Answers with a source behind them, and permissions in front.
A knowledge assistant is a retrieval-augmented system that answers questions from your own systems of record — documents, tickets, policies, tables — and cites the exact source behind every answer. It respects the permissions the person already has, so two employees asking the same question can correctly get different answers.
- The hard part is not the model. It is access control, freshness and proving the answer is current.
- Role-aware retrieval means the assistant can never surface a document the person couldn't already open.
- On a governed deployment in an enterprise team: 60–70% less time spent searching, 40% faster support resolution.
Six questions your systems already know the answer to
Every one of these is a question an employee asks weekly and answers by interrupting somebody senior.
"Has this customer hit this problem before?"
Tickets, release notes, runbooks and past resolutions in one answer, with the ticket number attached. New agents stop escalating the fifth repeat of a known issue.
"What did we commit to the last time we answered this?"
Past RFP responses, current security posture and approved claims, retrieved by question rather than by folder. Bid teams stop rewriting answers the company has already approved.
"Which policy applies to me, in my country?"
Leave, expense, benefits and mobility policies differ by entity and region. The assistant answers for the person asking, not for the handbook in general, and links the clause.
"What are our terms with this supplier?"
Contract clauses, payment terms, renewal dates and prior amendments, retrieved from the executed documents rather than from someone's memory of the negotiation.
"How do we do this here?"
Architecture decisions, runbooks and incident history, answered with the decision record attached. New engineers ramp against the written record instead of the nearest senior.
"Is this answer still current and authorized?"
Where a stale answer creates exposure, retrieval is scoped to approved and in-date sources, and superseded documents are visibly marked rather than quietly ranked lower.
Five mechanisms, in this order
- Ingestion that keeps the permissionsFolder- and role-level access controls travel with the content into the index. Access is checked at query time, not filtered after the fact.
- Retrieval tuned on your questionsWe collect the questions people actually ask, then tune chunking, hybrid search and reranking against that set rather than against a benchmark.
- Citations to the object, not the corpusEvery answer points at the page and source object it came from, so the reader can verify in one click and stop trusting the summary blindly.
- Freshness and supersessionRe-indexing on change, with superseded documents marked. An assistant that confidently quotes last year's policy is worse than no assistant.
- Evaluation and escalation loopsA labelled answer set, thumbs data from real users, and a path to a human when retrieval returns nothing good enough to answer with.
A knowledge assistant that survived the security review
Client names and locations are withheld; figures are drawn from delivery records and approved business cases.
A permission-aware knowledge assistant
Critical answers lived across SharePoint, Confluence, policies, tickets and tribal knowledge. Employees spent hours searching and still couldn't verify whether an answer was current or authorized.
- Multi-source ingestion with folder- and role-level access controls, spanning documents, tables and unstructured text
- Citations to the exact page and source object behind every answer
- Feedback, evaluation and escalation loops that improve the assistant over time
Delivered by a Production Pod
Read the full case →The four questions that decide the project
- Can it surface a document this person couldn't already open? No — access is enforced at query time.
- Where does the data sit? Inside your cloud account, region-bound, tenant-isolated.
- Is anything used to train a model? No. Your data never leaves your boundary.
- Can we audit an answer six months later? Yes — inputs, sources and version are logged per response.
Three situations where retrieval is the wrong tool
- The source of truth is contradictory. If four documents disagree and nobody owns which one wins, an assistant will surface the disagreement faster. That is useful, but it is a governance project first.
- The answer requires calculation, not recall. Questions that need a live query against a database belong in an interface over that database, not in a retrieval layer.
- The corpus is thin. Below a few thousand documents, better search and a tidy intranet usually beat anything we would build.
FAQ
What to ask before you build a knowledge assistant
What is RAG, and why does it matter here?
Retrieval-augmented generation means the model answers from documents retrieved out of your systems at question time rather than from what it memorized in training. That is what makes citations possible, keeps answers current when a policy changes, and stops the assistant inventing a plausible answer it has no source for.
How do you stop it showing people things they shouldn't see?
Access controls travel with the content into the index and are enforced at query time against the person asking. The assistant cannot surface a document that person couldn't already open. Two employees asking the same question can correctly get different answers.
What happens when the underlying document changes?
Sources are re-indexed on change and superseded versions are marked rather than quietly deprioritized. Freshness is part of the evaluation set, so a regression that starts quoting last year's policy fails the gate before it ships.
How do we know the answers are good enough to trust?
We build a labelled set from the questions your people actually ask and measure retrieval and answer quality against it before launch. That set becomes the regression gate for every release, alongside user feedback and an escalation path when retrieval finds nothing solid.
Most AI pilots die in the demo. Ours don't.
We define one measurable metric before we write a line of code. If it hasn't moved inside the impact window, the pod keeps building until it does, at no additional fee. You never pay for AI that just sits there.
Scope a Proof Pod →One workflow. Real data. A number that moves.
Bring the question your team asks most often and answers slowest. We will scope it in 30 minutes.
Scope a Proof Pod →