Verifiable RAG: cite or abstain, never from memory
Verifiable RAG cites or abstains: it delivers the whole procedure with its citation, or it tells you it has no support. It never answers from memory and never hands over half a procedure. In operations, where a missing step is a hazard, that’s the difference between a useful assistant and a dangerous one.
Why does generic RAG fail in operations?
RAG (retrieval-augmented generation) is the technique that lets an AI assistant search your documents and answer with what it finds. In an office demo it works. In a control room or next to a machine, it fails in three ways.
It answers with fragments
The usual approach chops documents into chunks and gives the model the ones that look most like the question. A 12-step procedure ends up split across several chunks, the assistant receives steps 3 to 7, and it answers with those.
Half a procedure is worse than none. Whoever reads it doesn’t know steps are missing, and neither does the assistant.
It makes things up with the same confidence it has when it’s right
When the search finds nothing relevant, the model answers anyway, from what it learnt in training. Same tone, same format. Nobody can tell the answer backed by the manual from the one that was invented.
It doesn’t say where the answer comes from
Without a citation to document, revision and section, nobody can check whether the answer comes from the current revision of the manual or from one withdrawn two years ago.
Add two conditions of real operations: many manuals can’t leave your network, and in the field there isn’t always coverage.
Why does this matter now?
Because the knowledge sits with people who are leaving. In energy, there are 2.4 workers close to retirement for every new entrant under 25 (IEA, December 2025). The urban water sector names the loss of know-how as one of the challenges of generational turnover (DAQUAS, October 2026).
Your technicians call “the one who knows”, and the one who knows is retiring. Much of what they know is in the manuals. An assistant that reads them properly makes it available to the whole shift. One that reads them badly is a new risk.
And because more and more organisations want their data kept close. 61% of CIOs in Western Europe will increase their reliance on local cloud providers because of geopolitics (Gartner, November 2025). If your manuals describe how to run an infrastructure, you probably don’t want to upload them to a cloud assistant either.
What rules should RAG follow in operations?
These four, without exception:
- Cite or abstain. Every answer carries document, revision and section. If there is no support, it says so.
- The whole procedure or nothing. If the answer is a procedure, it arrives complete.
- Never from memory. The assistant answers with what the index hands it, not with what it remembers from training.
- Never degrade silently. If something fails, the system says so.
And here is the difference, point by point, with the naive approach:
| Aspect | Naive RAG | Verifiable RAG |
|---|---|---|
| What the assistant receives | The chunks that look most like the question | The whole section |
| When there is no support | The model answers anyway | It says so |
| Citation | Optional | Document, revision and section |
| What goes into the index | Whatever is uploaded | What passes a quality check and your approval |
| When an update goes wrong | No way back | You roll back to the previous index version |
How does RAGNIX do it?
RAGNIX, our document retrieval engine, applies those rules in three steps.
Controlled ingestion
Today it takes Markdown, Word and HTML, and extracts the figures of each revision. Nothing goes in without passing a quality check and your approval, and every change is logged. You decide which documents feed the assistant, not whoever has access to a shared folder.
A local, versioned, immutable index
Search combines exact words and meaning. Exact words find an equipment code or an alarm number; meaning finds the answer even when the technician asks in their own words.
Every version of the index is immutable. If a document update goes wrong, you roll back to the previous one.
Answers over MCP, with the whole section
Your assistant connects over MCP, searches and receives the whole section with its citation. It can also browse the document’s table of contents to get to the section it needs. You don’t need a specific model: RAGNIX connects over MCP to whichever assistant you already use.
The citation isn’t decoration. It’s what lets your technician open the manual at the right section and check the answer before acting. The AI proposes; the person checks and decides.
Does it work offline and with small models?
Yes, that’s what it’s designed for. RAGNIX works with small, local language models, and indexing, search and serving all run inside your network. It sends us no telemetry.
It’s an 8.5 MB executable that runs on a laptop or a server, with no vector database, no Python and no containers. It fits on an air-gapped network or on a technician’s laptop.
Access control goes down to the fragment: collections are isolated and restricted content is hidden by default. You can keep several collections on the same system without them mixing. And there is no per-query cost and no dependence on a cloud AI provider.
How much does it cost compared with sending the whole manual to a long-context model?
The alternative to RAG is to put the whole manual into a long-context cloud model and ask it questions. It works, but you pay to re-read the entire manual on every question.
We have an estimate, not a measurement. With a 200,000-token manual, a two-question session costs about 24 times less with RAGNIX than sending the whole manual to a long-context model in the cloud. Assumptions: public prices from September 2026, output tokens not included. The experiment with a real manual is under way.
What we don’t promise: better answers than putting the whole manual into a giant cloud model. What we do promise is that, with several manuals, offline and with a small model, the answer still arrives with its citation.
What stage is RAGNIX at?
In pilot with a first customer. Today it indexes Markdown, Word and HTML and extracts the figures of each revision.
On the way: PDF, serving figures over MCP, encryption and authentication of the MCP channel, and access with your company’s SSO. We say so up front because, if your documentation lives in PDF, you should know before the first meeting.
Where do you start?
With a manual your technicians look up often and a list of real questions. It’s the same idea as accountable AI, applied to your documents: every answer comes with its evidence.
If you want to try it with one of your manuals, see how RAGNIX works; if it’s part of a bigger project, tell us about it on our services page.