A Retrieval-Augmented Architecture for Engineering Knowledge and CAD Models: An Industrial Case Study

Engineering reuse requires access to both design documents and CAD assets. Retrieval-augmented generation (RAG) combines retrieved evidence with language generation [1], but text-based retrieval alone cannot directly capture geometric characteristics such as hole placement or cavity structure. B-Rep learning provides a way to represent CAD geometry for retrieval, and UV-Net is used in this study as the learned CAD representation [2]. We propose an engineering knowledge architecture with separate document and CAD retrieval paths. In this study, we evaluate only the CAD retrieval path by comparing a learned CAD representation with a simple geometric descriptor. The industrial application motivates the study, while the evaluation in this study is limited to controlled STEP-model parts. Public Fusion data are used only to train the UV-Net representation. The evaluation considers two retrieval settings. The primary evaluation tests broad part-family retrieval, while the stress evaluation examines whether structurally different parts with similar overall geometry can be distinguished.

The proposed architecture separates document and CAD retrieval into independent indexes. Text passages are represented by semantic embeddings, whereas STEP-model solids are represented by geometric embeddings. Retrieved documents and CAD models can then be linked through asset metadata rather than a shared embedding space. Only the CAD retrieval path is evaluated in this study. In this study, we compare a simple geometric descriptor with embeddings learned by UV-Net. The baseline is an eight-component descriptor consisting of volume, surface area, three bounding-box dimensions, and the numbers of faces, edges, and vertices. Volume, surface area, and bounding-box dimensions are normalized using the largest dimension of the model's bounding box as the reference scale. The remaining three components use log(1 + count) for faces, edges, and vertices. For the learned CAD representation, we trained the official self-supervised UV-Net implementation [2,3] from random initialization using STEP models from the Fusion 360 Gallery Segmentation Dataset [4]. Training followed its self-supervised setting and did not use the dataset's segmentation labels. The training and validation sets contained 24,203 and 6,053 models, respectively. Training used 100 epochs with NT-Xent contrastive loss. The checkpoint from epoch 99 achieved the lowest validation loss and was selected as the final model before retrieval evaluation. The same UV-Net preprocessing was applied consistently to all CAD models. Evaluation used 128-dimensional UV-Net embeddings with L2 normalization and cosine distance. Labels assigned to the test sets were used only for evaluation and were not used for training or model selection.

The primary evaluation set contains 30 single-solid parts from five part families: bracket, flange, plate, shaft and housing, with six variants per family. Each part is used in turn as a query, and the remaining 29 parts are ranked according to their similarity to the query. The five other parts from the same family are treated as relevant results. To evaluate whether the representations can distinguish structurally different parts with similar overall geometry, we additionally use a stress evaluation set designed to make retrieval more challenging. The stress set contains 54 parts from three families: plate, housing, and flange. Each family includes three structural classes with six dimensional variants per class. For the stress evaluation, a part is considered relevant when it belongs to the same family and structural class as the query. Parts from the same family but a different structural class serve as hard negatives, because they retain similar overall geometry while differing in structural arrangement. The structural classes are defined by different feature arrangements within each part family. Plates have four holes in rectangular, linear or diamond layouts. Housings have one pocket or two pockets aligned along X or Y. Flanges have six bolt holes in regular hexagonal, alternating-radius or two-arc arrangements. Within each structural class, dimensions vary across the six variants, while corresponding variants across classes retain similar outer geometry. For each query, the remaining 53 parts are ranked. For consistency, the same trained UV-Net model and evaluation procedure were used for both evaluation sets. Retrieval performance is measured using mean Precision@K for K = 1, 3, and 5, defined as the average proportion of relevant parts among the top-K results. For the baseline, we additionally measure the separation margin between relevant and irrelevant parts:

where R and I denote the relevant and irrelevant parts for query q, respectively. A positive margin indicates that all relevant parts are closer to the query than any irrelevant part, while a negative margin indicates overlap.

Table I summarizes the retrieval results for both evaluation sets. UV-Net did not outperform the baseline on the primary task. In contrast, on the stress task, baseline margins were negative for 48 of 54 queries, with a median margin of -0.53875. Several structurally different parts also produced nearly identical baseline descriptors, indicating that the simple descriptor could not reliably separate different structural arrangements. These findings demonstrate the value of a learned geometric representation in distinguishing structurally different parts with similar overall geometry, but do not establish universal superiority over simpler geometric representations. The stress set was designed after the primary result and uses a different relevance criterion and a larger candidate set. The two evaluations should not be interpreted as a direct before-and-after comparison. More advanced hand-crafted geometric descriptors were not evaluated in this study. In addition, because the evaluation used a single trained UV-Net model and synthetic CAD parts, the results do not establish performance on real industrial data. Future evaluation should begin with Siemens NX single-part models to test export compatibility and retrieval under real engineering variation. It should then expand to industrial assembly-level retrieval, preserving component identity, placement, hierarchy and interfaces rather than treating an assembly as one solid. At that stage, relevance criteria should be defined with domain engineers at the part, subassembly and assembly levels. Evaluation on previously unseen product families will also be needed to assess generalization. Finally, linking retrieved CAD assets to revision-aware documents can support a separately evaluated engineering RAG system.