update

honeypot_api.py

@@ -149,9 +149,9 @@ def generate_agent_reply(history):
         "Keep replies to 1–2 sentences and end with a question.\n\n"
     )
 
-    convo=""
+    convo = ""
     for h in history[-6:]:
-        convo+=f"{h['sender']}: {h['text']}\n"
+        convo += f"{h['sender']}: {h['text']}\n"
 
     prompt = persona + convo + "user:"
 
@@ -167,7 +167,43 @@ def generate_agent_reply(history):
         )
 
     txt = agent_tokenizer.decode(out[0], skip_special_tokens=True)
-    return txt.split("user:")[-1].strip()
+    reply = txt.split("user:")[-1].strip()
+    reply = reply.replace("\n", " ").strip()
+
+    if "password is not your default email" in reply.lower():
+        reply = ""
+
+    sentences = [s.strip() for s in re.split(r"(?<=[.!?])\s+", reply) if s.strip()]
+    unique_sentences = []
+    seen = set()
+    for s in sentences:
+        key = s.lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        unique_sentences.append(s)
+        if len(unique_sentences) >= 2:
+            break
+
+    if not unique_sentences:
+        fallbacks = [
+            "I am a bit confused about this. What exactly do you want me to do?",
+            "I am worried about my account. What should I do now?",
+            "I am not sure I understand. Can you explain what I need to share?"
+        ]
+        idx = len(history) % len(fallbacks)
+        reply = fallbacks[idx]
+    else:
+        reply = " ".join(unique_sentences)
+
+    for forbidden in ["scam", "fraud", "phishing", "honeypot"]:
+        reply = re.sub(forbidden, "", reply, flags=re.IGNORECASE)
+    reply = " ".join(reply.split())
+
+    if "?" not in reply:
+        reply = reply.rstrip(".!") + "?"
+
+    return reply
 
 def compute_engagement_score(session_id, last_agent_reply):
     conv = conversation_store.get(session_id, [])