01
Internal knowledge assistant
Staff ask questions in plain English and get answers drawn from your handbooks, procedures and policies, with the source attached.
Knowledge assistants & RAG · UK
Knowledge assistants.Answers from your own information.
Give your team, or your customers, a reliable way to ask questions of your documents, policies and records. Answers are grounded in approved material, respect who can see what and show their sources, so people can check before they act.
Free first call · Fixed price after scoping · Sometimes the answer is a simpler tool
Can we accept payment by card for a new commercial client?
- Check the user’s access
- Search approved policies
- Draft an answer from matching sections
- Attach sources and flag gaps
Answer drafted from the finance policy, section 3.2.
Source linked. If the policy doesn’t cover it, the assistant says so.
Answers with sourcesAccess permissions respectedSays so when it doesn’t know
How I test
A question set,
run before anyone relies on it.
Before launch and after every change, I run the assistant against a set of real questions from your own work. Each has an expected source and an expected outcome, including questions it should decline. This is the format, not a result.
| Question | Expected source | Outcome |
|---|---|---|
| What information do we need before a project starts? | Project onboarding handbook, section 2 | Answered with source |
| Can a contractor see another client’s pricing? | Access policy, rule 4 | Answered with source, access respected |
| What is our policy on a topic the documents do not cover? | None expected | Declined as unsure |
Illustrative test format and data. Not a client system.
Read the free RAG guide for the question-set method.
Is it right for you?
When a knowledge
assistant helps.
The value comes from useful written material and repeated questions. Without those, a better search or a documentation project may be the right first step.
A good fit when
- Answers depend on a few experienced people or a long search through shared drives
- You have written material worth using: policies, procedures, manuals, contracts or past work
- Staff or customers ask similar questions again and again
- Getting an answer wrong has a cost, so sources need to be visible
Probably not yet when
- The knowledge isn’t written down yet. That’s a documentation project first
- Documents are badly out of date and nobody owns them
- Better search in a tool you already have would solve the problem
What I build
Company knowledge,
where people need it.
The same core technology serves different audiences. What changes is the source material, the permissions and where the answers appear.
02
Client and customer help
Help customers find answers from your approved product, service or support information, with a clear route to a person when needed.
03
Specialist document Q&A
Interrogate contracts, tenders, technical manuals or case files. Find the relevant clause, compare versions or check a document against a list.
04
Knowledge inside existing tools
Answers where work already happens: in Microsoft Teams, Slack, your CRM or intranet, rather than another tool to open.
How it works
Retrieval-augmented generation,
in plain English.
Most dependable knowledge assistants use an approach called retrieval-augmented generation, or RAG. The AI model doesn’t answer from memory; it answers from the passages it has just retrieved from your material.
Prepare
Gather the approved sources, clean them up, split them into sections and record who can see each one and which version is current.
Retrieve
For each question, find the most relevant passages that this person is allowed to see.
Answer
The model writes an answer using only those passages, and cites where each part came from.
Check
If the sources are missing or disagree, the assistant says so rather than filling the gap with a guess.
Free guide · PDF · 60 pages RAG: a practical guide The architecture, a worked example and a checklist for testing your own assistant. Get the free guide → Making answers trustworthy
Useful when it’s right.
Honest when it isn’t sure.
A knowledge assistant is only as good as its sources, its permissions and the testing behind it.
Sources that stay current
I connect to where documents already live. Updated documents are re-indexed and retired ones removed, so answers reflect the current version.
Permissions respected
Answers only draw on material the person asking is allowed to see, following the access rules you already have wherever possible.
Measured accuracy
I build a question set from real queries and measure how often answers are correct and properly sourced, before launch and after changes.
Designed to decline
When the material doesn’t support an answer, the assistant says so and points to who might know.
Ownership and running costs
Yours to run.
Costs you can see.
Deliverables
What you receive
You own the custom work I build for you.
- A working assistant, deployed where your team will use it
- Connectors to your agreed document sources
- Documentation for users and for maintenance, including how to add sources
- The evaluation question set and results
- A handover session with your team
Costs
What it costs to run
Running cost depends mainly on the volume of documents and questions. I estimate it during scoping.
- Hosting and search infrastructure
- AI model usage, which scales with the number of questions
- Any third-party licences or APIs
- Optional ongoing engineering and support
How a project runs
Map. Shadow. Own.
A focused first version. Tested in the real world. A clear path to what comes next.
01
Map.
I review the questions people ask, the sources that answer them and the access rules involved, then agree a defined first version and fixed price.
- Questions & sources
- Access & data handling
- Costed first version
02
Shadow.
A pilot group uses the assistant on real questions. I measure answers against the question set and improve sources and retrieval.
- Pilot with real users
- Answer quality measured
- Agreed acceptance criteria
03
Own.
I roll out to the wider team, document how to maintain the sources and hand over. Ongoing support is available if you want it.
- Rollout
- Source maintenance guide
- Support options agreed
Can it answer from confidential documents?
Yes, with the right design. The assistant only searches material you have approved, respects who is allowed to see what, and shows the source for every answer. Where the documents are hosted and processed is a project decision, agreed in writing during scoping.
What is RAG?
Retrieval-augmented generation. Instead of relying on what an AI model learned in training, the system first retrieves relevant passages from your own material, then asks the model to answer using only those passages. That makes answers specific to your business and lets them cite their sources.
Will it make things up?
Any AI system can produce a wrong answer, which is why the design matters. Answers are restricted to retrieved sources, citations are shown, and the assistant is built to say when it doesn’t know. I measure the error rate on your own questions rather than promise it's zero.
Can it work with SharePoint, Google Drive or Notion?
Usually, yes, through their APIs. I confirm your specific sources, and how permissions carry across, during scoping.
How is our data handled?
We agree where documents are stored and processed, which model providers are used and under what terms, and who can access what. Those decisions are made during discovery and documented.
How is this different from Copilot or ChatGPT with file upload?
Those tools are useful for individuals. A custom assistant is built around a defined set of approved sources and access rules, is tested on your questions and can appear inside the tools your team or customers already use.
What does it cost?
After scoping, we agree a fixed price for the defined build. Running costs, mainly hosting and model usage, are estimated separately. Talk to me about your project.
Related
Other ways
I can help.
A useful first conversation
Is your best knowledge hard to find?
Tell me what people ask, where the answers live and who needs them. I’ll talk through whether an assistant would help and what a sensible first version could cover.
Discuss your AI project Read the free RAG guide Or email hello@niroxai.comIndependent AI engineering. Based in Hull & East Yorkshire, working across Yorkshire and the UK.