app/agent/llm.py

from openai import OpenAI

from app.config import get_settings

_client: OpenAI | None = None

DEFAULT_MODEL = "z-ai/glm-5.2"  # swap to "meta/llama-3.1-70b-instruct" if tool-calling isn't confirmed for 5.2


def _get_client() -> OpenAI:
    global _client
    if _client is None:
        settings = get_settings()
        _client = OpenAI(
            api_key=settings.nvidia_api_key,
            base_url="https://integrate.api.nvidia.com/v1",
        )
    return _client


def ask(prompt: str, model: str = DEFAULT_MODEL) -> str:
    """
    Minimal single-turn call to NVIDIA NIM (OpenAI-compatible endpoint).
    No tools, no memory — just proves the API key + connection work end-to-end.
    """
    client = _get_client()

    response = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )

    return response.choices[0].message.content or ""