Create classifier_prompt.py

knowledge/classifier_prompt.py

@@ -0,0 +1,189 @@
+from typing import Dict, List
+
+
+GLOBAL_INTENTS = [
+    "restart",
+    "new_topic",
+    "complaint",
+    "direct_support",
+    "courses_info",
+    "children_courses",
+    "adults_courses",
+    "new_student",
+    "current_student",
+    "unclear",
+]
+
+STATE_ALLOWED_INTENTS: Dict[str, List[str]] = {
+    "WAITING_MAIN_MENU": [
+        "new_student",
+        "current_student",
+        "courses_info",
+        "complaint",
+        "direct_support",
+    ],
+    "WAITING_USER_TYPE": [
+        "new_student",
+        "current_student",
+    ],
+    "WAITING_AUDIENCE": [
+        "adults",
+        "children",
+    ],
+    "WAITING_PRIOR_STUDY": [
+        "prior_study_yes",
+        "prior_study_no",
+    ],
+    "WAITING_BEGINNER_SCHEDULE_CHOICE": [
+        "confirm_schedule_reviewed",
+        "proceed_booking",
+        "switch_to_prior_study_true",
+        "switch_to_prior_study_false",
+        "support_needed",
+    ],
+    "WAITING_PDF_102_CONFIRMATION": [
+        "confirm_pdf_reviewed",
+        "switch_to_prior_study_true",
+        "switch_to_prior_study_false",
+        "support_needed",
+    ],
+    "WAITING_PLACEMENT_TEST_CONFIRMATION": [
+        "confirm_placement_test_reviewed",
+        "switch_to_prior_study_true",
+        "switch_to_prior_study_false",
+        "support_needed",
+    ],
+    "WAITING_CURRENT_STUDENT_ACTION": [
+        "current_student_support",
+        "current_student_next_level",
+    ],
+    "WAITING_SUPPORT_QUESTION": [
+        "support_question_text",
+    ],
+    "WAITING_LEVEL_SELECTION": [
+        "level_selected",
+        "support_needed",
+    ],
+    "WAITING_PAYMENT_METHOD": [
+        "payment_method_selected",
+        "support_needed",
+    ],
+    "WAITING_COMPLAINT_FORM": [
+        "complaint_form_submitted",
+    ],
+    "HANDOFF_DONE": [
+        "thanks",
+        "courses_info",
+        "children_courses",
+        "adults_courses",
+        "new_student",
+        "current_student",
+        "direct_support",
+        "complaint",
+        "restart",
+        "new_topic",
+    ],
+}
+
+INTENT_DESCRIPTIONS: Dict[str, str] = {
+    "restart": "The user wants to restart the conversation from the beginning.",
+    "new_topic": "The user explicitly wants to ask about something else or start a new topic.",
+    "complaint": "The user is making or asking to make a complaint.",
+    "direct_support": "The user wants to talk to customer support or service directly.",
+    "courses_info": "The user is asking generally about available courses.",
+    "children_courses": "The user is asking about children courses.",
+    "adults_courses": "The user is asking about adult courses.",
+    "new_student": "The user indicates they are a new student.",
+    "current_student": "The user indicates they are a current student.",
+    "adults": "Direct answer: adults.",
+    "children": "Direct answer: children.",
+    "prior_study_yes": "Direct answer: user studied German before.",
+    "prior_study_no": "Direct answer: user did not study German before.",
+    "confirm_schedule_reviewed": "The user confirms they reviewed the beginner schedule.",
+    "proceed_booking": "The user wants to proceed with booking.",
+    "switch_to_prior_study_true": "The user changes to the path where they studied German before.",
+    "switch_to_prior_study_false": "The user changes to the beginner path.",
+    "support_needed": "The user needs help or wants support within the current topic.",
+    "confirm_pdf_reviewed": "The user confirms they reviewed the PDF/details file.",
+    "confirm_placement_test_reviewed": "The user confirms they reviewed placement test details.",
+    "current_student_support": "Current student wants support or has a question.",
+    "current_student_next_level": "Current student wants to book the next level.",
+    "support_question_text": "The message itself is the support question text.",
+    "level_selected": "The user selected a course level.",
+    "payment_method_selected": "The user selected a payment method.",
+    "complaint_form_submitted": "The user says they submitted the complaint form.",
+    "thanks": "The user is thanking the bot.",
+    "unclear": "The message is unclear and should not be confidently routed.",
+}
+
+
+def get_allowed_intents_for_state(state: str) -> List[str]:
+    state_specific = STATE_ALLOWED_INTENTS.get(state, [])
+    final = []
+    for item in GLOBAL_INTENTS + state_specific:
+        if item not in final:
+            final.append(item)
+    return final
+
+
+def _summarize_flow_data(flow_data: dict) -> str:
+    if not flow_data:
+        return "No extra context."
+
+    fields = []
+    for key in [
+        "customer_type",
+        "audience",
+        "prior_study",
+        "selected_level",
+        "payment_method",
+        "gender",
+    ]:
+        if key in flow_data and flow_data[key] is not None:
+            fields.append(f"{key}={flow_data[key]}")
+
+    return ", ".join(fields) if fields else "No extra context."
+
+
+def build_system_prompt(state: str, flow_data: dict, allowed_intents: List[str]) -> str:
+    descriptions = []
+    for intent in allowed_intents:
+        desc = INTENT_DESCRIPTIONS.get(intent, "")
+        descriptions.append(f"- {intent}: {desc}")
+
+    description_block = "\n".join(descriptions)
+
+    context_summary = _summarize_flow_data(flow_data)
+
+    return f"""
+You are a state-aware intent classifier for an Arabic WhatsApp bot.
+
+Your task:
+- Read the user's Arabic message.
+- Understand the CURRENT conversation state.
+- Choose exactly ONE intent label from the allowed list.
+- Return ONLY the label.
+- Do not explain.
+- Do not add punctuation.
+- Do not add extra words.
+
+Current state:
+{state}
+
+Known context:
+{context_summary}
+
+Allowed intent labels:
+{", ".join(allowed_intents)}
+
+Intent descriptions:
+{description_block}
+
+Decision rules:
+- If the message is a direct answer to the current question, choose the direct answer label.
+- If the user changes path but stays in the same broad topic, choose the appropriate state-switch label.
+- If the user changes to a new topic, choose the correct topic-switch label.
+- If you are not confident, return: unclear
+
+Return only one label from the allowed list.
+""".strip()