# ── LLM Provider ─────────────────────────────────────────────────────────────
# Set LLM_PROVIDER to one of: groq | openrouter | anthropic
LLM_PROVIDER=groq

# API key for the chosen provider (use one of the three lines below)
GROQ_API_KEY=your_groq_api_key_here
# OPENROUTER_API_KEY=your_openrouter_api_key_here
# ANTHROPIC_API_KEY=your_anthropic_api_key_here

# Optional: override the model (defaults shown per provider)
# Groq default:       llama-3.3-70b-versatile
# OpenRouter default: meta-llama/llama-3.3-70b-instruct
# Anthropic default:  claude-sonnet-4-20250514
# LLM_MODEL=llama-3.3-70b-versatile

# ── Transcription (optional, for Safari/iOS fallback) ────────────────────────
OPENAI_API_KEY=your_openai_api_key_here_optional
OPENAI_TRANSCRIBE_MODEL=gpt-4o-mini-transcribe

# ── TTS ───────────────────────────────────────────────────────────────────────
# Cartesia is the primary path (~200ms first byte). The backend reuses a
# per-turn `context_id` for streaming continuation so all chunks of one LLM
# turn share the same voice context — saves ~30-50ms of per-chunk overhead.
CARTESIA_API_KEY=your_cartesia_api_key_here
CARTESIA_VOICE_ID=db6b0ed5-d5d3-463d-ae85-518a07d3c2b4
CARTESIA_MODEL_ID=sonic-2

# Kokoro fallback (only used if Cartesia is missing/unreachable).
KOKORO_URL=http://localhost:8880
KOKORO_VOICE=af_heart

# ── STT (Deepgram streaming, lowest mic-release-to-final latency) ─────────────
DEEPGRAM_API_KEY=your_deepgram_api_key_here
# Trailing-silence (ms) before Deepgram fires `is_final`. Default 100ms — the
# frontend buffers all `is_final` segments through the hold and only commits
# on mic-release, so aggressive endpointing just makes the post-release flush
# land faster. Increase to 300+ if you see false-final fragmentation.
DEEPGRAM_ENDPOINTING_MS=100

# ── Frontend / CORS ───────────────────────────────────────────────────────────
VITE_BACKEND_URL=http://localhost:8000
ALLOWED_ORIGINS=*