updated project status

src/rag/agent_chain.py

@@ -28,7 +28,7 @@ from src.rag.models import ModelConfigurator as modelconf
 from src.rag.input_handler import InputHandler
 from src.rag.response_formatter import ResponseFormatter
 from src.rag.scope_guardian import ScopeGuardian
-from src.rag.quality_score_handler import QualityEvaluationResult, QualityScoreHandler
+# from src.rag.quality_score_handler import QualityEvaluationResult, QualityScoreHandler
 from src.rag.language_detection import LanguageDetector
 
 from src.utils.logging import get_logger
@@ -49,9 +49,10 @@ class ExecutiveAgentChain:
         self._conversation_history = [] 
         self._cache = Cache.get_cache()
 
-        # AI-middlewares
-        if config.chain.EVALUATE_RESPONSE_QUALITY:
-            self._quality_handler = QualityScoreHandler()
+        # Confidence scoring is intentionally disabled here because the extra
+        # model call adds latency and has not been reliable enough to justify it.
+        # if config.chain.EVALUATE_RESPONSE_QUALITY:
+        #     self._quality_handler = QualityScoreHandler()
         self._language_detector = LanguageDetector()
 
         # Generate unique user ID for this session
@@ -340,6 +341,160 @@ class ExecutiveAgentChain:
 
         return False
 
+    def _get_latest_ai_message_content(self, skip_latest: bool = False) -> str:
+        """Return the latest assistant message content from conversation history."""
+        ai_messages_seen = 0
+
+        for message in reversed(self._conversation_history):
+            if not isinstance(message, AIMessage):
+                continue
+
+            ai_messages_seen += 1
+            if skip_latest and ai_messages_seen == 1:
+                continue
+
+            content = getattr(message, "content", "") or getattr(message, "text", "")
+            if isinstance(content, list):
+                return " ".join(str(part) for part in content)
+            return str(content)
+
+        return ""
+
+    def _is_booking_preference_follow_up(self, query: str) -> bool:
+        """Detect short follow-up answers that continue an active booking flow."""
+        query_lower = query.lower().strip()
+        if not query_lower:
+            return False
+
+        preference_terms = [
+            "online",
+            "on-site",
+            "onsite",
+            "in person",
+            "in-person",
+            "st.gallen",
+            "st. gallen",
+            "morning",
+            "mornings",
+            "afternoon",
+            "afternoons",
+            "evening",
+            "beginning of the week",
+            "start of the week",
+            "end of the week",
+            "monday",
+            "tuesday",
+            "wednesday",
+            "thursday",
+            "friday",
+            "morgens",
+            "vormittag",
+            "vormittags",
+            "nachmittag",
+            "nachmittags",
+            "abends",
+            "wochenanfang",
+            "anfang der woche",
+            "ende der woche",
+            "montag",
+            "dienstag",
+            "mittwoch",
+            "donnerstag",
+            "freitag",
+            "vor ort",
+            "vor-ort",
+            "persönlich",
+            "persoenlich",
+            "hybrid",
+        ]
+
+        if any(term in query_lower for term in preference_terms):
+            return True
+
+        return False
+
+    def _previous_response_requested_booking_preferences(self) -> bool:
+        """Return True when the previous assistant turn asked clarifying booking questions."""
+        content_lower = self._get_latest_ai_message_content().lower()
+        if not content_lower:
+            return False
+
+        booking_context_terms = [
+            "appointment options",
+            "available appointments",
+            "available slots",
+            "appointment slots",
+            "online-terminoptionen",
+            "terminoptionen",
+            "verfügbare slots",
+            "verfügbare termine",
+            "beratungsgespräch",
+            "beratung",
+        ]
+        clarification_terms = [
+            "do you prefer",
+            "would you prefer",
+            "which programme",
+            "which program",
+            "one short question",
+            "final question",
+            "when i know this",
+            "bitte noch kurz",
+            "eine kurze rückfrage",
+            "eine kurze letzte frage",
+            "bevorzugen sie",
+            "haben sie eine tagespräferenz",
+            "sobald ich das weiss",
+            "damit die slots besser passen",
+        ]
+
+        return (
+            any(term in content_lower for term in booking_context_terms)
+            and any(term in content_lower for term in clarification_terms)
+        )
+
+    def _response_commits_to_showing_booking_widget(self, response: str) -> bool:
+        """Detect when the assistant says booking options are being shown now."""
+        response_lower = response.lower()
+
+        positive_terms = [
+            "i can show you",
+            "contact details and available appointment slots are shown below",
+            "appointment options are shown below",
+            "available slots are shown below",
+            "i can now show you",
+            "ich kann ihnen nun",
+            "ich kann ihnen jetzt",
+            "unten werden ihnen",
+            "unten finden sie",
+            "unten sehen sie",
+            "terminoptionen anzeigen",
+            "verfügbaren slots",
+            "verfügbaren termine",
+        ]
+        defer_terms = [
+            "if you would like",
+            "if you later wish",
+            "you can ask me",
+            "if that would be helpful",
+            "sobald ich das weiss",
+            "wenn ich das weiss",
+            "damit die slots besser passen",
+            "bitte noch kurz",
+            "eine kurze rückfrage",
+            "eine kurze letzte frage",
+            "bevorzugen sie",
+            "have you got a preference",
+            "do you prefer",
+            "would you prefer",
+            "which programme",
+            "which program",
+        ]
+
+        return (
+            any(term in response_lower for term in positive_terms)
+            and not any(term in response_lower for term in defer_terms)
+        )
     def _is_explicit_booking_intent(self, query: str) -> bool:
         """Detect whether the user is actively asking to book or accepting a booking offer."""
         query_lower = query.lower()
@@ -348,6 +503,8 @@ class ExecutiveAgentChain:
             "schedule",
             "appointment",
             "consultation",
+            "need a consultation",
+            "personal consultation",
             "speak with",
             "talk to an advisor",
             "talk to admissions",
@@ -360,6 +517,12 @@ class ExecutiveAgentChain:
             "termin vereinbaren",
             "beratungstermin",
             "beratungsgespräch",
+            "ich brauche eine beratung",
+            "ich möchte eine beratung",
+            "ich will eine beratung",
+            "beratung für",
+            "persönliche beratung",
+            "persoenliche beratung",
             "mit jemandem sprechen",
             "mit admissions sprechen",
             "mit der zulassung sprechen",
@@ -724,6 +887,11 @@ class ExecutiveAgentChain:
         
         response_language = self._stored_language
         explicit_booking_intent = self._is_explicit_booking_intent(preprocessed_query)
+        booking_preference_follow_up = (
+            self._conversation_state.get('handover_requested') is True
+            and self._previous_response_requested_booking_preferences()
+            and self._is_booking_preference_follow_up(preprocessed_query)
+        )
        
         # 1. History Update 
         self._conversation_history.append(HumanMessage(preprocessed_query))
@@ -752,18 +920,17 @@ class ExecutiveAgentChain:
 
         formatted_response = ResponseFormatter.clean_response(formatted_response)
 
-        # Step 7: Language fallback mechanisms and response quality evaluation
         confidence_fallback = False
-        if config.chain.EVALUATE_RESPONSE_QUALITY:
-            quality_evaluation: QualityEvaluationResult = self._quality_handler. \
-                evaluate_response_quality(preprocessed_query, formatted_response)
-            
-            chain_logger.info(f"Quality Score: {quality_evaluation.overall_score:1.2f}")
-
-            if quality_evaluation.overall_score < config.chain.CONFIDENCE_THRESHOLD:
-                confidence_fallback = True
-                formatted_response = CONFIDENCE_FALLBACK_MESSAGE[response_language]
-                chain_logger.info(f"Fallback Mechanism activated!")
+        # if config.chain.EVALUATE_RESPONSE_QUALITY:
+        #     quality_evaluation: QualityEvaluationResult = self._quality_handler. \
+        #         evaluate_response_quality(preprocessed_query, formatted_response)
+        #
+        #     chain_logger.info(f"Quality Score: {quality_evaluation.overall_score:1.2f}")
+        #
+        #     if quality_evaluation.overall_score < config.chain.CONFIDENCE_THRESHOLD:
+        #         confidence_fallback = True
+        #         formatted_response = CONFIDENCE_FALLBACK_MESSAGE[response_language]
+        #         chain_logger.info("Fallback Mechanism activated!")
 
         # Add to history
         self._conversation_history.append(AIMessage(formatted_response))
@@ -777,17 +944,25 @@ class ExecutiveAgentChain:
                 self._log_user_profile()
 
         formatted_response = ResponseFormatter.format_name_of_university(formatted_response, language=response_language)
-        appointment_requested = bool(explicit_booking_intent)
-        show_booking_widget = bool(explicit_booking_intent)
+        booking_flow_requested = explicit_booking_intent or booking_preference_follow_up
+        appointment_requested = bool(booking_flow_requested)
+        show_booking_widget = bool(
+            booking_flow_requested and (
+                structured_response.show_booking_widget
+                or self._response_commits_to_showing_booking_widget(formatted_response)
+            )
+        )
 
-        if structured_response.appointment_requested and not explicit_booking_intent:
-            chain_logger.info("Suppressed booking widget because no explicit booking intent was detected.")
+        if structured_response.appointment_requested and not booking_flow_requested:
+            chain_logger.info("Suppressed booking state because no user-led booking intent was detected.")
+        elif booking_preference_follow_up and show_booking_widget:
+            chain_logger.info("Continuing active booking flow and showing booking widget for a preference follow-up.")
         
         return LeadAgentQueryResponse(
             response = formatted_response,
             language = response_language,
             confidence_fallback = confidence_fallback,
-            should_cache = not any([confidence_fallback, appointment_requested, structured_response.is_context_dependent]),
+            should_cache = False if (confidence_fallback or appointment_requested or structured_response.is_context_dependent) else True,
             processed_query = preprocessed_query,
             appointment_requested = appointment_requested,
             show_booking_widget = show_booking_widget,

src/rag/models.py

@@ -76,7 +76,7 @@ class ModelConfigurator:
         
         from langchain_openai import ChatOpenAI
         cls._subagent_model_instance = ChatOpenAI(
-            model='gpt-5.1',
+            model='gpt-5.1-instant',
             openai_api_key=config.llm.get_api_key(),
             max_tokens=3072,
             temperature=0.01,

src/rag/prompts.py

@@ -300,7 +300,7 @@ RULES:
       - factual, static information (e.g. prices, durations, deadlines, program structure)
       - general definitions or explanations
       - publicly available information that does not vary by user
-    
+
     RULES:
     - Answer in the user's language. NEVER leave English terms untranslated in a German response. Key German translations:
       "tuition fee reduction" → "Studiengebührenreduktion", "tuition" → "Studiengebühr(en)", "included in tuition" → "in den Studiengebühren enthalten", "not included" → "nicht enthalten", "application deadline" → "Bewerbungsfrist".