Generative retrieval replaces item IDs with *Semantic IDs* (SIDs) decoded autoregressively by a sequence model. Playing the role of tokenizer, the codebook largely determines the quality of generative retrieval. Existing codebook metrics, however, only characterize the global uniformity of code usage and do not reveal whether an individual SID correctly represents its item or separates it from others. We address this problem with two metrics and a matching training objective. We build a query-centered triplet dataset q, a, p, n from large-scale user behavior, and based on which we define **Overlap**, a prefix-based distance between two SIDs, and **Pairwise Overlap Ranking (POR)**, the fraction of triplets where a codebook places the positive closer to the anchor than the hard negative. Both metrics correlate strongly with downstream Hit Rate in generative retrieval. We then introduce the **Differentiable Soft Overlap Loss (DSOL)** to make these two metrics differentiable during codebook learning and optimize them directly. On a public e-commerce benchmark (ESCI-us) and a large-scale industrial video search dataset, DSOL improves HR and POR over standard SID baselines.

Applying LLMs to query-document relevance estimation in industrial document retrieval faces two key challenges: the complexity and heterogeneity of Web content, and stringent latency and resource constraints. To address them, we propose an automated framework with two components: **R**eflection-**A**ware **C**ollaborative **A**gent (**RACA**) and **P**rogressive **O**ptimization **D**istillation **S**trategy (**PODS**). RACA uses collaborative reasoning and reflection to produce relevance labels and structured rationales. PODS stratifies training data by difficulty and combines supervised fine-tuning with Reasoning-Informed GRPO. Rationale supervision improves the student’s first-token relevance decision, while deployment stops after that token and avoids explicit rationale generation. Together, these mechanisms enable reasoning-informed first-token prediction that preserves the benefits of structured rationale supervision under stringent latency and resource constraints. Extensive offline experiments and a 10-day online A/B test in WeChat Search demonstrate the resulting accuracy–efficiency trade-off.