76 lines
2.3 KiB
Bash
76 lines
2.3 KiB
Bash
# ============================================
|
|
# WeLe Agentic AI — Environment
|
|
# ============================================
|
|
|
|
# --- Server ---
|
|
PORT=4000
|
|
NODE_ENV=development
|
|
PUBLIC_BASE_URL=http://localhost:4000
|
|
|
|
# --- LLM providers & failover ---
|
|
# Ordered chain of `provider:model`, tried left to right. If the first provider
|
|
# is rate-limited, out of quota, or down, the next takes over mid-turn.
|
|
# Providers: gmi | openrouter (both OpenAI-compatible)
|
|
LLM_CHAIN_SUPERVISOR=gmi:MiniMaxAI/MiniMax-M3,openrouter:nvidia/nemotron-3-ultra-550b-a55b:free
|
|
LLM_CHAIN_AGENT=gmi:MiniMaxAI/MiniMax-M3,openrouter:nvidia/nemotron-3-ultra-550b-a55b:free
|
|
LLM_DEFAULT_PROVIDER=gmi
|
|
|
|
# GMI Cloud — free MiniMax allocation (MiniMaxAI/MiniMax-M3, MiniMaxAI/MiniMax-M2.7)
|
|
GMI_API_KEY=
|
|
GMI_BASE_URL=https://api.gmi-serving.com/v1
|
|
|
|
# OpenRouter — fallback
|
|
OPENROUTER_API_KEY=
|
|
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
|
|
|
|
LLM_MAX_TOKENS=4000
|
|
|
|
# --- MongoDB (shared with CRM — READ path) ---
|
|
MONGODB_URI=
|
|
|
|
# --- CRM REST API (WRITE path) ---
|
|
CRM_API_BASE=http://localhost:3000
|
|
CRM_JWT_SECRET=
|
|
|
|
# --- Redis (memory + cache; optional, falls back to in-memory) ---
|
|
REDIS_ENABLED=true
|
|
REDIS_HOST=localhost
|
|
REDIS_PORT=6379
|
|
REDIS_PASSWORD=
|
|
REDIS_PREFIX=agentic:
|
|
|
|
# --- Session / memory ---
|
|
SESSION_TTL_SECONDS=86400
|
|
CACHE_TTL_SECONDS=300
|
|
MAX_HISTORY_MESSAGES=12
|
|
|
|
# --- Guardrails ---
|
|
REQUIRE_APPROVAL_FOR_WRITES=true
|
|
MAX_TOOL_CALLS_PER_TURN=12
|
|
SUPERVISOR_RECURSION_LIMIT=14
|
|
AGENT_RECURSION_LIMIT=8
|
|
RATE_LIMIT_WINDOW_MS=60000
|
|
RATE_LIMIT_MAX=40
|
|
|
|
# --- Artifacts ---
|
|
ARTIFACT_DIR=./storage/artifacts
|
|
ARTIFACT_TTL_HOURS=72
|
|
|
|
# --- Voice (speech-to-speech, CPU, in-process) ---
|
|
# Models are ONNX via Transformers.js — no GPU, no Python, no second service.
|
|
# STT onnx-community/whisper-tiny.en English only; half the RAM of base
|
|
# TTS Xenova/mms-tts-eng from the Hub
|
|
# Tamil is built and tested (assets/tts/mms-tts-tam, exported locally — no
|
|
# public ONNX exists). Enable it with VOICE_LANGUAGES=en,ta and a multilingual
|
|
# STT_MODEL, but budget ~200 MB more resident for the extra voice.
|
|
VOICE_LANGUAGES=en
|
|
SPEECH_WARMUP=false
|
|
STT_MODEL=onnx-community/whisper-tiny.en
|
|
# Below this confidence, the user's preferred language beats the detector.
|
|
DETECT_CONFIDENCE=0.6
|
|
|
|
# Endpointing
|
|
VAD_SILENCE_MS=700
|
|
VAD_MIN_SPEECH_MS=250
|
|
VAD_PREFIX_MS=300
|