The case for grounding AI in verified source documentation

The hallucination problem

Ask a general purpose AI model how to diagnose a hydraulic pressure fault on a haul truck and you'll get an answer. It will be well structured, technically plausible, and delivered with confidence. It might also be entirely fabricated. The model has no access to your maintenance documentation. It's generating a response based on patterns in its training data, which may include forum posts, generic maintenance guides, and content from machines that have nothing in common with yours.

In a consumer context, this is an inconvenience. In a mining maintenance context, it's a safety risk. A fabricated torque specification, an incorrect wiring diagram, or a diagnostic procedure that skips a critical isolation step could result in equipment damage, injury, or worse. The confident tone of AI generated text makes it particularly dangerous because it reads as authoritative even when it has no basis in fact.

Generic AI vs grounded AI

The distinction between generic and grounded AI is fundamental. Generic AI draws from its training data, which is broad but unverified for your specific context. It knows about hydraulic systems in general but not about your hydraulic system on your machine with your configuration. Grounded AI, by contrast, is constrained to answer only from a defined set of verified source documents. It searches your technical library first, then generates a response based on what it finds.

This is not a subtle difference. It's the difference between an AI that says “the typical hydraulic pressure for this class of truck is 3,500 psi” and one that says “according to the shop manual, section 30, page 47, the main relief pressure should be set to 32.4 MPa (4,700 psi) at 2,100 rpm.” The first sounds helpful. The second is actually useful because you can verify it.

Why source citation matters

In any safety critical environment, trust is not built on confidence. It's built on traceability. A maintenance procedure is only as reliable as its source. When AI generates a diagnostic workflow, every step needs to be traceable to a specific document, a specific section, and a specific specification. If the AI can't cite where it found the information, the technician has no way to verify it, and neither does the maintenance manager, the safety team, or an investigator after an incident.

Source citation also creates accountability. When you can trace an AI generated recommendation back to a specific paragraph in a verified source document, you know where the recommendation came from and whether it was applied correctly. This is essential for compliance, for continuous improvement, and for building organisational trust in AI as a tool.

How RAG pipelines work

The technical approach that makes grounded AI possible is called Retrieval Augmented Generation, or RAG. The concept is straightforward, even if the implementation is complex. Instead of asking an AI model to answer from memory, you first search a curated document library for relevant content, then provide that content to the model as context, and then ask the model to generate a response based only on what was provided.

The pipeline typically involves several stages. Documents are ingested, split into meaningful chunks, and converted into numerical representations called embeddings. When a query comes in, the system finds the most relevant chunks using vector similarity search. The best matches are then passed to a language model along with the original question. The model generates a response, and the system tags each part of the response with its source document and location.

The quality of the output depends heavily on how well the documents are processed. Technical maintenance documentation is complex. It contains tables, diagrams, cross references, and conditional logic. A naive chunking approach that splits documents every 500 words will lose the context that makes procedures understandable. Effective RAG pipelines need to understand document structure, preserve context across chunks, and handle the specific challenges of maintenance documentation.

What happens when grounding is insufficient

A well designed grounded AI system needs to know what it doesn't know. If the documentation doesn't contain enough information to confidently answer a question, the system should say so explicitly rather than filling the gap with assumptions. This is harder than it sounds. Language models are trained to be helpful, which means they have a tendency to generate answers even when the evidence is thin.

The countermeasure is confidence scoring. By measuring how well the retrieved documents match the query, and how much of the generated response can be traced to source material, the system can assign a grounding confidence score. If the score is below a threshold, the response is flagged as low confidence, and the technician knows to verify independently or escalate.

AI as a tool, not an oracle

The goal of grounded AI in maintenance isn't to replace human judgment. It's to put the right information in front of the right person at the right time, with a clear trail back to where that information came from. A technician who can see the relevant procedure, cited and sourced, makes better decisions faster than one who is guessing or scrolling through a 3,000 page PDF. That's the value proposition. Not AI that knows everything, but AI that can find exactly what your documentation says and show you where it found it.