tools/profanity_guard.py

# tools/profanity_guard.py
from typing import Any, Dict
from smolagents.tools import Tool
import json

class ProfanityGuardTool(Tool):
    name = "profanity_guard"
    description = "Detects profanity in English text and returns a label and confidence."
    inputs: Dict[str, Dict[str, Any]] = {
        "text": {"type": "string", "description": "English text to check for profanity."}
    }
    output_type = "string"  # return JSON string to match your web_search.py pattern

    def __init__(self, model_name: str = "tarekziade/pardonmyai", device: int | None = None, **kwargs: Any) -> None:
        """
        model_name options:
          - "tarekziade/pardonmyai" (default, DistilBERT-based, binary PROFANE/CLEAN)
          - "tarekziade/pardonmyai-tiny" (smaller, faster)
        """
        super().__init__()
        try:
            import torch  # noqa: F401
            from transformers import pipeline  # type: ignore
        except ImportError as e:
            raise ImportError(
                "You must install `transformers` (and optionally `torch`) to use ProfanityGuardTool.\n"
                "Example: pip install transformers torch --extra-index-url https://download.pytorch.org/whl/cu121"
            ) from e

        self.model_name = model_name
        # Pick device automatically if not specified
        try:
            import torch
            if device is None:
                device = 0 if torch.cuda.is_available() else -1
        except Exception:
            device = -1  # CPU fallback if torch not available/working

        # Build the pipeline once (fast subsequent calls)
        from transformers import pipeline
        self.pipe = pipeline(
            task="sentiment-analysis",           # model card uses this task name
            model=self.model_name,
            device=device,
            truncation=True
        )

    def forward(self, text: str) -> str:
        t = (text or "").strip()
        if not t:
            raise ValueError("`text` must be a non-empty string.")

        # Light normalization so profanity isn't split by odd whitespace
        t = " ".join(t.split())

        out = self.pipe(t)[0]  # e.g. {'label': 'PROFANE'|'CLEAN', 'score': 0.xx}

        payload = {
            "model": self.model_name,
            "label": str(out.get("label", "")),
            "score": float(out.get("score", 0.0)),
        }
        return json.dumps(payload)