🎯 Default to document search for ALL voice queries

- Changed default behavior: ALL non-greeting queries now search 23K documents
- Fixes: When ASR mishears 'pension rules' → 'challenge loans', still searches DB
- Previous: Only searched if query contained keywords like 'pension', 'rules'
- Now: Assumes ALL questions need document search (it's a gov rule assistant)
- Only skips search for clear greetings: hi, hello, thank you, bye
- Improves robustness when voice transcription is imperfect

groq_websocket_handler.py

@@ -498,20 +498,22 @@ class GroqWebSocketHandler:
                 "reason": f"Contains government/policy keywords: {', '.join(matched_keywords)}"
             }
         
-        # Default: treat as document search unless clearly casual
-        if len(query.split()) > 2:  # Multi-word queries likely need document search
+        # Default: ALWAYS search documents for non-casual queries
+        # This is a pension/government assistant, so most queries should search documents
+        if len(query.split()) >= 2:  # Multi-word queries likely need document search
             return {
                 "requires_documents": True,
                 "query_type": "information_request",
-                "confidence": 0.6,
-                "reason": "Multi-word query likely needs document search"
+                "confidence": 0.7,
+                "reason": "Multi-word query - defaulting to document search for better accuracy"
             }
         
+        # Even single-word queries should search documents (unless they're greetings)
         return {
-            "requires_documents": False,
-            "query_type": "general",
-            "confidence": 0.5,
-            "reason": "Simple query, may not need documents"
+            "requires_documents": True,
+            "query_type": "general_info",
+            "confidence": 0.6,
+            "reason": "Defaulting to document search - this is a government rule assistant"
         }
 
     async def _generate_audio_response(self, websocket: WebSocket, session_id: str, text: str):