LLM Changes

app/nodes/router.py

@@ -14,23 +14,31 @@ def router_node(state: AgentState) -> AgentState:
     tool_descriptions = "\n".join(
         f'- "{t.name}": {t.description}' for t in ALL_TOOLS
     )
-    router_prompt = f"""You are a query router for an AI assistant.
+    router_prompt = f"""You are a query router. Your ONLY job is to pick the correct route.
 
-Available tools:
-{tool_descriptions}
+    ROUTE RULES — follow strictly in this order:
 
-Knowledge base topics (for RAG):
-- LangGraph, RAG, Guardrails, HITL, Memory in AI agents, Tool calling, Conditional routing
+    1. "tool" — the query needs LIVE or REAL-TIME data that only a tool can provide.
+    Use this if the query mentions: weather, temperature, forecast, calculate, math, arithmetic.
+    Available tools:
+    {tool_descriptions}
 
-Given the user query below, decide the best route:
-  • "tool"    — if any available tool can directly answer or act on this query
-  • "rag"     — if the query asks for information that exists in the knowledge base
-  • "general" — for everything else (chit-chat, opinions, open-ended questions)
+    2. "rag" — the query asks about AI/ML concepts that exist in the knowledge base.
+    RAG topics: LangGraph, LangChain, RAG, Guardrails, HITL, Memory, Conditional routing.
+    Do NOT use rag for weather or calculations — those always go to tool.
 
-Respond ONLY with valid JSON — no explanation, no markdown fences:
-{{"route": "<tool|rag|general>", "reason": "<one sentence why>"}}
+    3. "general" — everything else (greetings, opinions, open-ended questions).
 
-User query: {state["query"]}"""
+    EXAMPLES:
+    "What is the weather in Mumbai?" → tool (needs live weather data)
+    "Calculate 25 * 48"             → tool (needs calculator)
+    "What is LangGraph?"            → rag  (AI concept in knowledge base)
+    "Tell me a joke"                → general
+
+    Respond ONLY with valid JSON, no markdown:
+    {{"route": "<tool|rag|general>", "reason": "<one sentence why>"}}
+
+    User query: {state["query"]}"""
 
     try:
         response = llm.invoke([HumanMessage(content=router_prompt)])