app.py

import gradio as gr
import pandas as pd
import re
from sklearn.ensemble import IsolationForest
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F

# ✅ Load Google-backed BERT model manually (no pipeline)
model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

def get_llm_score(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = F.softmax(outputs.logits, dim=1)
    return float(probs[0][0])  # Use first class probability as a proxy score

def detect_anomalies(log_text):
    lines = log_text.strip().split("\n")
    df = pd.DataFrame({"log": lines})
    df["length"] = df["log"].apply(len)
    df["digits"] = df["log"].apply(lambda x: sum(c.isdigit() for c in x))
    df["specials"] = df["log"].apply(lambda x: sum(not c.isalnum() for c in x))
    df["llm_score"] = df["log"].apply(get_llm_score)

    features = df[["length", "digits", "specials", "llm_score"]].fillna(0)
    model_iso = IsolationForest(contamination=0.1, random_state=42)
    preds = model_iso.fit_predict(features)
    df["anomaly"] = preds
    df["status"] = df["anomaly"].map({1: "Normal", -1: "Anomaly"})

    return df[["log", "status"]]

demo = gr.Interface(
    fn=detect_anomalies,
    inputs=gr.Textbox(lines=20, placeholder="Paste logs here..."),
    outputs=gr.Dataframe(label="Log Status", type="pandas"),
    title="🧠 Log Anomaly Detection (Google BERT)",
    description="Detect anomalies using Isolation Forest + Google-backed BERT. Fast, accurate, and deploy-safe."
)

demo.launch()