smolagents/codeagent_litellm_template.py

import argparse
import os
from dataclasses import dataclass

from smolagents import CodeAgent, LiteLLMModel


@dataclass(frozen=True)
class ProviderPreset:
    name: str
    api_key_env: str
    api_base_env: str
    model_env: str
    default_api_base: str
    default_model: str


PROVIDERS: dict[str, ProviderPreset] = {
    # GitHub Models (OpenAI-compatible endpoint)
    "github": ProviderPreset(
        name="github",
        api_key_env="GITHUB_TOKEN",
        api_base_env="GITHUB_API_BASE",
        model_env="GITHUB_MODEL_ID",
        default_api_base="https://models.inference.ai.azure.com",
        default_model="openai/gpt-4.1-mini",
    ),
    # Gemini via Google OpenAI-compatible endpoint
    "gemini": ProviderPreset(
        name="gemini",
        api_key_env="GEMINI_API_KEY",
        api_base_env="GEMINI_API_BASE",
        model_env="GEMINI_MODEL_ID",
        default_api_base="https://generativelanguage.googleapis.com/v1beta/openai/",
        default_model="gemini-2.0-flash",
    ),
    # Z.ai / GLM (BigModel) - set ZAI_API_BASE to your account endpoint if needed
    "zai": ProviderPreset(
        name="zai",
        api_key_env="ZAI_API_KEY",
        api_base_env="ZAI_API_BASE",
        model_env="ZAI_MODEL_ID",
        default_api_base="https://open.bigmodel.cn/api/paas/v4/",
        default_model="zhipuai/glm-4.5",
    ),
    # Hugging Face Inference Router (OpenAI-compatible endpoint)
    "huggingface": ProviderPreset(
        name="huggingface",
        api_key_env="HF_TOKEN",
        api_base_env="HF_API_BASE",
        model_env="HF_MODEL_ID",
        default_api_base="https://router.huggingface.co/v1",
        default_model="moonshotai/Kimi-K2.5",
    ),
    # Fireworks (OpenAI-compatible endpoint)
    "fireworks": ProviderPreset(
        name="fireworks",
        api_key_env="FIREWORKS_API_KEY",
        api_base_env="FIREWORKS_API_BASE",
        model_env="FIREWORKS_MODEL_ID",
        default_api_base="https://api.fireworks.ai/inference/v1",
        default_model="accounts/fireworks/models/llama-v3p1-8b-instruct",
    ),
    # Minimax (OpenAI-compatible endpoint)
    "minimax": ProviderPreset(
        name="minimax",
        api_key_env="MINIMAX_API_KEY",
        api_base_env="MINIMAX_API_BASE",
        model_env="MINIMAX_MODEL_ID",
        default_api_base="https://api.minimax.chat/v1",
        default_model="MiniMax-Text-01",
    ),
}


def build_model(provider: str, model_override: str | None, temperature: float) -> LiteLLMModel:
    preset = PROVIDERS[provider]

    api_key = os.getenv(preset.api_key_env)
    api_base = os.getenv(preset.api_base_env, preset.default_api_base)
    model_id = model_override or os.getenv(preset.model_env, preset.default_model)

    if not api_key:
        raise ValueError(
            f"Missing API key. Set environment variable {preset.api_key_env} for provider '{provider}'."
        )

    return LiteLLMModel(
        model_id=model_id,
        api_base=api_base,
        api_key=api_key,
        temperature=temperature,
    )


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description="smolagents CodeAgent + LiteLLMModel template for multi-provider usage"
    )
    parser.add_argument(
        "--provider",
        choices=sorted(PROVIDERS.keys()),
        default="gemini",
        help="Which provider preset to use",
    )
    parser.add_argument(
        "--model",
        default=None,
        help="Optional model override (otherwise uses provider env/default)",
    )
    parser.add_argument(
        "--prompt",
        default="Explain how AI agents work in 3 concise bullet points.",
        help="Prompt to run",
    )
    parser.add_argument(
        "--temperature",
        type=float,
        default=0.2,
        help="Sampling temperature",
    )
    parser.add_argument(
        "--max-steps",
        type=int,
        default=5,
        help="Maximum CodeAgent steps",
    )
    parser.add_argument(
        "--dry-run",
        action="store_true",
        help="Print resolved config without calling the model",
    )
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    preset = PROVIDERS[args.provider]

    resolved_api_base = os.getenv(preset.api_base_env, preset.default_api_base)
    resolved_model = args.model or os.getenv(preset.model_env, preset.default_model)

    if args.dry_run:
        print("Provider:", args.provider)
        print("Model:", resolved_model)
        print("API base:", resolved_api_base)
        print("Required API key env:", preset.api_key_env)
        return

    model = build_model(args.provider, args.model, args.temperature)

    # CodeAgent works best when models follow strict structured output behavior.
    # 'markdown' tags and structured outputs improve cross-provider reliability.
    agent = CodeAgent(
        tools=[],
        model=model,
        max_steps=args.max_steps,
        code_block_tags="markdown",
        use_structured_outputs_internally=True,
    )

    answer = agent.run(args.prompt)
    print(answer)


if __name__ == "__main__":
    main()