JapanManufacturing2025 – ongoingKomatsu
AI-assisted engineering knowledge retrieval
Komatsu chose search-focused AI and raised retrieval recall from 0.775 to 0.888 in tests
Problem
Komatsu's designers struggled to find relevant technical knowledge through exact-match searches. Less experienced engineers often lacked the terminology needed to search effectively.
Approach
Komatsu combined AI-OCR and language models to structure technical documents, with designers checking the results. It tested summary-based and search-focused interfaces with designers, then selected search support because summaries produced plausible errors. Hybrid retrieval followed by language-model reranking improved measured recall.
Results
Takeaway
Designers preferred ranked references that helped them decide which original documents to read. Generating an answer introduced errors that reduced its practical usefulness.
Categories
- Size
- Large enterprise
- Industry
- Manufacturing
- Business function
- IT and R&DInternal work and knowledge management
- Innovation type
- Process innovation
- AI technology
- Generative AIComputer vision
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
Komatsu documented designer evaluations and retrieval experiments for its engineering knowledge system. Recall increased from 0.775 to 0.888 with simple-prompt reranking; designer testing favored search support over generated summaries.
Full details are in the original sources.