services/llm.py

import logging
import os
from functools import lru_cache
from typing import Any, List, Optional

from huggingface_hub import InferenceClient
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
from langchain_core.outputs import ChatGeneration, ChatResult
from pydantic import Field

logger = logging.getLogger(__name__)

MODEL_ID = "allenai/Olmo-3-7B-Instruct"
FALLBACK_MODEL_ID = "HuggingFaceH4/zephyr-7b-beta"


def get_hf_token() -> str:
    token = (
        os.environ.get("HF_TOKEN")
        or os.environ.get("HUGGINGFACEHUB_API_TOKEN")
        or os.environ.get("HUGGING_FACE_HUB_TOKEN")
    )
    if not token:
        raise ValueError(
            "HF_TOKEN is not set. Add it as a Space secret (Settings → Repository secrets)."
        )
    # LangChain / huggingface_hub also read this name
    os.environ.setdefault("HUGGINGFACEHUB_API_TOKEN", token)
    return token


@lru_cache(maxsize=1)
def get_inference_client() -> InferenceClient:
    return InferenceClient(api_key=get_hf_token())


class HuggingFaceInferenceChat(BaseChatModel):
    """LangChain chat model using Hugging Face Inference API chat.completions."""

    model_id: str = Field(default=MODEL_ID)
    max_tokens: int = 512
    temperature: float = 0.2

    @property
    def _llm_type(self) -> str:
        return "huggingface-inference-chat"

    def _to_hf_messages(self, messages: List[BaseMessage]) -> list[dict[str, str]]:
        hf_messages: list[dict[str, str]] = []
        for msg in messages:
            if isinstance(msg, SystemMessage):
                hf_messages.append({"role": "system", "content": str(msg.content)})
            elif isinstance(msg, HumanMessage):
                hf_messages.append({"role": "user", "content": str(msg.content)})
            elif isinstance(msg, AIMessage):
                hf_messages.append({"role": "assistant", "content": str(msg.content)})
        return hf_messages

    def _generate(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Any = None,
        **kwargs: Any,
    ) -> ChatResult:
        client = get_inference_client()
        response = client.chat.completions.create(
            model=self.model_id,
            messages=self._to_hf_messages(messages),
            max_tokens=self.max_tokens,
            temperature=self.temperature,
        )
        if not response.choices:
            raise RuntimeError(f"No choices returned for model {self.model_id}")
        content = response.choices[0].message.content or ""
        return ChatResult(
            generations=[ChatGeneration(message=AIMessage(content=content))]
        )


_llm: HuggingFaceInferenceChat | None = None


def get_llm(model_id: str = MODEL_ID) -> HuggingFaceInferenceChat:
    global _llm
    if _llm is None or _llm.model_id != model_id:
        get_hf_token()
        _llm = HuggingFaceInferenceChat(model_id=model_id)
    return _llm


def invoke_chat(messages: List[BaseMessage], model_id: str = MODEL_ID) -> str:
    """Call primary model, then fallback if the provider rejects the request."""
    last_error: Exception | None = None
    for mid in (model_id, FALLBACK_MODEL_ID):
        try:
            llm = get_llm(mid)
            result = llm.invoke(messages)
            text = result.content if isinstance(result.content, str) else str(result.content)
            if text.strip():
                return text.strip()
        except Exception as exc:
            last_error = exc
            logger.warning("HF chat failed for model %s: %s", mid, exc)
            global _llm
            _llm = None
    raise RuntimeError(str(last_error) if last_error else "Unknown inference error")


if __name__ == "__main__":
    logging.basicConfig(level=logging.INFO)
    out = invoke_chat(
        [
            SystemMessage(content="You are an ATS resume analyst."),
            HumanMessage(
                content=(
                    "ATS Scores: Semantic 0.45, Keyword 0.70, Final 0.68. "
                    "Missing: docker, tensorflow. Skill overlap: 70%. "
                    "Write 3 short sections: Score Explanation, Weak Areas, Actionable Improvements."
                )
            ),
        ]
    )
    print(out)