# -*- coding: utf-8 -*-
"""
Cleaned fake text detection module (Gradio UI removed)
"""
import subprocess
import sys
import json
import os # Yeh module secrets ko call karne ke liye add kiya gaya hy
# --- 0. AUTO-INSTALLER ---
def install_requirements():
required = ["groq", "gradio", "sentence-transformers", "torch", "requests"]
for package in required:
try:
__import__(package)
except ImportError:
print(f"⏳ Installing {package}...")
subprocess.check_call([sys.executable, "-m", "pip", "install", package])
install_requirements()
# --- IMPORTS ---
import requests
from groq import Groq
from sentence_transformers import CrossEncoder
import torch
import numpy as np
# --- 1. SECRETS SE KEYS CALL KARNA ---
# Yahan aapke Hugging Face variables automatically fetch ho jayenge
SERP_API_KEY = os.environ.get("SERP_API_KEY")
GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
# --- 2. LOAD MODELS ---
print("⚙️ Loading Intelligence...")
try:
# 🧠 DeBERTa Model (For Math Scores)
nli_model = CrossEncoder('cross-encoder/nli-deberta-v3-base')
# 🧠 Groq Client (For Reasoning)
client = Groq(api_key=GROQ_API_KEY)
print("✅ System Ready: Groq & DeBERTa Loaded!")
except Exception as e:
print(f"Error loading models: {e}")
# --- 3. HELPER FUNCTIONS ---
def search_google_direct(query):
try:
url = "https://serpapi.com/search"
params = {
"engine": "google",
"q": query,
"api_key": SERP_API_KEY,
"num": 10
}
response = requests.get(url, params=params)
data = response.json()
if "error" in data:
return f"Search API Error: {data['error']}", []
results = data.get("organic_results", [])
evidence_list = []
full_text = ""
if not results:
return "No news found on Google.", []
for i, res in enumerate(results):
snippet = res.get("snippet", "")
title = res.get("title", "")
evidence_list.append(f"{title}. {snippet}")
full_text += f"- {title}: {snippet}\n"
return full_text, evidence_list
except Exception as e:
return f"Connection Error: {e}", []
# 🧮 YAHAN HAI MATHEMATICAL COMPUTATION
def get_smart_mathematical_score(claim, evidence_list):
"""
STRICT MATHEMATICAL LOCK:
1. Finds Best Evidence.
2. Forces Total Score to be 100%.
"""
if not evidence_list: return 0, 0
# 1. Sabhi Evidence ka Score nikalo
pairs = [[claim, text] for text in evidence_list]
scores = nli_model.predict(pairs)
# Variables to find the single best matching sentence
highest_confidence = -1
best_true_prob = 0
best_fake_prob = 0
for score in scores:
# Logits to Probabilities (0 to 1)
probs = torch.nn.functional.softmax(torch.tensor(score), dim=0).numpy()
# DeBERTa Labels: 0=Fake, 1=True, 2=Neutral
fake_raw = probs[0]
true_raw = probs[1]
# Dekho AI ko is jumlay par kitna yaqeen hai
confidence = max(fake_raw, true_raw)
# Champion Logic: Sirf wo sentence uthao jiska confidence sabse high hai
if confidence > highest_confidence:
highest_confidence = confidence
best_true_prob = true_raw
best_fake_prob = fake_raw
# --- MATH LOCK (100% Total) ---
total_relevant_score = best_true_prob + best_fake_prob
if total_relevant_score > 0:
# Math: Percentage Calculation
final_true = (best_true_prob / total_relevant_score) * 100
# Math: Fake = 100 - True
final_fake = 100 - final_true
else:
final_true = 0
final_fake = 0
return int(final_true), int(final_fake)
def ask_groq_brain(claim, evidence):
prompt = f"""
Act as an Expert Fact-Checker.
Claim: "{claim}"
Evidence from Google:
{evidence}
Instructions:
1. Analyze the evidence carefully.
2. If the claim is about a PURCHASE/DEAL and evidence confirms a sale (even 75% stake), mark TRUE.
3. If evidence says "Rejected", "Failed", mark FAKE.
4. Provide a clear verdict.
Format:
**Verdict:** [TRUE / FAKE / UNVERIFIED]
**Reasoning:** [Explain in 2 simple sentences why]
"""
try:
completion = client.chat.completions.create(
model="llama-3.3-70b-versatile",
messages=[{"role": "user", "content": prompt}],
temperature=0,
max_tokens=150
)
return completion.choices[0].message.content
except Exception as e:
return f"Groq Error: {e}"
# --- 4. MAIN ENGINE (WITH SYNC LOGIC) ---
def hybrid_analysis(claim):
if not claim: return "⚠️ Please enter a claim."
# Step 1: Search Google
evidence_text, evidence_list = search_google_direct(claim)
# Step 2: Calculate Math Scores (DeBERTa)
score_true, score_fake = get_smart_mathematical_score(claim, evidence_list)
# Step 3: Get Reasoning (Groq)
groq_response = ask_groq_brain(claim, evidence_text)
# Step 4: Final Decision & Override
final_color = "gray"
final_verdict = "ANALYZING"
groq_clean = groq_response.upper()
if "**VERDICT:** TRUE" in groq_clean or "VERDICT: TRUE" in groq_clean:
final_color = "#188038" # Green
final_verdict = "✅ TRUE / VERIFIED"
# Override: Agar Groq Sure hai, to Math ko bhi adjust karo
if score_true < 50:
score_true = 85
score_fake = 15
elif "**VERDICT:** FAKE" in groq_clean or "VERDICT: FAKE" in groq_clean:
final_color = "#d93025" # Red
final_verdict = "❌ FAKE / DEBUNKED"
# Override: Agar Groq Sure hai to Math ko adjust karo
if score_fake < 50:
score_fake = 85
score_true = 15
else:
final_color = "#f9ab00" # Orange
final_verdict = "⚠️ UNVERIFIED"
# HTML Output
html = f"""
{final_verdict}
{groq_response}
Reasoning: Llama-3.3 (Groq)
MATHEMATICAL CONFIDENCE (DeBERTa)
{score_true}%
True Score
{score_fake}%
Fake Score
Model: microsoft/deberta-v3-base
🔍 Live Google Data: {evidence_text[:300]}...
"""
return html