BSKLab Brief — August 2026

Most of the month went into the Industrial Knowledge Engine: rebuilding the retrieval layer so engineering documents, equipment records and process parameters are chunked around the structure of the document rather than a fixed token window.

We also spent time on the boundary question — which parts of an industrial AI stack can sit in a private cloud and which have to stay inside the plant network. That shaped the deployment options we now present in every architecture brief.

• Compared local embedding models on a small set of maintenance manuals and work instructions.
• Built a first pass at a manufacturing knowledge graph schema covering process, equipment, component and document nodes.
• Tested a reranking step in front of the generation model to reduce irrelevant passages in engineering answers.

Local inference keeps getting more practical. Quantised models running on a single workstation-class GPU are now good enough for retrieval-grounded question answering over technical documentation, which changes the conversation with plants that cannot send data outside their network.

• Extend the knowledge graph prototype to link quality records to process parameters.
• Start a predictive quality experiment using upstream process variables.
• Publish a research note on chunking strategies for engineering documents.