Retrieval-augmented generation is how a general model answers questions about your plant — without retraining it, and with a citation you can check.
The problem it solves
A general-purpose model knows nothing about your equipment, your procedures, or the revision of the work instruction currently on the floor. Retraining it on your material is expensive, has to be repeated every time a document changes, and produces a model whose answers still cannot be traced to a source.
Retrieval-augmented generation takes a different route. Leave the model alone. When a question arrives, find the relevant passages in your own documents first, put them into the prompt alongside the question, and ask the model to answer from that material. The knowledge lives in a library you maintain, not in weights someone else trained.
How it works, in four steps
Prepare the sources. Manuals, work instructions, specifications, maintenance records. Each is split into passages and converted by an embedding model into a numeric vector that captures its meaning, then stored in a vector database.
Retrieve. The incoming question is converted the same way, and the closest passages are pulled back. Closeness here is semantic, not keyword-based: "how much leave do I have" finds the annual leave policy without either phrase matching literally.
Augment. Those passages are inserted into the prompt as context, with an instruction to answer from them and to say so when they are insufficient.
Keep it current. When a document changes, its passages and their vectors have to be regenerated. This is the step teams skip, and it is the one that decides whether the system is trustworthy in month six.
A model grounded in a superseded revision is worse than no system at all — it is wrong in a well-formatted way, with a citation.
Why citations are the point
The output should carry references to the passages it drew on. This is not presentation polish. It is what converts an answer a reader has to trust into an answer a reader can check, and it changes who can safely use the system.
It also gives you a debugging surface. When an answer is wrong, the citation tells you whether the retrieval found the wrong passage or the model misread the right one. Those are different faults with different fixes, and without citations you cannot tell them apart.
What it does not fix
Retrieval controls what the model sees. It does not stop the model misreading what it was given, and it does not help when the answer is not in your documents at all.
It also inherits every problem in the source material. If two work instructions disagree, the system will confidently present whichever it retrieved. If access control is not enforced at retrieval time — so that the model can only reach documents the person asking is entitled to see — then you have built a mechanism for reading other people's files in natural language. Decide that before the pilot, not after.
Written by the software delivery team. Published articles carry a named author once attribution is confirmed.
