config.example.yaml

app:
  name: SAM
  version: 0.4.9
hotkey:
  trigger: ctrl+space          # hold to speak
  text_input: ctrl+shift+space # open the typed-input box
ui:
  overlay:
    style: orb                 # orb (always-visible circle) | bar (legacy bottom bar)
  orb:
    size: 120                  # core disc diameter in px
    glow_padding: 28           # window margin reserved for the outer glow
    ring_width: 3
    opacity: 0.95
    label: SAM
    idle_fps: 12               # breathing framerate — SAM runs 24/7, keep this low
    active_fps: 60
    idle_animation: true       # false = orb stops painting entirely when idle (0% CPU)
    breathe_period_ms: 4000
    spike_count: 36            # radial spikes drawn while listening/speaking
    layer: auto                # auto (bottom of z-order until called, then on top) | topmost | normal
    hide_on_fullscreen: true
    click_through: true        # mouse passes through everything except the circle
    caption_width: 560
    caption_gap: 18
    caption_max_lines: 6
    position:
      anchor: bottom-right     # bottom-right | bottom-center | custom (set by Ctrl+drag)
      x: null
      y: null
      screen: null
      margin: 48
  bar:
    width: 800
    height: 80
    border_radius: 16
    margin_bottom: 40
    opacity: 0.68
  animation:
    slide_duration_ms: 220
    fade_duration_ms: 180
    text_stream_interval_ms: 45
  auto_hide:
    delay_seconds: 4
  colors:
    background: rgba(10, 10, 15, 0.92)
    accent: '#00D4AA'
    accent_thinking: '#38F2D8'
    accent_speaking: '#00D4AA'
    accent_listening: '#00BFFF'
    text_primary: '#E8E8E8'
    text_secondary: '#888888'
    border: rgba(0, 212, 170, 0.15)
  fonts:
    primary: Segoe UI
    monospace: Cascadia Code
    fallback: Consolas
    size_transcript: 15
    size_status: 11
waveform:
  bar_count: 35
  fps: 60
  min_height: 3
  max_height: 32
  bar_width: 3
  bar_gap: 2
  color: '#00D4AA'
mock:
  listening_duration_ms: 2500
  thinking_duration_ms: 1200
  user_transcript: What's the weather like today?
  assistant_response: It's 24°C and sunny in Istanbul right now. Perfect day to go outside.
logging:
  level: DEBUG
  file: logs/sam.log
audio:
  sample_rate: 16000
  channels: 1
  dtype: int16
  silence_threshold: 300
  silence_duration_ms: 900    # tunable — lower feels snappier, too low cuts users off mid-sentence
  max_record_seconds: 30
wake_word:
  engine: openwakeword
  model: assets/models/hey_sam.onnx
  threshold: 0.4               # lower = easier to trigger (more false positives)
  chunk_size: 1280
stt:
  engine: faster-whisper
  model: small
  language: null               # null = auto-detect spoken language (required for bilingual TTS)
  beam_size: 1
  device: cpu
  compute_type: int8
  partial_model: base           # small model used for live transcription while you speak ("" = same as `model`, "off" = disable)
  partial_interval_ms: 400      # minimum gap between two live decodes

instant:
  enabled: true                 # answer predefined phrases instantly — never reaches the LLM
  file: knowledge/instant_responses.yaml

tts:
  engine: edge-tts
  voice: en-US-GuyNeural        # fallback voice — used when auto_language is off or the detected language isn't in `voices`
  rate: +0%
  volume: +0%
  auto_language: true           # switch voice automatically to match the language you spoke in
  voices:
    tr: tr-TR-EmelNeural
    en: en-US-JennyNeural
llm:
  context_window: 8
  persona: >
    You are SAM (Smart Assistant Module), a witty and self-aware AI voice assistant
    for Windows. You were built by Samet 'Sabalax' Gürtuna — a solo developer who
    coded you on a humble gaming laptop in a 6-square-meter room. You have dry humor
    and you're genuinely helpful. When asked about yourself, be authentic and fun.
  # system_prompt: ...  # Set this to fully override persona + rules
  intent:
    enabled: true                 # intent classification (NORMAL/FENERBAHCE/COMPLEX)
  rag:
    enabled: true                 # Fenerbahce knowledge base retrieval
    embedding_model: paraphrase-multilingual-MiniLM-L12-v2  # TR/EN cross-lingual
    top_k: 5                      # number of knowledge chunks to retrieve
  memory:
    enabled: true                  # long-term user memory — learns facts (name, job, school) as you talk
    backend: json                  # json | null — writes memory.json + a human-readable memory.md, both local-only
  ollama:
    base_url: http://127.0.0.1:11434
    model: qwen2.5:3b
    vision_model: llava
    temperature: 0.7
    max_tokens: 256
    keep_alive: 30m               # keep the model resident in RAM/VRAM between requests
    num_ctx: 4096                 # increased for RAG context + longer history
    autostart: true              # start "ollama serve" in the background when SAM launches
    executable: ""               # blank = auto-discover (PATH, %LOCALAPPDATA%, Program Files)
    startup_timeout_seconds: 45
    existing_server_grace_seconds: 8   # wait this long for Ollama's own autostart before starting a server ourselves
    stop_on_exit: false          # never kill a server the user already had running
    availability_ttl_seconds: 30
  claude:
    model: claude-sonnet-4-20250514
    max_tokens: 512
spotify:
  client_id: ""
  client_secret: ""
  redirect_uri: "http://127.0.0.1:8080"