Knowledge Graphs for Manufacturing RAG
Why vector similarity alone struggles with industrial questions, and what a structural layer over processes, equipment and documents adds to retrieval.
Research question
Does adding an explicit graph of processes, equipment, components and documents improve retrieval quality for engineering questions compared with vector search alone?
Background
Industrial questions are rarely about a single passage of text. "Which specification controls this weld parameter on this line?" requires knowing that a line contains a station, the station runs a process, and the process is governed by a document revision.
Vector retrieval finds text that reads similar to the question. It has no representation of that chain of relationships, so it retrieves plausible passages from the wrong equipment or an outdated revision.
Approach
We modelled a single production area as a graph with process, equipment, component, control-parameter and document nodes, then used the graph to constrain retrieval: resolve the entities in the question first, then retrieve passages only from documents attached to those entities.
- Manual schema definition for one production area.
- Semi-automated ingestion of documents and equipment records.
- Entity resolution step ahead of vector retrieval.
- Comparison against unconstrained vector retrieval on the same question set.
Observations
- Graph-constrained retrieval returned fewer passages, and the passages were far more often attached to the correct equipment.
- Questions that named equipment explicitly benefited most; open-ended questions benefited least.
- Document revision relationships turned out to matter as much as equipment relationships.
- Schema maintenance is the real cost — the ingestion, not the retrieval, is where the effort sits.
Limitations
- A single production area and a small internally-written question set; no external benchmark.
- Retrieval quality was assessed by engineering review, not by a formal scored evaluation.
- Ingestion was partly manual, so results do not reflect a fully automated pipeline.
Next steps
- Extend the schema to quality records and link them to process parameters.
- Automate ingestion for one document family and measure schema drift.
- Define a repeatable scored evaluation before making any comparative claims.
Related labs
Appeared in