Add 8999999999999999999999999999

enhanced_websocket_handler.py

@@ -94,19 +94,18 @@ def analyze_query_context(query: str) -> dict:
     doc_matches = sum(1 for kw in doc_keywords if kw in query_lower)
     general_matches = sum(1 for kw in general_keywords if kw in query_lower)
     
-    # Determine query type
+    # Determine query type - FIXED: Be more aggressive about document searches
     if doc_matches > 0 or detected_role:
         query_type = "document_related"
         confidence = max(min(doc_matches * 0.3, 1.0), role_confidence)
-    elif general_matches > 0:
+    elif general_matches > 0 and doc_matches == 0:
+        # Only treat as general if there are ZERO document keywords
         query_type = "general_conversation"
         confidence = min(general_matches * 0.4, 1.0)
-    elif len(query.strip().split()) < 3:
-        query_type = "general_conversation"  # Short queries likely general
-        confidence = 0.6
     else:
-        query_type = "document_related"  # Default to document search for longer queries
-        confidence = 0.3
+        # DEFAULT to document search - this is a government document system
+        query_type = "document_related"
+        confidence = 0.5  # Higher confidence for document search by default
     
     return {
         "type": query_type,
@@ -911,9 +910,17 @@ async def get_hybrid_response(user_message: str, context: str, config: dict, kno
         # Check if we have relevant documents
         has_relevant_docs = docs and any(doc.get("score", 0) > 0.5 for doc in docs)
         
-        # For general conversation queries, use LLM even if we have some documents
-        if query_context.get("type") == "general_conversation" and query_context.get("confidence", 0) > 0.6:
-            logger.info("📱 Detected general conversation, using LLM directly")
+        # FIXED: Always try document search first, even for apparent "general" queries
+        # This is a government document system - most queries should check documents
+        # Only use pure LLM for very clear greetings/thanks with NO document matches
+        very_general_keywords = ['hello', 'hi', 'thank you', 'thanks', 'goodbye', 'bye']
+        is_very_general = (query_context.get("type") == "general_conversation" and 
+                          query_context.get("confidence", 0) > 0.8 and
+                          any(keyword in user_message.lower() for keyword in very_general_keywords) and
+                          not docs)
+        
+        if is_very_general:
+            logger.info("📱 Detected pure greeting/thanks with no documents, using LLM directly")
             llm_response = await generate_llm_fallback_response(user_message, query_context)
             yield {
                 "clause_text": llm_response,