In DevelopmentIndustrial RAGUpdated August 19, 2026
Industrial Knowledge Engine
A prototype combining RAG, engineering documentation, structured manufacturing data and AI models to create an industrial knowledge assistant.
Problem
Engineering and operational knowledge is spread across manuals, specifications, work instructions, quality records and people. Finding the controlling information is often a larger daily cost than analysing it.
Hypothesis
If retrieval is grounded in both documents and structured process context, an assistant can answer engineering questions with citations that engineers are willing to trust.
Architecture
- Ingestion of engineering documents with structure-aware parsing.
- Structured process and equipment context from manufacturing systems.
- Embedding and reranking stages ahead of generation.
- Answer composition with explicit source citations and revision awareness.
- Deployment options spanning private cloud, customer infrastructure and local GPU.
Experiment
Whether structure-aware chunking plus a reranking stage reduces wrong-source and wrong-revision answers on a defined engineering question set.
Observations
- Keeping tables and numbered steps intact removed a recurring class of wrong-parameter answers.
- Citations mattered more to reviewer trust than answer fluency.
- Revision handling has to be explicit; otherwise superseded documents surface as valid answers.
Technologies
- Retrieval-augmented generation
- Vector search
- Rerankers
- Document parsing
- Local and hosted language models
Next steps
- Add quality records as a retrievable source.
- Define a repeatable scored evaluation before publishing comparative results.
- Test the full stack on local inference hardware end to end.
Related research
Related news
