Artificial intelligence is beginning to assist engineers in ways that go beyond general chat tools. A recent paper by Salahuddin Alawadhi and Noorhan Abbas, published in Computer Science & Information Technology (Vol. 15, No. 9, 2025), tested whether Retrieval-Augmented Generation (RAG) could help electrical engineers interpret technical manuals from ABB’s Emax series circuit breakers.
RAG combines two functions: a search system that retrieves relevant information from documents and a language model that generates an answer from that material. In principle, this should let an engineer ask a question in plain language and receive a response drawn directly from product documentation.
Why circuit breaker data was used
The study focused on ABB’s Emax E1.2 low-voltage air circuit breakers, equipment used in data centers and hospitals. Twelve ABB manuals and guides—together about 540 thousand tokens of text—were processed into a dataset. These documents included installation guides, ratings, and disassembly procedures that practicing engineers use to configure and maintain breakers.
How the researchers built the test system
The authors compared three RAG pipelines:
- Cohere using the command-xlarge-nightly model
- OpenAI GPT-4o with the text-embedding-ada-002 embedding model
- Anthropic Claude 3.5 Sonnet using Voyage AI’s voyage-3 embeddings
All models retrieved text from a Chroma vector database and generated answers with temperature = 0 to avoid random variation. To see how document segmentation affected accuracy, they tried three “chunking” methods:
- Basic – split text purely by length.
- Paragraph per page – break at page and paragraph boundaries.
- By title – start a new chunk when a section heading appears.
Performance was measured with RAGAS metrics: context precision, context recall, answer relevance, and faithfulness (how closely the answer matched the source text).
What performed best
Claude’s “paragraph per page” configuration achieved the highest factual accuracy, with a faithfulness score of 0.8556. GPT-4o’s “by title” method produced the most relevant answers but missed some supporting context. Cohere’s “paragraph per page” method retrieved the widest range of information yet occasionally lost precision. Simple length-based chunking performed worst across all models.
Example question and limits found
When asked how to activate the terminating resistance on an ABB Ekip Com Modbus RTU module, the best models correctly stated that a 120-ohm resistor is enabled with dip-switches. None specified which switches to use, though the answer was available in the source manual. This gap illustrated how retrieval errors or incomplete chunks can cause AI systems to omit critical details even when the data exists.
Technical improvements proposed
The paper recommended hybrid chunking (combining title- and paragraph-based methods), reranking retrieved passages with specialized models, and allowing iterative follow-up queries when the system is uncertain. Future versions could use knowledge graphs—structured maps of relationships among breaker models, ratings, and standards—and multimodal inputs so the AI can interpret tables or diagrams as well as text.
Why accuracy still matters
Engineering work tolerates little ambiguity. A system that is 80 percent correct may still produce unsafe recommendations. The authors noted ethical concerns around transparency and verification: engineers must be able to trace an AI’s statement back to a document and confirm it. Until that traceability improves, RAG should be treated as a reference aid rather than a source of final authority.
Source: Salahuddin Alawadhi and Noorhan Abbas, “Optimizing Retrieval-Augmented Generation for Electrical Engineering: A Case Study on ABB Circuit Breakers,” Computer Science & Information Technology, Vol. 15, No. 9 (2025), pp. 59–77. DOI: 10.5121/csit.2025.150905.
