FT-MMQS: Fault-Tolerant Multimodal Query Suggestion via Parameter-Based Fallback and Quality-Aware Inference

Multimodal Query Suggestion (MMQS) systems, such as RL4Sugg, enrich search experiences by generating textual queries directly from user-provided images. Yet in practice, these systems remain sensitive to runtime instabilities, including agent crashes or reward model failures, which can degrade output quality or interrupt service. In this work, we propose FT-MMQS, a novel and lightweight extension of RL4Sugg that enhances reliability through quality-aware inference. Rather than modifying the core RLHF pipeline, FT-MMQS introduces a runtime monitoring layer that evaluates two key signals, Reward (intent alignment) and Diversity (semantic coverage), and activates a parameter-based fallback (num_beams = 2, num_return_sequences = 2) when quality drops below empirically determined thresholds (e.g., Reward < 0.4 or Diversity < 0.15). This design enables graceful degradation: instead of failure, the system delivers robust, interpretable suggestions under stress. Evaluated on a real-world e-commerce dataset, including controlled simulations of RewardNet malfunction, FT-MMQS achieves 100% availability across all fault conditions, retains over 92% of the original DCG, and improves average reward from 0.32 to 0.55 in degraded scenarios. Notably, no additional models or infrastructure are required. By integrating principles from dependable computing into multimodal retrieval, FT-MMQS demonstrates a practical path toward systems that are not only intelligent, but also reliable in real-world deployment.