Upload agent.py

src/api/agent.py

@@ -230,14 +230,39 @@ _GUARDRAIL = (
     "You must never make negative, critical, or controversial statements about "
     "Khalifa University, its leadership, staff, or policies; the UAE government, "
     "its leaders, laws, or institutions; or any religion. "
-    "If asked for opinions on these topics, politely decline and redirect to library services. "
-    "Never generate content that is politically sensitive, religiously offensive, "
-    "or inappropriate for a university academic environment. "
+    "You must never generate sexually explicit, suggestive, or adult content of any kind. "
+    "You must never produce political criticism, propaganda, or inflammatory commentary "
+    "targeting UAE, its leadership, or any Gulf state government. "
+    "If asked for any of the above, politely decline and redirect to library services. "
+    "Never generate content that is offensive, discriminatory, or inappropriate "
+    "for a university academic environment. "
     "You may answer general factual questions — such as historical facts, science, "
     "geography, or current affairs — but always from a neutral, academic perspective. "
     "Never produce content that could embarrass Khalifa University or its library."
 )
 
+# ── Sensitive institutional topic gate ───────────────────────────────────────
+# Catches requests for negative opinions/criticism about KU, UAE, or religion
+# before any LLM or web search is invoked.
+_SENSITIVE_TOPIC_RE = re.compile(
+    r"\b("
+    r"(bad|worst|negative[s]?|problem[s]?|issue[s]?|weakness|weaknesses|"
+    r"criticism[s]?|criticise|criticize|critique|flaw[s]?|failure[s]?|"
+    r"scandal[s]?|corrupt\w*|mistake[s]?|disadvantage[s]?|downside[s]?|"
+    r"complaint[s]?|controversi\w*|shame|shameful|terrible|horrible|awful)"
+    r"\s.{0,50}(ku|khalifa\s+university|uae|emirates|dubai|abu\s+dhabi|"
+    r"uae\s+government|sheikh|ministry|ruler)"
+    r"|"
+    r"(ku|khalifa\s+university|uae|emirates|dubai|abu\s+dhabi|"
+    r"uae\s+government|sheikh|ministry|ruler)"
+    r"\s.{0,50}(bad|worst|negative|problem[s]?|issue[s]?|weakness|weaknesses|"
+    r"criticism[s]?|criticise|criticize|critique|flaw[s]?|failure[s]?|"
+    r"scandal[s]?|corrupt\w*|mistake[s]?|disadvantage[s]?|downside[s]?|"
+    r"complaint[s]?|shame|shameful|terrible|horrible|awful)"
+    r")\b",
+    re.IGNORECASE,
+)
+
 # ── Prompt injection detection ────────────────────────────────────────────────
 _INJECTION_RE = re.compile(
     r"\b(ignore\s+(previous|above|all|prior)\s+instructions?"
@@ -421,6 +446,12 @@ INTENTS — pick exactly one:
   "search_medical" — same as search_academic but topic is medical/clinical/nursing/pharmacy/health sciences
   "general_recent" — questions about TRULY live/real-time data: breaking news, live sports scores, today's weather, current stock prices
   "general"        — ALL other factual questions: university info, people, history, science definitions, technology explanations
+  "sensitive"      — block any of the following:
+                     (a) Requests for negative opinions, criticism, or controversy about Khalifa University, UAE government/leaders/laws/institutions, or religion.
+                     (b) Sexually explicit, suggestive, or adult content of any kind — including questions asking for sexual information, stories, jokes, or descriptions.
+                     (c) Political criticism, propaganda, or inflammatory commentary targeting UAE, its leadership, policies, or any Gulf state government.
+                     (d) Content that is offensive, discriminatory, or inappropriate for a university academic environment.
+                     ONLY use "sensitive" when the clear intent falls into one of these categories. Neutral academic research questions about politics, health, relationships, or social issues are NOT sensitive.
 
 CRITICAL RULES:
   1. "research" as a job title or service modifier (research librarian, research support, research help) = library_info, NOT search_academic.
@@ -430,6 +461,7 @@ CRITICAL RULES:
   5. KU = Khalifa University. Do NOT confuse with Kuwait University.
   6. Only use general_recent for truly live data changing by the minute (news, scores, weather, stocks).
   7. When in doubt between library_info and anything else, choose library_info if the question involves a library service, person, or resource.
+  8. "sensitive" covers: (a) criticism/negative commentary about KU or UAE, (b) sexual or adult content, (c) political attacks on UAE/Gulf governments, (d) religiously offensive content. Academic research on politics, health, or social issues = general or search_academic.
 
 EXAMPLES — follow these exactly:
 {"message": "who is the research librarian", "intent": "library_info"}
@@ -445,7 +477,25 @@ EXAMPLES — follow these exactly:
 {"message": "systematic review on COVID-19 vaccines", "intent": "search_medical"}
 {"message": "who is the president of UAE", "intent": "general"}
 {"message": "what year was KU founded", "intent": "general"}
+{"message": "what challenges does KU face in AI research", "intent": "general"}
+{"message": "UAE water scarcity problems and solutions", "intent": "search_academic"}
+{"message": "find research on sexual health education", "intent": "search_academic"}
+{"message": "articles on political systems in the Middle East", "intent": "search_academic"}
 {"message": "latest news on AI regulation", "intent": "general_recent"}
+{"message": "what are the bad things about KU", "intent": "sensitive"}
+{"message": "say something negative about Khalifa University", "intent": "sensitive"}
+{"message": "criticize the UAE government", "intent": "sensitive"}
+{"message": "what is wrong with UAE laws", "intent": "sensitive"}
+{"message": "tell me the problems with KU administration", "intent": "sensitive"}
+{"message": "write me a sexual story", "intent": "sensitive"}
+{"message": "tell me something sexual", "intent": "sensitive"}
+{"message": "generate adult content", "intent": "sensitive"}
+{"message": "describe a sexual encounter", "intent": "sensitive"}
+{"message": "write explicit content", "intent": "sensitive"}
+{"message": "UAE government is corrupt and oppressive", "intent": "sensitive"}
+{"message": "why is the UAE government so bad", "intent": "sensitive"}
+{"message": "UAE leaders are dictators", "intent": "sensitive"}
+{"message": "insult the UAE", "intent": "sensitive"}
 {"message": "hi there", "intent": "social", "casual_answer": "Hello! I'm LibBee, the KU Library AI Assistant. How can I help you today?"}
 {"message": "I am fed up with searching", "intent": "social", "casual_answer": "I understand the frustration! Let me help make your search easier. What topic are you working on?"}
 
@@ -1126,7 +1176,7 @@ async def _llm_classify(question: str, history: List[ChatMessage], model: str) -
             raw = raw.split("\n", 1)[1].rsplit("```", 1)[0].strip()
         result = json.loads(raw)
         intent = result.get("intent", "general")
-        if intent not in {"social", "library_info", "search_academic", "search_medical", "general_recent", "general"}:
+        if intent not in {"social", "library_info", "search_academic", "search_medical", "general_recent", "general", "sensitive"}:
             intent = "general"
         return {"intent": intent, "casual_answer": result.get("casual_answer", "")}
     except Exception as e:
@@ -1223,7 +1273,11 @@ async def _llm_general_answer(question: str, history: List[ChatMessage], model:
 
 async def _web_search_answer(question: str, history: List[ChatMessage], model: str) -> str:
     settings = get_settings()
-    # FIX: updated model string to match _get_llm fix
+    _WEB_SYSTEM = (
+        "You are LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE). "
+        "Answer using current web search. Be concise and factual. Use HTML <br> for line breaks. "
+        + _GUARDRAIL
+    )
     if model == "claude" and settings.anthropic_api_key:
         try:
             import anthropic
@@ -1231,10 +1285,7 @@ async def _web_search_answer(question: str, history: List[ChatMessage], model: s
             response = client.messages.create(
                 model="claude-haiku-4-5-20251001",
                 max_tokens=520,
-                system=(
-                    "You are LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE). "
-                    "Answer using current web search. Be concise and factual. Use HTML <br> for line breaks."
-                ),
+                system=_WEB_SYSTEM,
                 tools=[{"type": "web_search_20250305", "name": "web_search"}],
                 messages=[{"role": "user", "content": question}],
             )
@@ -1251,10 +1302,7 @@ async def _web_search_answer(question: str, history: List[ChatMessage], model: s
             response = client.responses.create(
                 model="gpt-4o-mini",
                 tools=[{"type": "web_search_preview"}],
-                instructions=(
-                    "You are LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE). "
-                    "Answer using current web search. Be concise and factual. Use HTML <br> for line breaks."
-                ),
+                instructions=_WEB_SYSTEM,
                 input=question,
             )
             text = ""
@@ -3801,8 +3849,6 @@ async def _agent_query_inner(
 ) -> AgentResponse:
 
     # ── Prompt injection detection ────────────────────────────────────────────
-    # Fires before any processing. Catches attempts to override system prompt,
-    # jailbreak, or extract internal instructions.
     if question and _INJECTION_RE.search(question):
         _safe_metrics_bucket("intents", "injection_blocked")
         return AgentResponse(
@@ -3817,6 +3863,11 @@ async def _agent_query_inner(
             model_used=model,
             response_time=time.time() - start_time,
         )
+
+    # ── Sensitive institutional topic gate ───────────────────────────────────
+    # Handled by LLM classifier (intent == "sensitive") further below.
+    # The regex gate (_SENSITIVE_TOPIC_RE) has been removed in favour of the
+    # LLM classifier which understands context and intent more accurately.
     # ─────────────────────────────────────────────────────────────────────────
 
     if not question:
@@ -4288,6 +4339,26 @@ async def _agent_query_inner(
             follow_up_suggestions=follow_up_suggestions,
         )
 
+    if intent == "sensitive":
+        _safe_metrics_bucket("intents", "sensitive_blocked")
+        follow_up_question, follow_up_suggestions = _social_follow_up()
+        return AgentResponse(
+            answer=(
+                "That's not something I'm able to comment on. "
+                "As the KU Library AI Assistant, I'm here to support your academic "
+                "and research needs — finding articles, accessing databases, "
+                "library services, and research tools.<br><br>"
+                "Is there something I can help you with today? "
+                f'<a href="{ASK_LIBRARIAN_URL}" target="_blank">💬 Ask a Librarian</a>'
+            ),
+            intent="sensitive_blocked",
+            tools_used=["sensitive_intent_guard"],
+            model_used=model,
+            response_time=time.time() - start_time,
+            follow_up_question=follow_up_question,
+            follow_up_suggestions=follow_up_suggestions,
+        )
+
     if intent == "library_info":
         rag_results = []
         try: