Semantic clause retrieval for trademark law using transformer encoders and lexical baselines: a cross-domain agri-robotics compliance case study

Clause-level retrieval is a recurring bottleneck in legal research and compliance workflows: relevant obligations, exceptions, procedures, and enforcement conditions are often buried in long statutes and regulatory texts, and users may not know the exact terminology needed for keyword search. We present an application-oriented semantic clause retrieval pipeline that indexes documents at the clause level and ranks candidates using off-the-shelf sentence-transformer encoders with cosine similarity. Standard lexical baselines are included to contextualize performance under the same top-k retrieval and expert relevance judgment protocol. We evaluate the approach in a cross-domain setting spanning, trademark statute retrieval on Trademark Ordinance data and a scoped agri-robotics compliance corpus covering regulatory and standards-oriented requirements. The trademark benchmark serves as the primary quantitative evaluation, while the agri-robotics component is used to assess cross-domain transfer under a bounded query set without overstating generalization. In addition to aggregate ranking metrics, we report query-level analysis to characterize model behavior and common failure modes, including high-similarity but decision-irrelevant matches that arise from procedural or definitional overlap.