interview/llm_router.py

from .models import LLMProviderConfig

# ponytail: wrapper ultra-fino para litellm. Si falla, boom.
def chat_completion(messages):
    config = LLMProviderConfig.objects.filter(is_active=True, is_embedding=False).first()
    if not config:
        raise Exception("No active LLM provider configured.")
        
    if config.provider_type == 'ollama':
        import litellm
        endpoint = config.api_key if config.api_key else "http://127.0.0.1:11434"
        return litellm.completion(
            model=f"ollama/{config.model}",
            messages=messages,
            api_base=endpoint
        )
    elif config.provider_type == 'talentai':
        # Para TalentAI (HuggingFace Space microservicio), usamos formato OpenAI 
        # config.endpoint tiene la URL del Space, config.api_key tiene la clave secreta
        ep = config.endpoint.strip().rstrip("/") if config.endpoint else ""
        import litellm
        return litellm.completion(
            model=f"openai/{config.model}",
            messages=messages,
            api_base=ep,
            api_key=config.api_key if config.api_key else "no-key-needed",
            custom_llm_provider="openai"
        )
    elif config.provider_type == 'openai':
        import litellm
        return litellm.completion(
            model=config.model,
            messages=messages,
            api_key=config.api_key
        )
    elif config.provider_type == 'anthropic':
        import litellm
        return litellm.completion(
            model=config.model,
            messages=messages,
            api_key=config.api_key
        )
    elif config.provider_type == 'nvidia':
        import openai
        client = openai.OpenAI(
            base_url="https://integrate.api.nvidia.com/v1",
            api_key=config.api_key
        )
        return client.chat.completions.create(
            model=config.model,
            messages=messages,
            temperature=1,
            top_p=0.95,
            max_tokens=4000
        )
    else:
        # Fallback genérico para Litellm (que se encarga de rutear según el modelo)
        import litellm
        return litellm.completion(
            model=config.model,
            messages=messages,
            api_key=config.api_key,
            api_base=config.endpoint if config.endpoint else None
        )