agent/llm_factory.py

import os
from langchain_openai import ChatOpenAI
from dotenv import load_dotenv

load_dotenv()

_DEFAULT_TIMEOUT = int(os.getenv("LLM_TIMEOUT_SECONDS", "30"))


def get_llm(timeout: int = _DEFAULT_TIMEOUT):
    """
    Returns a configured LLM instance based on environment variables.
    Supports OpenAI (default) and OpenAI-compatible endpoints (DeepInfra, OpenRouter, Groq).

    All providers are given an HTTP timeout to prevent hung LLM calls from
    blocking the FastAPI server indefinitely.
    """
    provider   = os.getenv("LLM_PROVIDER", "OPENAI").upper()
    api_key    = os.getenv("LLM_API_KEY") or os.getenv("OPENAI_API_KEY")
    base_url   = os.getenv("LLM_BASE_URL")
    model_name = os.getenv("LLM_MODEL_NAME")

    common_kwargs = {
        "request_timeout": timeout,
    }

    if provider == "OPENAI":
        return ChatOpenAI(
            model=model_name or "gpt-4-turbo",
            api_key=api_key,
            base_url=base_url,
            **common_kwargs,
        )

    elif provider == "QWEN":
        # Qwen 2.5 Coder via OpenRouter or DeepInfra
        return ChatOpenAI(
            model=model_name or "qwen/qwen-2.5-coder-32b-instruct",
            openai_api_key=api_key,
            openai_api_base=base_url or "https://openrouter.ai/api/v1",
            max_tokens=2048,
            temperature=0.2,
            **common_kwargs,
        )

    else:
        # Generic OpenAI-compatible endpoint
        return ChatOpenAI(
            model=model_name,
            openai_api_key=api_key,
            openai_api_base=base_url,
            **common_kwargs,
        )