Selection Pipeline:
-
Receive three candidate questions from the Step Agent.
-
Extract structured feature vectors, including:
-
Topic Embeddings (semantic vector representations)
-
Entity Links (conceptual and factual mappings)
-
Context Frequency (local and global recurrence metrics)
-
-
Score each question using the defined function.
-
Apply historical penalties for redundancy and rewards for semantic novelty.
-
Introduce domain-specific and urgency modifiers.
-
Trigger stochastic fallback mode if scoring spread criteria are met.
-
Select the top candidate based on final adjusted score.
-
Log the full decision trace for auditing, review, and adaptive training.