Update src/api/agent.py

src/api/agent.py

@@ -118,7 +118,7 @@ MEDICAL_SEARCH_RE = re.compile(
     re.IGNORECASE,
 )
 
-# FIX: removed "president" — was catching KU-specific questions incorrectly
+# FIX: removed "president" — was incorrectly catching KU leadership questions
 CURRENT_INFO_RE = re.compile(
     r"\b(current|currently|latest|today|recent|recently|now|this week|this month"
     r"|news|update|updated|prime minister|ceo|minister|king|ruler|who won)\b",
@@ -136,7 +136,8 @@ MEDICAL_KEYWORDS = [
     "crispr", "gene therapy", "gene editing", "genomic medicine", "precision medicine",
 ]
 
-# LLM Classifier system prompt
+# ── LLM Classifier System Prompt ──────────────────────────────────────────────
+
 _CLASSIFIER_SYSTEM = """You are the routing brain of LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE).
 Your only job: read the user message and return a single JSON object routing it to the right handler.
 
@@ -145,8 +146,8 @@ 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 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
+  "general_recent" — questions about TRULY live/real-time data: breaking news, live sports scores, today's weather, current stock prices, events happening right now
+  "general"        — ALL other factual questions: university info, people, history, science, technology, definitions, explanations, stable facts like who holds a position
 
 RULES:
   1. When in doubt between library_info and social, choose library_info if ANY library keyword is present.
@@ -155,7 +156,8 @@ 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).
+  7. Questions about people, university leadership, presidents, provosts, history, science, technology = general (answered via web search).
+  8. Only use general_recent for truly live data that changes by the minute/hour.
 
 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".
@@ -210,7 +212,7 @@ def _looks_research_question(question: str) -> bool:
 
 
 def _rule_based_classify(question: str) -> Dict:
-    """Fallback classifier used when no API key is available."""
+    """Fallback classifier when no API key is available."""
     q = question.lower()
     if SOCIAL_RE.match(question.strip()):
         return {
@@ -403,6 +405,92 @@ async def _rag_answer(
         )
 
 
+async def _web_search_answer(
+    question: str,
+    history: List[ChatMessage],
+    model: str,
+) -> str:
+    """
+    Answer using live web search for BOTH models.
+
+    Claude → Anthropic web search tool (web_search_20250305)
+    GPT    → OpenAI Responses API with web_search_preview tool
+
+    Both fall back to _llm_general_answer() if web search fails.
+    """
+    settings = get_settings()
+
+    # ── Claude: Anthropic web search tool ─────────────────────────────────
+    if model == "claude" and settings.anthropic_api_key:
+        try:
+            import anthropic
+            client = anthropic.Anthropic(api_key=settings.anthropic_api_key)
+
+            response = client.messages.create(
+                model="claude-3-haiku-20240307",
+                max_tokens=600,
+                system=(
+                    "You are LibBee, the Khalifa University Library AI Assistant (Abu Dhabi, UAE). "
+                    "Answer the question using current web search results. "
+                    "Be concise and factual. Use HTML <br> for line breaks. "
+                    "KU = Khalifa University. "
+                    "If the topic is academic, briefly mention you can help find scholarly articles."
+                ),
+                tools=[{"type": "web_search_20250305", "name": "web_search"}],
+                messages=[{"role": "user", "content": question}],
+            )
+
+            text = ""
+            for block in response.content:
+                if hasattr(block, "text"):
+                    text += block.text
+
+            if text.strip():
+                logger.info("Claude web search succeeded")
+                return text.strip()
+
+        except Exception as e:
+            logger.warning(f"Claude web search failed: {e} — trying GPT web search")
+
+    # ── GPT: OpenAI Responses API with web_search_preview ─────────────────
+    if settings.openai_api_key:
+        try:
+            from openai import OpenAI
+            client = OpenAI(api_key=settings.openai_api_key)
+
+            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 the question using current web search results. "
+                    "Be concise and factual. Use HTML <br> for line breaks. "
+                    "KU = Khalifa University — do NOT confuse with Kuwait University. "
+                    "If the topic is academic, briefly mention you can help find scholarly articles."
+                ),
+                input=question,
+            )
+
+            # Extract text from Responses API output items
+            text = ""
+            for item in response.output:
+                if hasattr(item, "content"):
+                    for block in item.content:
+                        if hasattr(block, "text"):
+                            text += block.text
+
+            if text.strip():
+                logger.info("GPT web search succeeded")
+                return text.strip()
+
+        except Exception as e:
+            logger.warning(f"GPT web search failed: {e} — falling back to LLM knowledge")
+
+    # ── Final fallback: LLM training knowledge ────────────────────────────
+    logger.warning("Both web search tools failed — answering from LLM training knowledge")
+    return await _llm_general_answer(question, history, model)
+
+
 async def _llm_general_answer(
     question: str,
     history: List[ChatMessage],
@@ -410,7 +498,7 @@ async def _llm_general_answer(
 ) -> str:
     """
     Answer general factual questions from LLM training knowledge.
-    Used for: university info, people, places, science, technology, etc.
+    Used as fallback when web search is unavailable or fails.
     """
     settings = get_settings()
     if not settings.openai_api_key:
@@ -425,8 +513,9 @@ async def _llm_general_answer(
         "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."
+        "Note: your training data has a cutoff date so some facts may be outdated — "
+        "mention this only if the question is about something that changes frequently. "
+        "If the topic is academic, add one short sentence offering to find library resources."
     )
 
     try:
@@ -445,56 +534,6 @@ async def _llm_general_answer(
         )
 
 
-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)
@@ -655,24 +694,24 @@ async def agent_query(request: Request, req: AgentRequest):
         )
 
     elif intent == "general_recent":
-        # Real-time/live questions → Claude web search or GPT knowledge fallback
+        # Truly live/real-time questions → web search (both models)
         answer = await _web_search_answer(question, history, model)
-        tools_used = ["web_search" if model == "claude" else "llm_knowledge"]
         return AgentResponse(
             answer=answer,
             intent="general_recent",
-            tools_used=tools_used,
+            tools_used=["web_search"],
             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)
+        # Factual questions (people, places, science, university info, etc.)
+        # → web search for current accurate answers (both models)
+        answer = await _web_search_answer(question, history, model)
         return AgentResponse(
             answer=answer,
             intent="general",
-            tools_used=["llm_knowledge"],
+            tools_used=["web_search"],
             model_used=model,
             response_time=time.time() - start_time,
         )