This paper targets e-commerce search relevance. While Large Language Models (LLMs) have demonstrated significant potential in this field, they often encounter performance bottlenecks in persistent 'corner cases' within complex industrial scenarios. Existing research primarily focuses on optimizing reasoning trajectories via Reinforcement Learning. However, real-world observations suggest that the primary bottleneck stems from knowledge boundaries, where the absence of domain-specific intelligence in the model's parametric memory creates a contextual void. This void persists when interpreting idiosyncratic queries or niche products and cannot be resolved solely through reasoning-path optimization. To bridge this gap, we propose K-CARE, a framework that extends the model's cognitive reach by grounding reasoning in external knowledge. K-CARE comprises two synergistic components: (1) Symmetrical Contextual Anchoring (SCA), which fills the contextual void by anchoring queries and products with behavior-derived implicit knowledge; and (2) Analogical Prototype Reasoning (APR), which leverages expert-curated prototypical knowledge to calibrate decision boundaries through in-context analogy. Extensive offline evaluations and online A/B tests on a leading e-commerce platform demonstrate that K-CARE significantly outperforms state-of-the-art baselines, delivering substantial commercial impact by resolving knowledge-intensive relevance challenges.

翻译：暂无翻译