# Copy to .env and edit. Do not commit your real .env (git-ignored).

# --- Bind address (the app connects here: wss://<host-lan-ip>:<PORT>) ---
HOST=0.0.0.0
PORT=8080

# --- LLM (the brain) via Ollama's OpenAI-compatible endpoint ---
OLLAMA_BASE_URL=http://localhost:11434/v1
# Brain (swappable — the host owns the model, no app change to switch). The LLM dominates turn latency.
# qwen3.5:27b (~17GB) is the default: qwen3.5:9b fakes actions and misses portal.end_conversation.
# Run Ollama with OLLAMA_KEEP_ALIVE=-1 so the first turn isn't a cold reload.
OLLAMA_MODEL=qwen3.5:27b
# Cap model output length (reasoning is off, so this is spoken text and tool calls only). 128 is too
# low when 20+ portal/external tools are registered (the turn ends with no tool_call and no speech).
# 1024 is a safe default for tool turns; spoken replies stay short via the system prompt.
OLLAMA_MAX_TOKENS=1024
# Return only Tavily's synthesized answer (skip page snippets) to shrink the post-search LLM prefill.
WEB_SEARCH_ANSWER_ONLY=true
# Preload the model in the background at boot (cuts first-turn cold-start latency).
WARM_OLLAMA_AT_STARTUP=true
# Pre-synthesize one Kokoro phrase at boot (cuts first-turn TTS TTFB).
WARM_KOKORO_AT_STARTUP=true
# Prime web_search at boot (Tavily connection / DuckDuckGo path).
WARM_TAVILY_AT_STARTUP=true
# Load Whisper weights into MLX at boot (cuts first-turn ASR cold-start latency).
WARM_WHISPER_AT_STARTUP=true

# --- ASR (Whisper on MLX, Apple Silicon) ---
WHISPER_MODEL=mlx-community/whisper-large-v3-turbo   # latency pick; mlx-community/whisper-large-v3 for max accuracy

# --- TTS ---
KOKORO_VOICE=af_heart             # any installed Kokoro voice

# --- Web search grounding (host-side tool; weather uses Open-Meteo via get_weather) ---
# Backend: tavily (default, needs API key) or duckduckgo (free, no key).
WEB_SEARCH_BACKEND=tavily
TAVILY_API_KEY=tvly-xxxxxxxxxxxxxxxxxxxx

# --- Audio contract (must match the app: 16k mono in / 24k mono out). Change only if you know why. ---
INPUT_SAMPLE_RATE=16000
OUTPUT_SAMPLE_RATE=24000
# Silero VAD: seconds of silence before the host ends your turn (Pipecat default is 0.2).
# Raise if the host jumps in during short pauses or clips your last word; lower if it feels sluggish.
VAD_STOP_SECS=0.8