Model Configuration

LingXiao allows separate configuration of Leader and Worker LLM models, with support for multiple providers and compatibility layers.

Configuration Methods

Environment Variables

export LINGXIAO_LLM_PROVIDER=openai
export LINGXIAO_LEADER_MODEL=gpt-4o
export LINGXIAO_AGENT_MODEL=gpt-4o-mini

Config File

Edit ~/.lingxiao/settings.json:

{
  "llm": {
    "provider": "openai",
    "leader_model": "gpt-4o",
    "agent_model": "gpt-4o-mini",
    "enable_streaming": true
  }
}

LLM Config Group

Key settings in the llm group:

Setting	Description
`provider`	Provider selection: `auto` / `openai` / `anthropic`
`leader_model`	Leader model
`agent_model`	Worker model
`wiki_model`	Wiki knowledge summary model
`model_providers`	Provider model registry
`gateway_routes`	Multi-model gateway routes
`gateway_fallback_models`	Gateway fallback models
`request_timeout_s`	Request timeout (seconds)
`connect_timeout_s`	Connect timeout (seconds)
`max_retries`	Max retry count
`backoff_base_ms`	Backoff base milliseconds
`context_max_tokens`	Context max tokens
`capped_max_tokens`	Capped max tokens
`escalated_max_tokens`	Escalated max tokens
`thinking_budget_tokens`	Thinking budget tokens
`enable_streaming`	Enable streaming output
`show_thinking_content`	Show thinking content
`enable_thinking_instruction`	Enable thinking instruction
`enable_extended_thinking`	Enable extended thinking
`reasoning_effort`	Reasoning effort

Supported Providers

Provider	provider value	API format	Notes
OpenAI	`openai`	OpenAI API	Native support
Anthropic	`anthropic`	Anthropic API	Native support
DeepSeek	`openai`	OpenAI compatible	Set `LINGXIAO_OPENAI_BASE_URL`
Qwen	`openai`	OpenAI compatible	Set `LINGXIAO_OPENAI_BASE_URL`
Moonshot/Kimi	`openai`	OpenAI compatible	Set `LINGXIAO_OPENAI_BASE_URL`
Gemini	`openai`	OpenAI compat layer	Via compatibility layer
Groq	`openai`	OpenAI compatible	Set `LINGXIAO_OPENAI_BASE_URL`
SiliconFlow	`openai`	OpenAI compatible	Set `LINGXIAO_OPENAI_BASE_URL`

Compatible Provider Examples

DeepSeek example:

export LINGXIAO_LLM_PROVIDER=openai
export LINGXIAO_OPENAI_API_KEY=sk-your-deepseek-key
export LINGXIAO_OPENAI_BASE_URL=https://api.deepseek.com/v1
export LINGXIAO_LEADER_MODEL=deepseek-chat
export LINGXIAO_AGENT_MODEL=deepseek-chat

Qwen example:

export LINGXIAO_LLM_PROVIDER=openai
export LINGXIAO_OPENAI_API_KEY=sk-your-qwen-key
export LINGXIAO_OPENAI_BASE_URL=https://dashscope.aliyuncs.com/compatible-mode/v1
export LINGXIAO_LEADER_MODEL=qwen-plus
export LINGXIAO_AGENT_MODEL=qwen-turbo

Local LLM Gateway

LingXiao includes a local LLM gateway (/llm/* routes) as a unified endpoint, supporting both OpenAI and Anthropic formats.

Gateway Config

{
  "llm_gateway": {
    "enabled": true,
    "provider": "openai",
    "model": "gpt-4o-mini",
    "inject_env": true,
    "default_rpm": 60,
    "default_tpm": 200000,
    "default_daily_token_budget": 2000000,
    "trace_enabled": true
  }
}

Virtual Keys

The gateway supports virtual key management:

{
  "llm_gateway": {
    "virtual_keys": [
      {
        "id": "key-001",
        "key": "sk-virtual-xxx",
        "label": "Dev Test",
        "enabled": true,
        "model": "gpt-4o-mini",
        "rpm": 30,
        "tpm": 100000,
        "daily_token_budget": 500000
      }
    ]
  }
}

Gateway routes authenticate with virtual keys (Authorization: Bearer <key> or x-api-key), not the server token.

Leader & Worker Model Strategy

Leader model: Responsible for task decomposition and orchestration decisions; recommend a more capable model
Worker model: Responsible for task execution; can use a more cost-effective model
Wiki model: Used for knowledge summaries; can use a lightweight model

{
  "llm": {
    "leader_model": "gpt-4o",
    "agent_model": "gpt-4o-mini",
    "wiki_model": "gpt-4o-mini"
  }
}