How it works, no smoke
This page is the training part. If you only want to know what it does, go to Use cases. If you want to understand why it works, keep reading.
What an AI assistant (really) is
At the centre, the language model: it reasons and writes, that's all. Around it, what makes it useful. Context: documents and data. Tools: APIs, SQL, OPC UA. Memory: the conversation and the history. Guardrails: permissions, limits and audit. Interface: chat, HMI, Teams or voice.
The model is the most famous part, but not the most important one. A mediocre model with good context and good tools always beats a brilliant model that doesn't know where it is.
From classic RAG to agentic RAG
RAG is an ugly name for a simple idea: before answering, look it up. Classic RAG is a straight path: question, index lookup, answer. For looking up a manual it works beautifully.
The problem shows up when the answer isn't in any single document but scattered across several places. “Why did the line stop last night?” isn't in a PDF: it's in the historian, the alarm list and a ticket someone closed in March.
Agentic RAG plans first: it breaks the question into steps and decides which source to query at each one. Then it verifies: is there evidence? If not, it queries again. It's what a senior technician does: checks the manual, checks the history, and if it doesn't add up, checks again.
| Classic RAG | Agentic RAG | |
|---|---|---|
| Best for | FAQ, looking up a manual | Cross-referencing, diagnosing |
| Model calls | 1 | Several |
| Cost and latency | Low | Medium |
| Rule of thumb | Start here | Move here when the answer requires cross-referencing |
Rule of thumb: start classic where the question is simple and go agentic where the answer requires cross-referencing. You don't need an agent to look up a torque value.
An ordinary Monday: line 3's OEE
Translates the question into queries
text → GraphQL / SQL
query { oee(line: 3) }Runs and cross-references
OEE, micro-stops, alarms, work orders
Alarms 23:10–05:40Correlates and detects the pattern
station 4 after every changeover
OEE_L3_turnoExplains it with evidence
in operator language
SOP-114.pdf p.6OEE by shift
Micro-stops per station · night shift
Illustrative data · the assistant builds these views from the queries; they are not pre-made.
Nobody designed that chart beforehand. The shift report someone writes half-asleep at six in the morning will soon be drafted by the assistant and validated by the person. Value shifts from knowing how to pull the data to knowing how to ask the question.
What it feeds on
Six families of sources you already have. What's new isn't the list: it's crossing them. The magic is the moment the assistant pairs a historian trend with a seven-month-old ticket and tells you: “this happened before, and this is how it was fixed.”
Before the model
Four conditions that come before choosing any LLM. Many AI projects fail because they start with the icing.
- Synchronised data: same time base and batch context, with a plant data model.
- Inherited permissions: the assistant sees what the user sees.
- A network that allows it: reads from the DMZ or IT level, never from the fieldbus.
- Data quality: without it there is no reliable root cause.
The model is the icing. This is the cake.