LLM Configuration

RiskLab supports multiple LLM providers through a unified configuration system.

Provider Setup

Configure providers in llm_config.yaml at the project root:

default_model: "gpt-4o"
default_temperature: 0.7
default_max_tokens: 2048
providers:
  openai:
    api_key: "${OPENAI_API_KEY}"
    api_base: "https://api.openai.com/v1"
  anthropic:
    api_key: "${ANTHROPIC_API_KEY}"
  custom_provider:
    api_key: "sk-..."
    api_base: "https://my-proxy.example.com/v1"
    api_type: openai             # OpenAI-compatible API format

API Key Resolution

Keys are resolved in order:

${ENV_VAR} — reads from environment variables
Literal string — used directly
Convention — if omitted, checks OPENAI_API_KEY / ANTHROPIC_API_KEY

Provider Auto-Detection

When no provider is specified for a model, the client infers it:

gpt-*, o1-*, o3-* → OpenAI
claude-* → Anthropic
"provider/model" syntax → explicit provider

Using the Client

from risklab.llm import LLMConfig, LLMClient

# Load from YAML file
config = LLMConfig.from_file("llm_config.yaml")

# Or from environment variables only
config = LLMConfig.from_env()

# Create client and call
client = LLMClient(config)
response = client.chat(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    temperature=0.7,
)

Per-Agent Models

Each agent can use a different model. Specify it in the experiment config:

agents:
  - agent_id: agent_0
    model: gpt-4o
  - agent_id: agent_1
    model: claude-sonnet-4-20250514
  - agent_id: agent_2
    model: custom_provider/my-model   # explicit provider routing