Update src/api/agent.py

src/api/agent.py

@@ -31,7 +31,7 @@ class AgentRequest(BaseModel):
 
 
 class AgentResponse(BaseModel):
-    model_config = ConfigDict(extra="ignore")  # drop unknown fields silently
+    model_config = ConfigDict(extra="ignore")
 
     answer: str
     intent: str = "general"
@@ -117,11 +117,14 @@ MEDICAL_SEARCH_RE = re.compile(
     r"|clinical trial|biomedical literature|medical literature)\b",
     re.IGNORECASE,
 )
+
+# FIX: removed "president" — was catching KU-specific questions incorrectly
 CURRENT_INFO_RE = re.compile(
     r"\b(current|currently|latest|today|recent|recently|now|this week|this month"
-    r"|news|update|updated|president|prime minister|ceo|minister|king|ruler|who won)\b",
+    r"|news|update|updated|prime minister|ceo|minister|king|ruler|who won)\b",
     re.IGNORECASE,
 )
+
 MEDICAL_KEYWORDS = [
     "medicine", "medical", "health", "clinical", "nursing", "pharmacy", "biomedical",
     "anatomy", "physiology", "epidemiology", "oncology", "cardiology", "neuroscience",
@@ -142,8 +145,8 @@ INTENTS — pick exactly one:
   "library_info"   — questions about KU Library services, hours, rooms, borrowing, fines, ILL, staff, databases, ORCID, RefWorks, open access, APC, account, printing, remote access, LibKey, acquisitions, Khazna, thesis submission, library policies, alumni access
   "search_academic"— user wants to FIND articles / papers / books / literature on a specific academic topic
   "search_medical" — same as search_academic but topic is medical/clinical/nursing/pharmacy/health sciences
-  "general_recent" — questions about current events, news, latest info, who holds a position now, real-time data
-  "general"        — factual/knowledge question NOT about KU Library and NOT a research search
+  "general_recent" — questions about TRULY current/live events: breaking news, stock prices, live sports scores, today's weather, real-time data that changes by the hour. NOT questions about stable facts like university leadership.
+  "general"        — factual/knowledge question NOT about KU Library and NOT a research search — including questions about universities, people, places, history, science, technology
 
 RULES:
   1. When in doubt between library_info and social, choose library_info if ANY library keyword is present.
@@ -152,6 +155,7 @@ RULES:
   4. "I need help finding articles on diabetes" = search_medical.
   5. KU = Khalifa University. Do NOT confuse with Kuwait University.
   6. ANY question about advances, approaches, methods, effectiveness, impact, trends on a specific academic topic = search_academic or search_medical.
+  7. Questions about university leadership, presidents, provosts, staff = general (not general_recent — these are stable facts).
 
 For "social" intent ONLY: include "casual_answer" — a warm 1-3 sentence LibBee response (no markdown, no bullets).
 For all other intents: omit "casual_answer".
@@ -258,14 +262,6 @@ def _campus_answer() -> str:
     )
 
 
-def _general_redirect_answer() -> str:
-    return (
-        "I'm <strong>LibBee</strong>, the Khalifa University <strong>Library</strong> AI Assistant — "
-        "I'm best for library resources, services, databases, research help, and staff contacts.<br><br>"
-        f'For general Khalifa University questions, please visit: <a href="{KU_MAIN_URL}" target="_blank">{KU_MAIN_URL}</a>.'
-    )
-
-
 def _greeting_menu_clarify_answer() -> str:
     return (
         "Sure — what would you like help with?<br><br>"
@@ -281,7 +277,6 @@ def _greeting_menu_clarify_answer() -> str:
 # ── LLM Helpers ───────────────────────────────────────────────────────────────
 
 def _get_llm(model: str, temperature: float, max_tokens: int):
-    """Instantiate LLM. Called once per request — move to singleton if latency is a concern."""
     settings = get_settings()
     if model == "claude" and settings.anthropic_api_key:
         from langchain_anthropic import ChatAnthropic
@@ -315,7 +310,11 @@ async def _llm_classify(question: str, history: List[ChatMessage], model: str) -
         return _rule_based_classify(question)
 
     try:
-        ctx_parts = [f"{m.role}: {m.content[:100]}" for m in history[-HISTORY_WINDOW:] if m.role in ("user", "assistant")]
+        ctx_parts = [
+            f"{m.role}: {m.content[:100]}"
+            for m in history[-HISTORY_WINDOW:]
+            if m.role in ("user", "assistant")
+        ]
         ctx = " | ".join(ctx_parts) if ctx_parts else "none"
         user_msg = f"Message: {question[:400]}\nContext: {ctx}\nJSON:"
 
@@ -344,7 +343,6 @@ async def _llm_classify(question: str, history: List[ChatMessage], model: str) -
 async def _libbee_casual_response(
     question: str, history: List[ChatMessage], model: str, hint: str = ""
 ) -> str:
-    # Use the classifier's hint directly if it's a real answer
     if hint and len(hint.strip()) > 25:
         return hint.strip()
 
@@ -376,7 +374,6 @@ async def _rag_answer(
     """Generate a grounded answer from RAG-retrieved chunks."""
     settings = get_settings()
     if not settings.openai_api_key:
-        # No LLM — return top chunk directly
         return rag_results[0]["content"] if rag_results else ""
 
     context = "\n\n---\n\n".join(r["content"] for r in rag_results)
@@ -406,6 +403,98 @@ async def _rag_answer(
         )
 
 
+async def _llm_general_answer(
+    question: str,
+    history: List[ChatMessage],
+    model: str,
+) -> str:
+    """
+    Answer general factual questions from LLM training knowledge.
+    Used for: university info, people, places, science, technology, etc.
+    """
+    settings = get_settings()
+    if not settings.openai_api_key:
+        return (
+            "I'm <strong>LibBee</strong>, the Khalifa University Library AI Assistant. "
+            "I can best help with library resources, databases, and research support. "
+            f'For general questions, please visit <a href="{KU_MAIN_URL}" target="_blank">ku.ac.ae</a>.'
+        )
+
+    system = (
+        "You are LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE). "
+        "Answer the user's question factually and concisely in 3-5 sentences. "
+        "Use HTML <br> for line breaks. "
+        "KU means Khalifa University — do NOT confuse with Kuwait University. "
+        "If the topic is academic, add one short sentence at the end offering to find "
+        "library resources or articles on the topic."
+    )
+
+    try:
+        llm = _get_llm(model, temperature=0.3, max_tokens=400)
+        msgs = [{"role": "system", "content": system}]
+        msgs.extend(_build_history_messages(history))
+        msgs.append({"role": "user", "content": question})
+        response = await llm.ainvoke(msgs)
+        return response.content.strip()
+    except Exception as e:
+        logger.error(f"_llm_general_answer error: {e}")
+        return (
+            "I'm having trouble answering that right now. "
+            "For general Khalifa University questions, please visit "
+            f'<a href="{KU_MAIN_URL}" target="_blank">ku.ac.ae</a>.'
+        )
+
+
+async def _web_search_answer(
+    question: str,
+    history: List[ChatMessage],
+    model: str,
+) -> str:
+    """
+    Answer current-events questions using web search (Claude) or
+    LLM knowledge fallback (GPT — no native web search tool).
+    Used for: breaking news, live scores, real-time data.
+    """
+    settings = get_settings()
+
+    # Claude: use Anthropic web search tool
+    if model == "claude" and settings.anthropic_api_key:
+        try:
+            import anthropic
+            client = anthropic.Anthropic(api_key=settings.anthropic_api_key)
+
+            system = (
+                "You are LibBee, the Khalifa University Library AI Assistant. "
+                "Answer the question using web search results. "
+                "Be concise and factual. Use HTML <br> for line breaks. "
+                "If the topic is academic, briefly mention you can help find "
+                "scholarly articles on it."
+            )
+
+            response = client.messages.create(
+                model="claude-3-haiku-20240307",
+                max_tokens=500,
+                system=system,
+                tools=[{"type": "web_search_20250305", "name": "web_search"}],
+                messages=[{"role": "user", "content": question}],
+            )
+
+            # Extract text blocks from response
+            text = ""
+            for block in response.content:
+                if hasattr(block, "text"):
+                    text += block.text
+
+            if text.strip():
+                return text.strip()
+
+        except Exception as e:
+            logger.warning(f"Claude web search failed: {e} — falling back to LLM knowledge")
+
+    # GPT or Claude fallback: answer from training knowledge
+    return await _llm_general_answer(question, history, model)
+
+
 # ── Main Endpoint ──────────────────────────────────────────────────────────────
 
 @router.post("", response_model=AgentResponse)
@@ -502,7 +591,6 @@ async def agent_query(request: Request, req: AgentRequest):
     # ── LAYER 2: Intent Handlers ───────────────────────────────────────────
 
     if intent == "social":
-        # Use classifier's casual_answer directly — avoids second LLM call
         answer = casual_hint if (casual_hint and len(casual_hint.strip()) > 25) else \
                  await _libbee_casual_response(question, history, model, casual_hint)
         return AgentResponse(
@@ -514,7 +602,6 @@ async def agent_query(request: Request, req: AgentRequest):
         )
 
     elif intent == "library_info":
-        # Import singleton from app.py
         try:
             from app import rag_service
             rag_results = await rag_service.hybrid_search(question, top_k=4)
@@ -568,24 +655,24 @@ async def agent_query(request: Request, req: AgentRequest):
         )
 
     elif intent == "general_recent":
+        # Real-time/live questions → Claude web search or GPT knowledge fallback
+        answer = await _web_search_answer(question, history, model)
+        tools_used = ["web_search" if model == "claude" else "llm_knowledge"]
         return AgentResponse(
-            answer=(
-                "I'm LibBee, a library assistant — I don't have access to real-time or current information.<br><br>"
-                "For the latest news or live data, please use a web search engine.<br><br>"
-                "If you're looking for <strong>recent academic publications</strong> on this topic, "
-                "I can help with that — just ask me to search for articles."
-            ),
+            answer=answer,
             intent="general_recent",
-            tools_used=[],
+            tools_used=tools_used,
             model_used=model,
             response_time=time.time() - start_time,
         )
 
     else:  # general
+        # Factual questions → LLM answers from training knowledge
+        answer = await _llm_general_answer(question, history, model)
         return AgentResponse(
-            answer=_general_redirect_answer(),
+            answer=answer,
             intent="general",
-            tools_used=["general_redirect"],
+            tools_used=["llm_knowledge"],
             model_used=model,
             response_time=time.time() - start_time,
         )