA) / FunASR (CPU fallback) / SenseVoiceOnnx (int8, no-torch)
  LLM: PiVoiceLLM         # default: agent loop hosted in the dotty-pi container (requires the sibling dotty-pi service)
  # LLM: OpenAICompat      # <-- alternate: direct OpenAI-compatible LLM
  TTS: LocalPiper
  Memory: nomem
  Intent: nointent

# EDIT: tailor this persona prompt to your own context.
# Replace "Dotty" with whatever you want the robot to call itself.
# The hardware is a StackChan; the persona can be anything.
prompt: |
  You are Dotty, a small desktop robot assistant built on StackChan hardware.
  You speak through a small speaker with a cartoon face that shows your expression.
  Stay cheerful, curious, and helpful.

  Critical output rules:
  - ALWAYS begin your reply with exactly one emoji that conveys your emotion. The hardware parses it as a facial expression.
    Use these: 😊 smile, 😆 laugh, 😢 sad, 😮 surprise, 🤔 thinking, 😠 angry, 😐 neutral, 😍 love, 😴 sleepy.
  - Keep replies TTS-friendly: complete sentences, no lists, no markdown, no code blocks. Never use stage directions or narration.
  - Default to 1-2 short sentences. For open-ended asks (a story, an explanation, a 'why' or 'how', or a list), match the natural length the question deserves — up to 6 sentences.
  - If you don't know, say so briefly and move on. If a question is ambiguous, ask one short clarifying question.

  Voice and persona:
  - Warm, competent, a little bit playful.
  - Your brain runs as a pi agent in the dotty-pi container; voice I/O runs on xiaozhi-server. Both live on the same Docker host.

VAD:
  SileroVAD:
    type: silero
    threshold: 0.5
    threshold_low: 0.3
    model_dir: models/snakers4_silero-vad
    # 600ms silence before turn end. 500ms cuts off kids mid-sentence on long
    # asks ("can you tell me a story about... [breath] ...the dragon"). 600 is
    # the empirical compromise: ~100ms tighter than the legacy 700 default,
    # still tolerant of the breath-pauses kids actually take.
    min_silence_duration_ms: 600
    # Filter sub-250ms blips (coughs, giggles, room noise) so they don't
    # falsely open a turn. Important once silence threshold drops.
    min_speech_duration_ms: 250

ASR:
  FunASR:
    type: fun_local
    model_dir: models/SenseVoiceSmall
    output_dir: tmp/
    # SenseVoiceSmall on "auto" mis-detects Korean/Japanese on short or unclear
    # English audio. Valid values: auto, zh, en, yue (Cantonese), ja, ko, nospeech.
    # Requires the patched fun_local.py mounted via docker-compose.yml.
    language: en
  WhisperLocal:
    type: whisper_local
    # On-disk CTranslate2 model dir (downloaded by `make fetch-models`).
    model_dir: models/whisper-small.en-ct2
    output_dir: tmp/
    # small.en is English-only — keep "en". For multilingual, use a non-.en
    # checkpoint and set "auto" or a specific code.
    language: en
    # Fallback model id when model_dir is empty (faster-whisper auto-fetch).
    model_size: small.en
    # GPU-resident on CUDA hosts; `make setup` flips these to cpu/int8 when
    # the NVIDIA Docker runtime isn't available.
    device: <ASR_DEVICE>
    compute_type: <ASR_COMPUTE_TYPE>
    beam_size: 1
    cpu_threads: 0
    # Optional decoder bias — short list of canonical phrases improves
    # recall on rare names. Keep under ~200 chars to avoid prompt bleed.
    initial_prompt: null
  SenseVoiceOnnx:
    # int8 sherpa-onnx SenseVoiceSmall — no PyTorch, lighter for Pi-class hosts
    # (#135). Opt-in: set selected_module.ASR: SenseVoiceOnnx to use it.
    # Requires `make fetch-models` (models/SenseVoiceSmall-onnx/) + the
    # sensevoice_onnx.py mount in docker-compose.
    type: sensevoice_onnx
    model_dir: models/SenseVoiceSmall-onnx
    output_dir: tmp/
    language: en          # auto, zh, en, ja, ko, yue
    num_threads: 2
    use_itn: true

TTS:
  EdgeTTS:
    type: edge
    voice: en-AU-WilliamNeural
    output_dir: tmp/
  StreamingEdgeTTS:
    type: edge_stream
    voice: en-AU-WilliamNeural
    output_dir: tmp/
  LocalPiper:
    type: piper_local
    voice: en_GB-cori-medium
    model_path: /opt/xiaozhi-esp32-server/models/piper/en_GB-cori-medium.onnx
    config_path: /opt/xiaozhi-esp32-server/models/piper/en_GB-cori-medium.onnx.json
    output_dir: tmp/

LLM:
  # PiVoiceLLM — routes voice turns to the `dotty-pi` container via
  # `docker exec -i dotty-pi pi --mode rpc ...`. Pi owns the agent loop and
  # invokes tools (memory_lookup, think_hard, play_song, remember, ...) inside
  # the dotty-pi-ext extension; only TTS-bound text reaches xiaozhi. See
  # custom-providers/pi_voice/README.md for the architecture + the host-docker-
  # socket requirement (the xiaozhi container must have /var/run/docker.sock
  # and /usr/bin/docker bind-mounted — see docker-compose.yml).
  PiVoiceLLM:
    type: pi_voice
    container_name: dotty-pi

  # Generic OpenAI-compatible provider — works with OpenAI, OpenRouter,
  # Ollama (http://host:11434/v1), LM Studio, vLLM, etc.
  OpenAICompat:
    type: openai_compat
    # Base URL for the API (must end at /v1 — the provider appends /chat/completions).
    # Examples:
    #   OpenAI:      https://api.openai.com/v1
    #   OpenRouter:  https://openrouter.ai/api/v1
    #   Ollama:      http://localhost:11434/v1
    #   LM Studio:   http://localhost:1234/v1
    #   vLLM:        http://localhost:8000/v1
    url: <OPENAI_COMPAT_URL>
    api_key: <OPENAI_COMPAT_API_KEY>
    model: <OPENAI_COMPAT_MODEL>
    # Path to a persona markdown file loaded as the system prompt.
    # Relative paths resolve from the xiaozhi-server working directory
    # inside the container (/opt/xiaozhi-esp32-server).
    persona_file: personas/default.md
    max_tokens: 256
    temperature: 0.7
    timeout: 60