Version_3/main.py

import math
import os
import sys
from pydantic import BaseModel, Field

# ── Import your existing cli.py functions directly ───────────────────────────
try:
    from cli import clean_text, predict, load_models as cli_load_models
except ImportError:
    # Fallback to handle relative imports if invoked directly by master gateway
    from Version_3.cli import clean_text, predict, load_models as cli_load_models

# ─────────────────────────────────────────────────────────────────────────────
# 1. REQUIRED SCHEMAS FOR MASTER GATEWAY COMPATIBILITY
# ─────────────────────────────────────────────────────────────────────────────

class EmailScanRequest(BaseModel):
    """
    Must use 'email_body' to match app.js and the validation expected by main.py
    """
    email_body: str = Field(..., description="The raw email text or headers to scan")


class EmailScanResponse(BaseModel):
    """
    Must match the exact structure enforced by response_model in master main.py
    """
    verdict: str
    confidence: float


# ─────────────────────────────────────────────────────────────────────────────
# 2. REQUIRED GLOBAL CONTEXT VARIABLES
# ─────────────────────────────────────────────────────────────────────────────
_vectorizer = None
_model      = None
_load_error = None


def load_models():
    """
    Called automatically by the master gateway bootstrapper.
    """
    global _vectorizer, _model, _load_error
    try:
        _vectorizer, _model = cli_load_models()
        print("[Chimera V3] SVM Models loaded successfully via master handshake.")
        return _vectorizer, _model
    except SystemExit:
        _load_error = "Model files not found. Check models/SVM_model.pkl and models/vectorizer.pkl exist."
        print(f"[Chimera V3] WARNING: {_load_error}")
    except Exception as e:
        _load_error = str(e)
        print(f"[Chimera V3] WARNING: Could not load models — {_load_error}")
    return None, None


# ─────────────────────────────────────────────────────────────────────────────
# 3. CORE ROUTING EXECUTION ENTRYPOINT
# ─────────────────────────────────────────────────────────────────────────────

def scan_email(payload: EmailScanRequest) -> EmailScanResponse:
    """
    Direct functional endpoint executed by the master route handler.
    """
    global _model, _vectorizer, _load_error

    # Lazy loading safe-state fallback for mounted instances
    if _model is None or _vectorizer is None:
        _vectorizer, _model = load_models()

    # If models are genuinely unavailable, fallback gracefully
    if _model is None or _vectorizer is None:
        return EmailScanResponse(
            verdict="Suspicious (Model Offline)",
            confidence=0.50
        )

    text_to_analyze = payload.email_body

    # Guard: empty body text check
    if not text_to_analyze.strip():
        return EmailScanResponse(verdict="Clean", confidence=0.0)

    # ── Run your underlying SVM model prediction ─────────────────────────────
    # Emulating a blank display_name context to feed into your cli pipeline
    combined_text = f"\n{text_to_analyze}"
    label, confidence, spam_score = predict(combined_text, _vectorizer, _model)

    # Calculate normalized confidence float (0.0 to 1.0)
    risk_score = spam_score if spam_score is not None else (100.0 if label == "Spam" else 0.0)
    final_confidence = (confidence / 100.0) if (confidence is not None) else (risk_score / 100.0)

    # ── Map results to match standard frontend badge expectation ──────────────
    if risk_score >= 70:
        assigned_verdict = "phishing"
    elif risk_score >= 40:
        assigned_verdict = "suspicious"
    else:
        assigned_verdict = "clean"

    return EmailScanResponse(
        verdict=assigned_verdict,
        confidence=final_confidence
    )