Fix the intent prompt

main.py

@@ -155,39 +155,88 @@ class HealthcareChatbot:
         self.json_parser = JsonOutputParser(pydantic_object=EndpointRequest)
 
         # Intent classification prompt
+    #     self.intent_classifier_template = PromptTemplate(
+    #       template="""
+    #       You are an intent classifier. Your job is simple: understand what the user wants and check if any API endpoint can do that.
+
+    #       User Message: {user_query}
+    #       Language: {detected_language}
+    #       API Endpoints: {endpoints_documentation}
+
+    #       Think step by step:
+
+    #       1. What does the user want from this message?
+    #       Read the user's message carefully. What is the user trying to say or accomplish? What would a human understand from this message?
+
+    #       2. Can any API endpoint fulfill what the user wants?
+    #       Look at each API endpoint. Does any endpoint do what the user is asking for? Be very precise - only say yes if there's a clear match.
+
+    #       Important rules:
+    #       - Focus ONLY on the current message, ignore conversation history for classification
+    #       - If the user is just talking, being social, or saying something casual, that's CONVERSATION
+    #       - Only choose API_ACTION if the user is clearly asking for something an API endpoint can do
+    #       - When you're not sure, choose CONVERSATION
+
+    #       Answer in this format:
+    #       {{
+    #           "intent": "API_ACTION" or "CONVERSATION",
+    #           "confidence": [0.0 to 1.0],
+    #           "reasoning": "What does the user want? Can any API do this?",
+    #           "requires_backend": true or false
+    #       }}
+    #       """,
+    #       input_variables=["user_query", "detected_language", "conversation_history", "endpoints_documentation"]
+    #   )
+
         self.intent_classifier_template = PromptTemplate(
-          template="""
-          You are an intent classifier. Your job is simple: understand what the user wants and check if any API endpoint can do that.
-
-          User Message: {user_query}
-          Language: {detected_language}
-          API Endpoints: {endpoints_documentation}
-
-          Think step by step:
-
-          1. What does the user want from this message?
-          Read the user's message carefully. What is the user trying to say or accomplish? What would a human understand from this message?
-
-          2. Can any API endpoint fulfill what the user wants?
-          Look at each API endpoint. Does any endpoint do what the user is asking for? Be very precise - only say yes if there's a clear match.
-
-          Important rules:
-          - Focus ONLY on the current message, ignore conversation history for classification
-          - If the user is just talking, being social, or saying something casual, that's CONVERSATION
-          - Only choose API_ACTION if the user is clearly asking for something an API endpoint can do
-          - When you're not sure, choose CONVERSATION
-
-          Answer in this format:
-          {{
-              "intent": "API_ACTION" or "CONVERSATION",
-              "confidence": [0.0 to 1.0],
-              "reasoning": "What does the user want? Can any API do this?",
-              "requires_backend": true or false
-          }}
-          """,
-          input_variables=["user_query", "detected_language", "conversation_history", "endpoints_documentation"]
-      )
+            template="""
+            You are a strict intent classification system. Your only task is to determine if the user message requires an API action or is general conversation.
+
+            === ABSOLUTE RULES ===
+            1. OUTPUT FORMAT MUST BE EXACTLY:
+            {{
+                "intent": "API_ACTION" or "CONVERSATION",
+                "confidence": 0.0-1.0,
+                "reasoning": "clear justification",
+                "requires_backend": true or false
+            }}
+            2. Never invent custom intent types
+            3. Never output endpoint names in the intent field
+            4. "requires_backend" must match the intent (true for API_ACTION)
+
+            === CLASSIFICATION CRITERIA ===
+            API_ACTION must meet ALL of:
+            - The message contains a clear, actionable request
+            - The request matches a documented API endpoint's purpose
+            - The request requires specific backend functionality
+
+            CONVERSATION applies when:
+            - The message is social/greeting/smalltalk
+            - The request is too vague for API action
+            - No API endpoint matches the request
 
+            === INPUT DATA ===
+            User Message: {user_query}
+            Detected Language: {detected_language}
+            API Endpoints: {endpoints_documentation}
+
+            === DECISION PROCESS ===
+            1. Analyze the message literally - what is the explicit request?
+            2. Check endpoints documentation - is there an exact functional match?
+            3. If uncertain, default to CONVERSATION
+            4. Validate against rules before responding
+
+            === OUTPUT VALIDATION ===
+            Before responding, verify:
+            - Intent is ONLY "API_ACTION" or "CONVERSATION"
+            - Confidence reflects certainty (1.0 = perfect match)
+            - Reasoning explains the endpoint match (for API_ACTION)
+            - requires_backend aligns with intent
+
+            Respond ONLY in the exact specified format.
+            """,
+            input_variables=["user_query", "detected_language", "conversation_history", "endpoints_documentation"]
+        )