Explainer
RAG on the Hangar Floor: What Airworthy Retrieval Requires
The Demo vs the Hangar
Correct Isn't the Same as Approved
Retrieval-augmented generation is the right architecture for “ask our manuals a question,” and vendor claims of drastic search-time reductions aren't fantasy — finding the right procedure in a large library genuinely is slow, and retrieval genuinely is fast. But aviation maintenance raises a bar most demos never meet: the answer's pedigree. A usable answer must come from approved data at the current revision, with service bulletins incorporated; must respect effectivity — this aircraft, this configuration; and must be traceable after the fact, because “what was the technician looking at” is a question investigations ask with dates attached.
Point a RAG stack at a folder of exported PDFs and it will fail all three — confidently, with citations, quoting a superseded revision to a mechanic working the wrong configuration. The model didn't fail. The library did.
The Foundation
Airworthy Retrieval Is a Content Property
Everything airworthy retrieval needs is a property of the content layer, and the list should look familiar: one approved source rather than a scraped share drive; version and supersession awareness, so retrieval targets current content and can answer about historical states; applicability honored so answers respect configuration; page-level addresses so a citation opens the actual page; permissions inherited from the source; events captured so the record exists. That is, precisely, a content services platform doing its ordinary job — our Content Portal API exposes exactly this shape, which is why we describe it as ready for grounded AI rather than shipping a chatbot.
So the sequencing advice stays what it's been all series: foundation first, model last. The platform pays for itself in ordinary operations before any AI arrives; the retrieval pilot then runs on a bounded corpus with citations mandatory and the source page one tap away. And the framing that keeps everyone honest: the assistant is a faster index. The manual remains the authority — the win is making the authority faster to reach, not optional to consult.
FAQ
Questions We Hear
What is RAG, in one paragraph?
Retrieval-Augmented Generation: instead of answering from its training data, the AI first retrieves relevant passages from a defined content set — your manuals — and generates its answer from those passages, ideally with citations. It's the standard architecture for 'ask questions of our documents,' and it genuinely works. The catch is that the generation is only as trustworthy as the retrieval, and the retrieval is only as trustworthy as the content set underneath it.
Why is aviation maintenance the hardest test for it?
Because the answer's pedigree matters as much as its correctness. A maintenance answer must come from approved data — the current revision, with applicable service bulletins and airworthiness directives incorporated — must respect effectivity (which aircraft, which configuration), and must be traceable, because 'what was the technician looking at' is a question incident investigations actually ask. A RAG demo over a folder of exported PDFs satisfies none of that: it will happily retrieve a superseded revision or the wrong effectivity, with total confidence and a nice citation.
What does the content layer need to provide?
Exactly what a good delivery platform provides anyway: one approved source (not a scraped folder), version and supersession awareness so retrieval targets what is current — and can answer about what was current on a date; applicability so answers respect configuration; page-level addressability so citations point at a real, openable location; permission scoping inherited from the source; and event capture so there's a record of what was retrieved and shown. Get those, and the AI layer on top is almost routine engineering.
Should operators wait until the tools mature?
The models are already good enough for the retrieval-and-summarize job; waiting doesn't change the real gating item, which is the content foundation. The pragmatic order: put the library on a platform with versions, applicability, permissions, and APIs — which pays for itself in ordinary operations immediately — then pilot retrieval on a bounded corpus with citations mandatory and the source page one tap away. Treat the assistant as a faster index, not an authority: the manual remains the authority, and a good implementation makes reaching it faster, not optional.
Get In Touch
Grounded, or Just Confident?
If your AI pilot answers from a folder of PDFs, it's one superseded revision away from a very bad day. We build the approved-content foundation first — then the assistant earns its place.