app.py

import os
import re
import torch
from fastapi import FastAPI
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# =====================================================
# ✅ Safe Hugging Face Cache Configuration
# =====================================================
CACHE_DIR = "/tmp/hf_cache"
os.environ["HF_HOME"] = CACHE_DIR
os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
os.environ["HF_DATASETS_CACHE"] = CACHE_DIR
os.environ["HF_HUB_CACHE"] = CACHE_DIR
os.makedirs(CACHE_DIR, exist_ok=True)

# =====================================================
# ✅ Load Model and Tokenizer
# =====================================================
MODEL_NAME = "roberta-base-openai-detector"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, cache_dir=CACHE_DIR)

app = FastAPI(title="AI Text Detector")

# =====================================================
# ✅ Input Schema
# =====================================================
class InputText(BaseModel):
    text: str

# =====================================================
# ✅ Helper Functions
# =====================================================
def split_into_paragraphs(text: str):
    """Split text into paragraphs by double newlines or long single breaks."""
    paragraphs = re.split(r'\n\s*\n', text.strip())
    paragraphs = [p.strip() for p in paragraphs if len(p.strip()) > 0]
    return paragraphs

def analyze_text_block(text: str):
    """Analyze a single paragraph and return AI/Human probability."""
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=1)[0].tolist()

    return {
        "label_scores": {
            model.config.id2label[0]: round(probs[0], 4),
            model.config.id2label[1]: round(probs[1], 4)
        },
        "ai_generated_score": probs[1],
        "human_written_score": probs[0],
        "is_ai": probs[1] > probs[0]
    }

# =====================================================
# ✅ Routes
# =====================================================
@app.get("/")
def root():
    return {"message": "AI Text Detector is running. Use POST /analyze with {'text': 'your text'}"}

@app.post("/analyze")
async def analyze(data: InputText):
    text = data.text.strip()
    if not text:
        return {"success": False, "code": 400, "message": "Empty input text"}

    paragraphs = split_into_paragraphs(text)
    results = []

    ai_words, total_words = 0, 0

    for paragraph in paragraphs:
        res = analyze_text_block(paragraph)
        results.append({
            "paragraph": paragraph,
            "ai_generated_score": res["ai_generated_score"],
            "human_written_score": res["human_written_score"]
        })

        word_count = len(paragraph.split())
        total_words += word_count
        ai_words += word_count * res["ai_generated_score"]

    fake_percentage = round((ai_words / total_words) * 100, 2) if total_words > 0 else 0
    feedback = (
        "Most of Your Text is AI/GPT Generated"
        if fake_percentage > 50
        else "Most of Your Text Appears Human-Written"
    )

    return {
        "success": True,
        "code": 200,
        "message": "detection result passed to proxy",
        "data": {
            "sentences": [],
            "isHuman": round(100 - fake_percentage, 2),
            "additional_feedback": "",
            "h": [r["paragraph"] for r in results],
            "hi": [],
            "textWords": total_words,
            "aiWords": int(total_words * (fake_percentage / 100)),
            "fakePercentage": fake_percentage,
            "specialIndexes": [],
            "specialSentences": [],
            "originalParagraph": text,
            "feedback": feedback,
            "input_text": text,
            "detected_language": "en"
        }
    }