Asymmetric cross-source retrieval transfer reveals the role of corpus semantic richness in industrial RAG systems for wind-turbine maintenance

Retrieval-Augmented Generation (RAG) systems for specialized industrial domains face a persistent cold-start problem when high-quality, target-domain training data is scarce. We investigate cross-source transfer in this setting using wind-turbine technical documentation as an industrial testbed. A 3 × 4 factorial experiment over twelve embedding-model configurations reveals an asymmetric transfer pattern: embeddings fine-tuned on a semantically rich bilingual corpus from an alternative manufacturer outperform domain-matched models on the target manufacturer's own test set (MRR 0.7902 vs. 0.7012). A seven-category semantic analysis shows the effect is largest for safety-critical queries (+ 48.2%; n = 21, post-hoc power 0.91), and we report a data-dilution effect in which naive multi-source aggregation degrades performance. Because the two corpora differ along several correlated axes - monolingual lexical density, raw volume, and cross-lingual breadth - we characterize semantic richness with independent, pre-specified metrics (MATTR, MTLD, hapax, Jaccard) and present it as a multi-dimensional, testable construct rather than a single factor. To probe robustness, we evaluated five multi-source mixing protocols (naive concatenation, balanced down-sampling, inverse-frequency oversampling, similarity-based quality filtering, and K-means diversity-aware sampling) against three single-brand baselines on a fourth-manufacturer zero-shot test set (Test D, n = 99). No mixing protocol significantly outperformed the strongest single-brand fine-tune, confirming that the asymmetric transfer effect is robust to data-combination strategy and extends to out-of-distribution generalization. We interpret the observed asymmetry as an empirical pattern within this two-manufacturer benchmark whose broader applicability to other industrial domains remains to be validated.