config.yaml.example

# OpenMux 配置示例
# 开箱即用：只需在 .env 中填入对应平台的 API Key 即可启动
# 支持的免费模型平台：智谱GLM、硅基流动、OpenRouter、魔搭ModelScope、阿里云百炼、DeepSeek、美团LongCat

# 服务配置
server:
  port: 8080
  host: "0.0.0.0"
  timeout: 120s
  read_timeout: 60s
  write_timeout: 60s

# 认证配置
auth:
  enabled: false
  # api_keys:
  #   - key: "sk-your-service-key-1"
  #     name: "client-1"
  #     rate_limit:
  #       rpm: 500
  #       tpm: 200000
  #       concurrent: 50

# ============================================================
# Provider 配置
# 每个 provider 只需配置对应的环境变量 API Key 即可使用
# 不需要的 provider 可以注释掉或删除
# ============================================================
providers:
  # ----------------------------------------------------------
  # 智谱 GLM (多款免费模型)
  # 申请 Key: https://open.bigmodel.cn
  # 免费模型: glm-4-flash, glm-4-flash-250414, glm-4v-flash,
  #           glm-4.7-flash, glm-z1-flash, glm-4.5-flash
  # 限制: 30 并发
  # ----------------------------------------------------------
  zhipu:
    base_url: "https://open.bigmodel.cn/api/paas/v4"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 60
      tpm: 50000
      concurrent: 30
    api_keys:
      - "${ZHIPU_API_KEY}"

  # ----------------------------------------------------------
  # 硅基流动 SiliconFlow (多款免费模型)
  # 申请 Key: https://cloud.siliconflow.cn
  # 免费模型: Qwen/Qwen2.5-7B-Instruct, Qwen/Qwen2.5-14B-Instruct,
  #           deepseek-ai/DeepSeek-V3, deepseek-ai/DeepSeek-R1,
  #           deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, Qwen/QwQ-32B
  # 注意: Pro/ 前缀为付费加速版
  # ----------------------------------------------------------
  siliconflow:
    base_url: "https://api.siliconflow.cn/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 100
      tpm: 50000
      concurrent: 10
    api_keys:
      - "${SILICONFLOW_API_KEY}"

  # ----------------------------------------------------------
  # OpenRouter (约 30 款免费模型，模型名带 :free 后缀)
  # 申请 Key: https://openrouter.ai/keys
  # 免费模型: google/gemini-2.5-flash:free,
  #           meta-llama/llama-3.3-70b-instruct:free,
  #           deepseek/deepseek-v3.2-20251201:free,
  #           qwen/qwen3-coder:free 等
  # 限制: ~20 RPM, ~200 请求/天
  # ----------------------------------------------------------
  openrouter:
    base_url: "https://openrouter.ai/api/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 20
      tpm: 50000
      concurrent: 5
    api_keys:
      - "${OPENROUTER_API_KEY}"

  # ----------------------------------------------------------
  # 魔搭 ModelScope (免费，每天 2000 次调用)
  # 申请 Token: https://modelscope.cn/my/myaccesstoken
  # 免费模型: Qwen/Qwen3-235B-A22B-Instruct-2507,
  #           Qwen/Qwen3-32B, deepseek-ai/DeepSeek-V3-0324,
  #           deepseek-ai/DeepSeek-R1-0528, ZhipuAI/GLM-4.6 等
  # 限制: 2000 次/天，每日 00:00 (UTC+8) 重置
  # ----------------------------------------------------------
  modelscope:
    base_url: "https://api-inference.modelscope.cn/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 60
      rpd: 2000     # 每天 2000 次调用限制
      tpm: 100000
      concurrent: 10
    api_keys:
      - "${MODELSCOPE_API_KEY}"

  # ----------------------------------------------------------
  # 阿里云百炼 DashScope (qwen-turbo 等有免费额度)
  # 申请 Key: https://dashscope.console.aliyun.com
  # ----------------------------------------------------------
  aliyun:
    base_url: "https://dashscope.aliyuncs.com/compatible-mode/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 200
      tpm: 100000
      concurrent: 20
    api_keys:
      - "${ALIYUN_API_KEY}"

  # ----------------------------------------------------------
  # DeepSeek
  # 申请 Key: https://platform.deepseek.com
  # ----------------------------------------------------------
  deepseek:
    base_url: "https://api.deepseek.com/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 200
      tpm: 100000
      concurrent: 20
    api_keys:
      - "${DEEPSEEK_API_KEY}"

  # ----------------------------------------------------------
  # 美团 LongCat (公测免费，每日 50 万 tokens)
  # 申请 Key: https://longcat.chat/platform
  # 免费模型: LongCat-Flash-Chat, LongCat-Flash-Thinking,
  #           LongCat-Flash-Lite (每日 5000 万 tokens)
  # ----------------------------------------------------------
  longcat:
    base_url: "https://api.longcat.chat/openai/v1"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 60
      tpm: 50000
      concurrent: 10
    api_keys:
      - "${LONGCAT_API_KEY}"

  # ----------------------------------------------------------
  # 火山引擎 Volcengine / 豆包 Doubao (新用户每模型 50 万 tokens 免费)
  # 申请 Key: https://www.volcengine.com
  # 免费模型: doubao-lite-32k, doubao-pro-32k 等
  # 注意: API Key 格式为 ark-xxxxx
  # ----------------------------------------------------------
  volcengine:
    base_url: "https://ark.cn-beijing.volces.com/api/v3"
    type: "openai"
    timeout: 60s
    rate_limit:
      rpm: 120
      tpm: 100000
      concurrent: 10
    api_keys:
      - "${VOLCENGINE_API_KEY}"


# ============================================================
# 模型路由配置
# 通过别名将请求路由到不同的 provider/model 组合
# 客户端只需使用别名即可，无需关心具体 provider
# ============================================================
model_routes:
  # 免费聊天模型 - 多平台 fallback
  chat-free:
    description: "免费聊天模型 (多平台自动 fallback)"
    targets:
      - provider: zhipu
        model: glm-4-flash
        weight: 3
      - provider: siliconflow
        model: Qwen/Qwen2.5-7B-Instruct
        weight: 2
      - provider: longcat
        model: LongCat-Flash-Chat
        weight: 2
      - provider: modelscope
        model: Qwen/Qwen3-8B
        weight: 1
      - provider: openrouter
        model: meta-llama/llama-3.3-70b-instruct:free
        weight: 1
    strategy: weighted_round_robin

  # 免费推理模型 - 思考链
  reasoning-free:
    description: "免费推理模型 (thinking)"
    targets:
      - provider: siliconflow
        model: deepseek-ai/DeepSeek-R1
        weight: 3
      - provider: modelscope
        model: deepseek-ai/DeepSeek-R1-0528
        weight: 2
      - provider: longcat
        model: LongCat-Flash-Thinking
        weight: 2
      - provider: zhipu
        model: glm-z1-flash
        weight: 1
    strategy: weighted_round_robin

  # 免费大参数模型 (质量优先)
  chat-free-large:
    description: "免费大参数模型 (质量更高)"
    targets:
      - provider: modelscope
        model: Qwen/Qwen3-235B-A22B-Instruct-2507
        weight: 3
      - provider: siliconflow
        model: deepseek-ai/DeepSeek-V3
        weight: 2
      - provider: openrouter
        model: deepseek/deepseek-v3.2-20251201:free
        weight: 1
    strategy: weighted_round_robin

  # 通用大模型 (付费，质量更高)
  chat:
    description: "通用大模型 (付费)"
    targets:
      - provider: deepseek
        model: deepseek-chat
        weight: 3
      - provider: aliyun
        model: qwen-plus
        weight: 2
      - provider: zhipu
        model: glm-4-plus
        weight: 1
    strategy: weighted_round_robin

  # 免费高吞吐轻量模型
  chat-lite:
    description: "轻量高速模型 (免费)"
    targets:
      - provider: longcat
        model: LongCat-Flash-Lite
        weight: 3
      - provider: zhipu
        model: glm-4-flash
        weight: 2
      - provider: siliconflow
        model: Qwen/Qwen2.5-7B-Instruct
        weight: 1
    strategy: weighted_round_robin

# ============================================================
# 智能路由 (Auto Route)
# 根据请求复杂度自动分流到不同层级的模型
# 模型层级由 discovery 从模型名自动解析参数量并分级:
#   lite (<10B), standard (10-72B), large (>72B), reasoning (thinking/R1)
# 用法: curl -d '{"model":"auto", ...}'
# ============================================================
auto_route:
  enabled: true
  alias: "auto"                     # 触发别名
  classifier: "zhipu/glm-4-flash"   # 可选: 用小模型辅助分类，不配则纯规则
  # 可选: 手动覆盖某个层级的路由 (不配则自动使用 discovery 生成的 auto:* 路由)
  # lite: "chat-lite"
  # standard: "chat-free"
  # large: "chat-free-large"
  # reasoning: "reasoning-free"

# 直通模式 - 允许 provider/model 格式直接访问
passthrough:
  enabled: true
  allowed_providers:
    - zhipu
    - siliconflow
    - openrouter
    - modelscope
    - aliyun
    - deepseek
    - longcat
    - volcengine

# ============================================================
# 模型自动发现
# 定时从各平台 /v1/models 接口获取最新免费模型列表
# 发现的模型会自动注册路由，可通过 provider/model 格式访问
# 同时自动生成聚合别名 "free"，从每个平台选最佳免费模型组成 fallback 组
# 用法: curl -d '{"model":"free", ...}' 即可调用
# ============================================================
discovery:
  enabled: true
  interval: 24h       # 刷新间隔
  free_alias: "free"  # 聚合别名，所有免费模型的入口
  providers:           # 需要自动发现的 provider
    - openrouter
    - siliconflow
    - modelscope

# 监控配置
monitoring:
  enabled: true
  metrics_path: "/metrics"
  log_level: "info"

# 缓存配置
cache:
  enabled: false
  ttl: 3600s
  max_size: 1000