Update app.py

app.py

@@ -348,20 +348,62 @@ def get_risk_stage(patterns, sentiment):
     return 1
 
 def generate_risk_snippet(abuse_score, top_label, escalation_score, stage):
+    # Extract aggression score from the top label if it's aggression
+    if isinstance(top_label, str) and "aggression" in top_label.lower():
+        try:
+            aggression_score = int(top_label.split("–")[1].replace("%", "").strip()) / 100
+        except:
+            aggression_score = 0
+    else:
+        aggression_score = 0
+
     if abuse_score >= 85 or escalation_score >= 16:
         risk_level = "high"
-    elif abuse_score >= 60 or escalation_score >= 8:
+    elif abuse_score >= 60 or escalation_score >= 8 or aggression_score >= 0.75:
         risk_level = "moderate"
     elif stage == 2 and abuse_score >= 40:
-        risk_level = "moderate"  # 🔧 New rule for escalation stage
+        risk_level = "moderate"
     else:
         risk_level = "low"
+
     if isinstance(top_label, str) and " – " in top_label:
         pattern_label, pattern_score = top_label.split(" – ")
     else:
         pattern_label = str(top_label) if top_label is not None else "Unknown"
         pattern_score = ""
 
+    WHY_FLAGGED = {
+        "control": "This message may reflect efforts to restrict someone’s autonomy, even if it's framed as concern or care.",
+        "gaslighting": "This message could be manipulating someone into questioning their perception or feelings.",
+        "dismissiveness": "This message may include belittling, invalidating, or ignoring the other person’s experience.",
+        "insults": "Direct insults often appear in escalating abusive dynamics and can erode emotional safety.",
+        "threat": "This message includes threatening language, which is a strong predictor of harm.",
+        "blame shifting": "This message may redirect responsibility to avoid accountability, especially during conflict.",
+        "guilt tripping": "This message may induce guilt in order to control or manipulate behavior.",
+        "recovery phase": "This message may be part of a tension-reset cycle, appearing kind but avoiding change.",
+        "projection": "This message may involve attributing the abuser’s own behaviors to the victim.",
+        "contradictory statements": "This message may contain internal contradictions used to confuse, destabilize, or deflect responsibility.",
+        "obscure language": "This message may use overly formal, vague, or complex language to obscure meaning or avoid accountability.",
+        "default": "This message contains language patterns that may affect safety, clarity, or emotional autonomy."
+    }
+
+    explanation = WHY_FLAGGED.get(pattern_label.lower(), WHY_FLAGGED["default"])
+
+    base = f"\n\n🛑 Risk Level: {risk_level.capitalize()}\n"
+    base += f"This message shows strong indicators of **{pattern_label}**. "
+
+    if risk_level == "high":
+        base += "The language may reflect patterns of emotional control, even when expressed in soft or caring terms.\n"
+    elif risk_level == "moderate":
+        base += "There are signs of emotional pressure or verbal aggression that may escalate if repeated.\n"
+    else:
+        base += "The message does not strongly indicate abuse, but it's important to monitor for patterns.\n"
+
+    base += f"\n💡 *Why this might be flagged:*\n{explanation}\n"
+    base += f"\nDetected Pattern: **{pattern_label} ({pattern_score})**\n"
+    base += "🧠 You can review the pattern in context. This tool highlights possible dynamics—not judgments."
+    return base
+
     WHY_FLAGGED = {
         "control": "This message may reflect efforts to restrict someone’s autonomy, even if it's framed as concern or care.",
         "gaslighting": "This message could be manipulating someone into questioning their perception or feelings.",