Domain-adaptive contrastive embeddings for frugal skill routing in agentic e-commerce evaluation

Large-language-model (LLM) agents increasingly operate and evaluate e-commerce workflows, and a recurring building block is a skill router : a semantic-retrieval layer that maps a free-form question to the correct skill in a catalog. The same layer lets an LLM-as-judge evaluator decide which skill should have answered and whether the catalog covers the question at all. In practice it runs on general-purpose embedders that mishandle the dense, abbreviation-heavy vocabulary of retail supply-chain operations, and the reflexive fixes, a larger general embedder or a frontier LLM prompted to route, are metered per call. We instead adapt a small open-source encoder to the domain and deliver the adaptation in the two forms a deployed system needs. The static recipe fully fine-tunes a 22M-parameter MiniLM-class encoder on typed contrastive pairs under a leakage-free skill-holdout split, lifting overall routing Hit@1 from 0.474 (a production 1024-d general embedder) to 0.656, and in-distribution Hit@1 from 0.393 to 0.746, over a compact 384-d index at zero marginal inference cost. The continual-learning recipe keeps that encoder current as the catalog grows from 23 to 43 skills: warm-starting plus a 30% replay sample holds overall Hit@1 near 0.76, matching a full retrain at roughly half the compute while avoiding catastrophic forgetting. We state the main caveat plainly: the gains concentrate on skills seen in training, and on genuinely unseen skills the general embedder still leads. A score-fusion ablation tops every arm but reintroduces a per-query general-embedder call, so we keep the standalone specialist central and point to a lightweight learned gate as future work. For the retrieval substrate of agentic e-commerce, a small specialized embedder is both more accurate in domain and far cheaper than a larger general model or a frontier LLM.