model_api.py

import os
from huggingface_hub import InferenceClient


def _get_client() -> InferenceClient:
    return InferenceClient(
        model="Qwen/Qwen2.5-7B-Instruct",
        token=os.getenv("HF_TOKEN")
    )


def generate_weekly_plan(prompt: str) -> str:
    """
    Calls the LLM with the full 7-day blueprint prompt.
    Returns the raw text output which will be split on [NUTRITION_MATRIX_START].
    """
    try:
        client = _get_client()
        response = client.chat_completion(
            messages=[
                {
                    "role": "system",
                    "content": (
                        "You are a certified elite strength & conditioning coach "
                        "and clinical Indian sports nutritionist. "
                        "Produce structured, precise, professional output. "
                        "Never use informal language, filler phrases, or excessive punctuation. "
                        "Follow the exact format requested without deviation."
                    )
                },
                {"role": "user", "content": prompt}
            ],
            max_tokens=3500,
            temperature=0.65
        )
        return response.choices[0].message.content

    except Exception as e:
        return f"[MODEL_ERROR] Unable to connect to inference endpoint: {str(e)}"


def generate_flexspace_patch(prompt: str) -> str:
    """
    Calls the LLM with a targeted FlexSpace alteration prompt.
    Expects a compact, surgical modification output.
    """
    try:
        client = _get_client()
        response = client.chat_completion(
            messages=[
                {
                    "role": "system",
                    "content": (
                        "You are an adaptive fitness coach performing a precision plan modification. "
                        "Output only the modified day section in the original structured format. "
                        "Be surgical — change only what conflicts with the constraint."
                    )
                },
                {"role": "user", "content": prompt}
            ],
            max_tokens=900,
            temperature=0.55
        )
        return response.choices[0].message.content

    except Exception as e:
        return f"[MODEL_ERROR] FlexSpace patch failed: {str(e)}"