sync: backend from GitHub Actions

core/inference_engine.py

@@ -1,14 +1,5 @@
 """
 Inference Engine — Akinator-style, context-aware, guaranteed-to-terminate AI.
-
-Fixes in this version:
-- get_next_question() returns None when question_selector finds nothing useful
-  (active-item grounding removes impossible questions like "Russia border?" after
-  Indian Subcontinent confirmed). This immediately triggers ready_to_guess=True
-  in app.py, ending the game — no infinite loop.
-- _should_stop_asking() is the single source-of-truth for stopping; it is called
-  both after process_answer() and inside get_next_question().
-- soft_filter() no longer takes top_k from GAME_CONFIG (removed aggressive top_k=3).
 """
 
 import logging
@@ -22,7 +13,7 @@ from algorithms.information_gain import InformationGain
 from algorithms.bayesian_network import BayesianNetwork
 from models.game_state import GameState
 from models.item_model import Item
-from backend.config import GAME_CONFIG
+from config import GAME_CONFIG
 from services.firebase_service import FirebaseService
 
 logger = logging.getLogger(__name__)
@@ -32,17 +23,17 @@ class InferenceEngine:
     """Main AI Engine — Akinator-style, guaranteed termination."""
 
     def __init__(self):
-        self.question_selector    = QuestionSelector()
-        self.probability_manager  = ProbabilityManager()
+        self.question_selector     = QuestionSelector()
+        self.probability_manager   = ProbabilityManager()
         self.confidence_calculator = ConfidenceCalculator()
-        self.information_gain     = InformationGain()
-        self.bayesian_network     = BayesianNetwork()
-        self.firebase_service     = FirebaseService()
+        self.information_gain      = InformationGain()
+        self.bayesian_network      = BayesianNetwork()
+        self.firebase_service      = FirebaseService()
         self.active_games: Dict[str, GameState] = {}
         self.session_stats = {
-            'games_played': 0,
+            'games_played':       0,
             'successful_guesses': 0,
-            'average_questions': 0,
+            'average_questions':  0,
         }
         logger.info("InferenceEngine ready (v3.2 — context-aware, guaranteed termination)")
 
@@ -88,14 +79,6 @@ class InferenceEngine:
     # ── Question flow ─────────────────────────────────────────────────────────
 
     def get_next_question(self, game_state: GameState) -> Optional[Dict]:
-        """
-        Returns the best next question, or None if the game should end.
-        None can mean:
-          (a) confidence/item-count threshold reached → ready to guess
-          (b) no useful question left after active-item grounding → ready to guess
-        Both cases set ready_to_guess=True in app.py.
-        """
-        # Check stopping conditions first
         if self._should_stop_asking(game_state):
             logger.info(f"[{game_state.session_id}] Stop condition met before question select.")
             return None
@@ -114,11 +97,6 @@ class InferenceEngine:
             game_state_history=game_state.answer_history,
         )
 
-        # ── KEY FIX: selector returned None → no useful question remains ─────
-        # This happens when all remaining questions are about attributes that
-        # none of the active items actually have (e.g., "Russia border?" when
-        # only Indian Subcontinent countries are left).
-        # Returning None here makes app.py set ready_to_guess=True immediately.
         if question is None:
             logger.info(
                 f"[{game_state.session_id}] Selector found no useful question "
@@ -143,7 +121,6 @@ class InferenceEngine:
         question = game_state.current_question
         game_state.record_answer(answer)
 
-        # Bayesian probability update
         active_items = game_state.get_active_items()
         for item in active_items:
             item.probability = self.probability_manager.update_item_probability(
@@ -151,7 +128,6 @@ class InferenceEngine:
             )
 
         self.probability_manager.normalize_probabilities(game_state.items)
-        # Conservative soft filter — keeps top-20, threshold 0.1% of top item
         self.probability_manager.soft_filter(game_state.items)
         self.bayesian_network.update_beliefs(question, answer, game_state.items)
 
@@ -168,10 +144,10 @@ class InferenceEngine:
         )
 
         return {
-            'confidence':        confidence,
+            'confidence':         confidence,
             'active_items_count': len(current_active),
-            'top_prediction':    top_item.to_dict() if top_item else None,
-            'should_stop':       should_stop,
+            'top_prediction':     top_item.to_dict() if top_item else None,
+            'should_stop':        should_stop,
         }
 
     # ── Final prediction ──────────────────────────────────────────────────────
@@ -208,14 +184,9 @@ class InferenceEngine:
     # ── Stopping logic ────────────────────────────────────────────────────────
 
     def _should_stop_asking(self, game_state: GameState) -> bool:
-        """
-        Single source of truth for stopping.
-        Returns True when the engine should make its final guess.
-        """
         active_items = game_state.get_active_items()
         active_count = len(active_items)
 
-        # 1. Only 1 or 2 items remain → guess immediately
         force_at = GAME_CONFIG.get('force_guess_at_items', 2)
         if active_count <= force_at:
             logger.info(
@@ -224,12 +195,10 @@ class InferenceEngine:
             )
             return True
 
-        # 2. No questions left at all (all asked)
         if not game_state.get_available_questions():
             logger.info(f"[{game_state.session_id}] No available questions left.")
             return True
 
-        # 3. Adaptive confidence threshold
         confidence = self.confidence_calculator.calculate(active_items)
         return self.confidence_calculator.should_make_guess(
             confidence,
@@ -252,8 +221,8 @@ class InferenceEngine:
         games   = self.session_stats['games_played']
         success = self.session_stats['successful_guesses']
         return {
-            'games_played':     games,
+            'games_played':       games,
             'successful_guesses': success,
-            'success_rate':     (success / games * 100) if games > 0 else 0,
-            'average_questions': self.session_stats['average_questions'],
+            'success_rate':       (success / games * 100) if games > 0 else 0,
+            'average_questions':  self.session_stats['average_questions'],
         }