Retrieval-Augmented Generation (RAG) systems for industrial documentation suffer from a critical preprocessing bottleneck: when hierarchically structured technical documents are segmented into flat text chunks for embedding, individual fragments lose the contextual information necessary for precise semantic matching. This context loss is particularly severe in CNC controller documentation, where the meaning of parameters, alarm codes, and operational procedures is distributed across nested section hierarchies. As a result, conventional chunking pipelines fail to retrieve the correct document for most user queries.

This paper proposes a Graph-Contextualized RAG pipeline that addresses this limitation through three integrated stages. First, the hierarchical structure of the source HTML documentation is parsed and stored in a Neo4j graph database, preserving parent–child relationships between chapters, sections, and content nodes. Second, an LLM analyzes each chunk together with its graph neighborhood, including parent nodes, sibling nodes, and the hierarchical path, to generate a structurally informed context description that is prepended to the chunk prior to embedding. Third, the enriched chunks are indexed and evaluated against a benchmark of 1,000 expert-generated test questions, each mapped to a known ground-truth document, using top-15 retrieval success as the primary metric.

The results demonstrate that structural context is a decisive factor for retrieval performance. The baseline pipeline using conventional flat chunking achieves a top-15 retrieval success rate of only 21%, while the graph-contextualized pipeline achieves 76%, representing an improvement by a factor of 3.6 in retrieval accuracy. These findings provide empirical evidence that explicitly modeling document hierarchy through graph databases and translating structural relationships into LLM-generated chunk context can overcome the semantic ambiguity inherent in flat chunking, enabling reliable RAG systems for complex industrial knowledge bases.

Due to the increasing complexity of production processes, systematic problem-solving is gaining growing relevance.
At the same time, its effectiveness is hampered by structural changes and organizational obstacles.
Advances in the field of artificial intelligence, particularly large language models, offer a promising opportunity to address these obstacles.
The present work focuses on the problem description step of the systematic problem-solving process, as errors in this step can have a particularly negative impact on both the quality and the sustainability of the developed solution.
It is investigated whether large language models are suitable for the evaluation of problem descriptions and can reliably detect missing information therein.
To this end, various LLM models and prompting strategies are compared against human-level performance in a benchmark.
Finally, the practical utility in production is assessed through a usability test.

The semiconductor sector is currently undergoing a dual transition: scaling production through Artificial Intelligence (AI) while navigating stringent global sustainability mandates, such as the EU Carbon Border Adjustment Mechanism (CBAM). To address the fragmented nature of technical, environmental, and governance approaches in this field, this paper presents a scoping review of 1,465 documents from Web of Science and Scopus across thematic domains, including AI-integrated metrology, supply chain ESG, and federated data spaces. Our bibliometric analysis identifies a "core-periphery" structure where machine learning and virtual metrology (VM) are technically central but often decoupled from macro-scale lifecycle orchestration. The study highlights a critical need for interoperable, Safe and Sustainable by Design (SSbD) architectures that move beyond isolated technical fixes to system-wide governance. We propose a six-layer SSbD architecture—integrating substitution, RegTech compliance, and federated data—to align industrial analytics with AI safety and Scope 3 emissions reporting. By synthesizing technology management perspectives with systems engineering, this research establishes a foundation for provenance-aware data fabrics and risk-aware model governance, providing a pathway for climate-neutral and circular innovation in semiconductor manufacturing.

With innovative concepts, Indian startups have been making good strides. However, the facts revealed that majority of start-ups fail before of their limited focus on Minimum Viable Product (MVP) development. This study aims to investigate important aspects of Minimum Viable Product (MVP) development and its importance to make startup successful. Drawing on interviews with founders of startup, this study provides a comprehensive understanding of the various elements that contribute to Minimum Viable Product (MVP) development for success of Indian start-ups. Using thematic analysis approach, this study identified the key components of MVP development such as Capability - trust based network evaluation, Customer-centric market validation, AI enabled Product Development, Human - AI Collaborative Capability and Market sensing and customer adoption. The findings of this study offer new insights for improving MVP development, thereby enhancing the survival chances of Indian startups. A future study could be carried out using different methodology in varied setting to extend the findings of the study.