Knowledge search your staff can rely on

01Why the demo convinces and production does not

Nearly every knowledge assistant uses one pattern, retrieval-augmented generation (opens external site): look up the passages that fit a question, then let a language model compose an answer based on them. Point it at a tidy folder of fifty documents and it impresses within days.

Real estates are messier. Think of a State agency with two decades of network drives, circulars, intranet pages, case notes and guidance, much of it duplicated, superseded or restricted. There the assistant can fail twice over. A wrong answer makes people give up on it. A wrong answer delivered confidently is worse, because people act on it.

So reliability depends less on the model than on five qualities of the system built around it. Each one can be designed, and each one can be tested.

02Access rights: only what this user may open

The principle fits in one line: never show a person a passage they could not open in the system it came from. Where you enforce that is the hard part. Anything that reaches the model can end up in the answer, so rights have to be applied before generation, during retrieval, for the person asking.

OWASP ranks sensitive information disclosure (opens external site) and vector and embedding weaknesses (opens external site), such as leaks between permission groups, among the main risks of language model applications. For a build team that translates into four rules.

• Attach the source document’s access rights to every chunk in the index, read from the source system instead of being rebuilt by hand.
• Filter by those rights at query time with the signed-in user’s identity, before ranking and before the model sees anything.
• Sync permission changes at least as fast as content changes, and agree how quickly a withdrawn right must stop working.
• Test with actual roles. Whether a new clerical officer can surface the pay scales is a test case, not a lesson for the incident review.

03Sources a reader can check

Every claim should lead back to its passage: title, version or date, owner, and a link that lands on the right spot. Display the passage, not just the file name, so the reader sees at a glance whether it backs up the sentence. Where a policy exists in Irish and English, cite the version the reader asked in, and link the other.

Shape the answer so that unsupported wording is obvious. When no retrieved passage backs a sentence, the system drops it or flags it. Check that automatically too, by comparing each cited passage with the claim it is meant to support and sampling the results for review by a person. A citation that is present but proves nothing is a frequent failure, and an easy one to miss.

04Currency and the authoritative source

Organisations keep many versions of the same rule. Last year’s travel and subsistence circular and this year’s both match a question about mileage rates, and they disagree. Index what separates them: effective date, status (draft, approved, superseded) and owner. Put current approved material first, leave superseded versions out unless someone asks for the history, and show the date in each citation.

Agree for each subject which source wins. HR policy comes from the HR portal, not from an email quoting an older copy. Set how often each source is re-indexed, from close to real time for case data to once a day for an intranet, and watch it: an outdated index gives answers that look plausible and are wrong.

05Admitting a gap

Sometimes the best answer is that there is no answer. Language models can produce smooth statements that nothing supports, which NIST’s generative AI profile calls confabulation (opens external site). Grounding the model in retrieved passages lowers that risk without removing it. OWASP also lists the consequence, people acting on false output, as a risk of its own under misinformation (opens external site).

Pick a relevance threshold under which the assistant says it found no reliable source and points to who can help. On questions of policy, entitlement or obligation, stop the model from filling holes with general knowledge. Track how often it declines, per topic. Early on, that list is frequently the most valuable thing the project delivers, because it shows content owners precisely where guidance is missing.

06What the pipeline looks like

You do not need unusual technology for any of this. You need access rights, metadata and checks to be built into the pipeline from the start, not bolted on once the demo has impressed.

Treat personal data deliberately. Indexing documents still counts as processing the personal data inside them, so the GDPR principles of purpose limitation and data minimisation (opens external site) in Article 5 apply, under the supervision of the Data Protection Commission. Exclude sources the use case does not need, make sure a deletion in a source system also reaches the index, and cover the index when you handle a subject access request. Keep the index hosted in the EU.

07Quality you measure, before go-live and after

A handful of questions tried in a meeting tells you nothing about reliability. Set quality criteria with the content owners and measure every layer on its own, so a problem points to its fix.

Draw the test set from questions people really ask: search logs, service desk tickets, what new starters want to know. A few hundred questions with answers agreed by the content owners is a sound beginning. Rerun all of them after each change, whether a new model, a different chunking approach or an extra source. Quality slips quietly as content changes, and only a stable test set shows the slippage.

08How reliability shows up for the reader

For staff, all of the above becomes a few things they can see for themselves. Answers come only from documents they are entitled to open. Each claim carries a source they can click. They can tell how recent that source is. They hear plainly when no reliable answer exists. And they can flag a wrong answer and watch it get corrected, which tells them someone is measuring.

Answers will still be wrong at times, and some questions belong with a colleague rather than a search box. Being open about that does not weaken the assistant. It is the reason people trust it with everything else.