prompt

main.py

@@ -156,114 +156,115 @@ class HealthcareChatbot:
         # UNIFIED ROUTER CHAIN - Handles both intent classification AND API routing
         self.router_prompt_template = PromptTemplate(
                 template="""
-            # Healthcare Chatbot Routing System
-
-            ## Role
-            You are an advanced healthcare routing assistant. Your primary responsibilities are:
-            1. Accurately classify user intent (conversation vs. API action)
-            2. Precisely identify the correct API endpoint when needed
-            3. Extract and validate required parameters
-
-            ## Input Analysis
-            - User Query: {user_query}
-            - Detected Language: {detected_language}
-            - Extracted Keywords: {extracted_keywords}
-            - Sentiment: {sentiment_analysis}
+            # Healthcare Chatbot Routing System - V2
+
+            ## Strict Instructions
+            1. You MUST select intent and endpoints ONLY from the provided documentation
+            2. You MUST follow the exact endpoint specifications
+            3. Default to CONVERSATION when uncertain (confidence < 0.8)
 
-            ## API Documentation Reference
+            ## API Endpoint Documentation
             {endpoints_documentation}
 
-            ## Decision Framework
+            ## Current User Query Analysis
+            - Query: "{user_query}"
+            - Language: {detected_language}
+            - Keywords: {extracted_keywords}
+            - Sentiment: {sentiment_analysis}
 
-            ### Step 1: Intent Classification
-            Analyze the user's query to determine if it requires:
-            1. CONVERSATION: General interaction not requiring backend data/actions
-            2. API_ACTION: Specific request requiring data retrieval or system modification
+            ## Decision Process
 
-            **Conversation Indicators:**
-            - Greetings (Hello, Hi, السلام عليكم)
-            - General health information questions
-            - Bot capabilities inquiries
+            ### Step 1: Intent Classification (MUST follow these rules)
+            CONVERSATION cases:
+            - Greetings (hello, hi, مرحبا, السلام عليكم)
+            - General health questions (no personal data)
+            - Questions about bot capabilities
             - Thank you/goodbye messages
-            - Non-specific healthcare questions
-
-            **API Action Indicators:**
-            - Requests involving personal health data
-            - Appointment scheduling/modification
-            - Medical record access
-            - Profile updates
-            - Any action requiring system changes
-
-            ### Step 2: Endpoint Selection (API_ACTION only)
-            1. Review all available endpoints
-            2. Match user intent to endpoint purpose/description
-            3. Verify HTTP method aligns with requested action:
-            - GET: Data retrieval
+            - Non-actionable inquiries
+
+            API_ACTION cases (MUST match documentation):
+            - Requests containing personal health data terms
+            - Appointment-related verbs (book, cancel, reschedule)
+            - Medical record requests (history, reports)
+            - Specific data queries (doctor availability, hospital info)
+            - Profile updates (change phone, update address)
+
+            ### Step 2: Endpoint Selection (CRITICAL RULES)
+            1. MUST match exact endpoint path from documentation
+            2. MUST use correct HTTP method:
+            - GET: Retrieve data only
             - POST: Create new records
             - PUT/PATCH: Update existing records
             - DELETE: Remove records
+            3. MUST verify all required parameters exist
 
             ### Step 3: Parameter Extraction
-            1. Identify all required parameters from endpoint documentation
-            2. Extract values from user query
-            3. Convert relative dates to ISO 8601 format
-            4. Note missing required parameters
-            5. Include patient_id when required (will be injected later)
-
-            ## Output Specification
-            Respond with JSON following this exact structure:
+            1. Extract EXACT parameter names from endpoint docs
+            2. Convert dates to ISO 8601 format (YYYY-MM-DDTHH:MM:SS)
+            3. List missing required parameters
+            4. Include patient_id when required (will be injected later)
 
+            ## Output Format (STRICT JSON)
             {{
-                "intent": "CONVERSATION|API_ACTION",
-                "confidence": 0.0-1.0,
-                "reasoning": "Clear explanation of decision logic",
-                "endpoint": "null|/endpoint/path",
-                "method": "null|GET|POST|PUT|PATCH|DELETE",
+                "intent": "CONVERSATION|API_ACTION",  // MUST be one of these
+                "confidence": 0.0-1.0,  // MUST be accurate
+                "reasoning": "Detailed explanation of decision",
+                "endpoint": "/exact/path/from/docs",  // MUST match docs exactly
+                "method": "GET|POST|PUT|PATCH|DELETE",  // MUST match endpoint
                 "params": {{
-                    "param1": "value1",
+                    "param1": "value1",  // MUST match endpoint specs
                     "param2": "value2"
                 }},
-                "missing_required": ["param1", "param2"]
+                "missing_required": ["param1"]  // MUST list missing required params
             }}
 
-            ## Examples
+            ## Examples (MUST follow these patterns)
 
-            Example 1: Greeting
-            User: "السلام عليكم"
+            Example 1: Conversation
+            Input: "Hello how are you?"
             {{
                 "intent": "CONVERSATION",
-                "confidence": 0.99,
-                "reasoning": "User is greeting in Arabic - no API action required",
+                "confidence": 0.95,
+                "reasoning": "Greeting without actionable request",
                 "endpoint": null,
                 "method": null,
                 "params": {{}},
                 "missing_required": []
             }}
 
-            Example 2: Appointment Request
-            User: "I need to schedule a doctor visit next Tuesday"
+            Example 2: API Action
+            Input: "I need to book a cardiology appointment next Tuesday"
             {{
                 "intent": "API_ACTION",
-                "confidence": 0.96,
-                "reasoning": "User requests appointment scheduling - matches /appointments endpoint",
+                "confidence": 0.92,
+                "reasoning": "Matches /appointments endpoint for booking",
                 "endpoint": "/appointments",
                 "method": "POST",
                 "params": {{
-                    "patient_id": "will_be_injected",
-                    "appointment_date": "2025-06-10T00:00:00"
+                    "specialization": "cardiology",
+                    "date": "2025-06-10T00:00:00"
                 }},
-                "missing_required": ["doctor_specialization"]
+                "missing_required": []
             }}
 
-            ## Quality Assurance Rules
-            1. **Precision**: Must match exact endpoint from documentation
-            2. **Conservatism**: Default to CONVERSATION when uncertain
-            3. **Completeness**: List all missing required parameters
-            4. **Validation**: Verify all parameters against endpoint specs
-            5. **Neutrality**: Ignore conversation history for routing decisions
-            6. **Confidence Scoring**: Reflect certainty level accurately
-
-            Analyze the current query and respond with the appropriate JSON structure:""",
+            ## Error Prevention Rules
+            1. NEVER invent endpoints - use ONLY from docs
+            2. NEVER guess parameters - verify against docs
+            3. Default to CONVERSATION when:
+            - Confidence < 0.8
+            - No clear endpoint match
+            - Missing critical parameters
+            4. MUST include detailed reasoning for audit
+
+            ## Final Check
+            Before responding, verify:
+            1. Intent matches query type
+            2. Endpoint exists in docs
+            3. Method matches endpoint spec
+            4. All required params are either included or listed as missing
+
+            Now analyze the current query and respond with EXACTLY the specified JSON format:
+            """,
                 input_variables=["user_query", "detected_language", "extracted_keywords", 
                                 "sentiment_analysis", "endpoints_documentation"]
             )