app.py

import gradio as gr
import requests
import os

HF_TOKEN = os.getenv("HF_TOKEN")

API_URL = "https://router.huggingface.co/hf-inference/models/joeddav/xlm-roberta-large-xnli"


def classify(payload):
    try:
        # Gradio JSON input bazen list içinde gelir
        if isinstance(payload, list):
            payload = payload[0]

        invoice_text = payload.get("invoice_text", "")
        categories = payload.get("categories", [])

        if not invoice_text or not categories:
            return "NO_INPUT"

        # Metni kırp ama Türkçe karakterleri bozmadan
        invoice_text = invoice_text.strip()[:3000]

        candidate_labels = [c["cat_name"].strip() for c in categories]

        if not candidate_labels:
            return "NO_CATEGORIES"

        response = requests.post(
            API_URL,
            headers={
                "Authorization": f"Bearer {HF_TOKEN}",
                "Content-Type": "application/json"
            },
            json={
                "inputs": invoice_text,
                "parameters": {
                    "candidate_labels": candidate_labels,
                    "multi_label": False
                }
            },
            timeout=120
        )

        print("STATUS:", response.status_code)
        print("RESPONSE:", response.text[:500])

        if response.status_code != 200:
            return f"HTTP_ERROR_{response.status_code}"

        result = response.json()

        # xlm-roberta ve deberta her ikisi de dict döner:
        # {"sequence": "...", "labels": [...], "scores": [...]}
        # Ama HF bazen list of dict döner, ikisini de handle ediyoruz.

        if isinstance(result, list):
            # list of {"label": ..., "score": ...} formatı
            if not result:
                return "NO_RESULT"
            result.sort(key=lambda x: x.get("score", 0), reverse=True)
            return result[0].get("label", "NO_LABEL")

        elif isinstance(result, dict):
            labels = result.get("labels", [])
            scores = result.get("scores", [])

            if not labels:
                return "NO_LABEL"

            # En yüksek skorlu kategoriyi direkt döndür
            # (random veya penalty yok — model kararına güven)
            best_index = scores.index(max(scores)) if scores else 0
            return labels[best_index]

        else:
            return "UNEXPECTED_RESPONSE"

    except requests.exceptions.Timeout:
        return "TIMEOUT"
    except Exception as e:
        import traceback
        traceback.print_exc()
        return f"ERROR: {str(e)}"


gr.Interface(
    fn=classify,
    inputs="json",
    outputs="text"
).launch(ssr_mode=False)