Evaluating Source-Grounded AI Technical Assistants for Higher Education: Technical Accuracy, Knowledge Boundaries, and Repeatability in Printing Technology

Source-grounded generative AI assistants can improve access to disciplinary literature, but citations do not ensure technical correctness. We evaluated two textile- and offset-printing assistants in Google NotebookLM (now Gemini Notebook), using 20 controlled prompts per assistant. One subject-matter expert rated technical accuracy, completeness, source support, and contextualization and manually checked citations. Mean overall scores were 4.73/5 and 4.61/5, respectively; because prompt sets differed, these are descriptive, not comparative. Source support exceeded technical accuracy in 35 of 40 original responses. In eight explicitly prompted knowledge-boundary tests, six responses showed appropriate abstention and two showed abstention with reservations; this evaluates prompted handling, not spontaneous recognition. Of seven repeated pairs with preserved identical wording, six received unchanged overall scores (mean absolute difference, 0.036/5); content comparison found recurring errors and differences across the ten pairs, although three offset pairs were only same-topic comparisons. A recurring textile error labeled source-reported heat-setting temperatures as glass-transition temperatures; a reported fabric-temperature rise applies only to vat-dye all-in cotton printing in saturated steam. With one expert evaluator and 15 original offset prompts preserved only as topic labels, this is an exploratory case study, not a validated benchmark. Student learning outcomes were not assessed.