An AI assistant that answers from your knowledge, not the internet's
Centricone Technologies builds retrieval-augmented assistants over your documents, tickets, wikis, and records — answering with citations, respecting the permissions the source systems already enforce, and scored against a test set you own.
The hard part of a knowledge assistant is not the answer, it is the citation and the permission. We build both from the first week, because retrofitting either one means starting again.
support coverage across US and Canada time zones
of code reviewed, tested, and documented before release
Everyone has the documents. Nobody can find the answer
The knowledge is spread across six places
A wiki, a drive, the ticket history, and three people's inboxes. The answer exists; finding it costs an hour and somebody else's interruption.
The pilot answered confidently and wrongly
No citations, no retrieval over the real corpus, and no way to tell a good answer from a fluent one — so trust broke on the first bad response and never came back.
It would show people things they should not see
Permissions live in the source systems, the index flattened them, and the security review stopped the rollout at exactly the right moment.
for teams whose answers are already written down
Centricone covers the whole assistant: connecting and indexing the sources, retrieval and ranking, permission-aware answering with citations, the evaluation set, and the reporting that shows what people actually ask.
Deliverables
What you get
- A working assistant over your own corpus
- Connectors and an index that stays in sync
- Permission checks enforced at retrieval time
- Citations on every answer
- A scored evaluation set, in your repository
- Usage, cost, and failed-question reporting
How an assistant engagement runs
Pick the corpus and the questions
The questions worth answering and the sources that hold them. Narrow beats complete — one team answered well is worth more than everything indexed badly.
Index, retrieve, and measure
Ingestion, chunking, and retrieval built against an evaluation set from the first week, before any effort goes into the prompt.
Permissions and citations before the pilot
Access checks and source links wired in early, because these are what the security review and the first sceptical user will both test.
Pilot with one team, then widen
A group whose questions we can score, feedback read weekly, and the corpus grown where the failures point rather than where the drive is biggest.
When a knowledge assistant is the wrong build
This is the most requested AI build in the market right now, which is exactly why it keeps getting bought when something else was needed.
- The knowledge is not written down anywhere. Retrieval cannot find what does not exist — that is a documentation project first.
- You need the system to take actions rather than answer questions. That is an AI agent.
- The real task is pulling structured fields out of forms and invoices at volume. Document intelligence does that far better than a chat box.
- One team, one folder, and a decent search would solve it. Sometimes the honest answer is better search, not a model.
Why teams build assistants with Centricone
An evaluation set from week one, not after launch
Permissions enforced at retrieval, never bolted on
Every answer traceable to a source you can open
Your corpus, your index, your cloud accounts
The capability behind this engagement
Retrieval architecture, model selection and evaluation, guardrails, and the cost controls that decide whether a pilot becomes a product — covered in more depth on the AI development services page.
Frequently asked questions
Not sure this is the right shape? Tell us the situation and we’ll say which engagement fits.