src/infrastructure/adapters/huggingface_adapter.py

import json
import logging
from typing import AsyncGenerator

from src.core.ports.llm_port import LlmPort
from src.core.config import settings

logger = logging.getLogger(__name__)

# HF Inference Providers router — OpenAI-compatible
# Only models with chat/instruct fine-tuning work on the chat completions endpoint
_ROUTER_URL = "https://router.huggingface.co/v1"

# Chat-compatible instruct models available on HF router (free tier)
_DEFAULT_MODEL = "meta-llama/Llama-3.1-8B-Instruct"
_FALLBACK_MODELS = [
    "meta-llama/Llama-3.1-8B-Instruct",
    "mistralai/Mistral-Nemo-Instruct-2407",
    "Qwen/Qwen2.5-7B-Instruct",
    "microsoft/Phi-3.5-mini-instruct",
    "HuggingFaceH4/zephyr-7b-beta",
]


class HuggingFaceAdapter(LlmPort):
    """
    HuggingFace Inference Providers adapter — free tier with HF token.
    Uses the new router.huggingface.co OpenAI-compatible endpoint.

    Get token: https://huggingface.co/settings/tokens
      → New token → Fine-grained → enable "Make calls to Inference Providers"
    Set HF_TOKEN env var. Optionally set HF_MODEL to choose a specific model.

    Free limits: generous free tier, no credit card needed.
    """

    def __init__(self):
        self.token = settings.HF_TOKEN
        self.model = (settings.HF_MODEL or _DEFAULT_MODEL).strip()
        self.llm = None

        if self.token and self.token != "your-hf-token-here":
            try:
                from langchain_openai import ChatOpenAI
                self.llm = ChatOpenAI(
                    api_key=self.token,
                    base_url=_ROUTER_URL,
                    model=f"{self.model}:fastest",   # :fastest = auto-select best provider
                    temperature=0.2,
                    max_tokens=1024,
                )
                logger.info(f"✅ HuggingFace adapter ready — model: {self.model}:fastest")
            except Exception as e:
                logger.error(f"Failed to initialize HuggingFace adapter: {e}")
        else:
            logger.warning("HF_TOKEN not set — HuggingFace adapter disabled.")

    def generate(self, prompt: str) -> str:
        if not self.llm:
            return "HuggingFace token not configured."
        try:
            return self.llm.invoke(prompt).content
        except Exception as e:
            error_msg = str(e)
            if "429" in error_msg or "rate" in error_msg.lower():
                return "HuggingFace rate limit reached. Please try again shortly."
            logger.error(f"HuggingFace generate error: {e}")
            return f"HuggingFace error: {error_msg}"

    async def generate_stream(self, prompt: str) -> AsyncGenerator[str, None]:
        if not self.llm:
            yield f"data: {json.dumps({'token': 'HuggingFace token not configured.'})}\n\n"
            yield "data: [DONE]\n\n"
            return
        try:
            for chunk in self.llm.stream(prompt):
                if hasattr(chunk, "content") and chunk.content:
                    yield f"data: {json.dumps({'token': chunk.content})}\n\n"
            yield "data: [DONE]\n\n"
        except Exception as e:
            error_msg = str(e)
            msg = "HuggingFace rate limit reached." if "429" in error_msg else f"HuggingFace error: {error_msg}"
            yield f"data: {json.dumps({'token': msg})}\n\n"
            yield "data: [DONE]\n\n"