Products · Axistwelve RAG
The answer is on your site. Your search box is burying it.
Axistwelve RAG is governed semantic search. It reads the whole question and answers it in plain English, with the published document behind every claim shown underneath. It draws only on content you have approved.
- UK
- Knowledge base and model on our own estate
- IL3–4
- Infrastructure rated at Impact Levels 3 and 4
- Since 2013
- Answerable for where public-sector software runs
We build and run digital services for
People stopped searching in keywords.
Since 2022 people have been putting whole questions to an assistant, and it has changed what they type into a search box. They arrive with a sentence now, not a phrase.
Keyword search punishes them for it. It matches strings and counts how often they appear, so a longer question gives it more words to match on and more ways to match the wrong document. The page that actually answers it sinks below a news item that happens to repeat the words more often.
Your search today
Matches the words. Counts how often they appear.
With Axistwelve RAG
Reads what is being asked. Answers it.
How governed semantic search works.
Four stages. You control what goes into the first.
Stage one
Point it at your content
You pick the sources. Your own publishing rules decide what makes it into the index, so nothing gets in that you have not approved.
Stage two
It tells you what it understood
It works out what the question means, then shows the reader that interpretation before answering. No guessing on either side.
Stage three
It shows its working
The answer comes back short and in plain English, with every claim traceable to a document and the full results underneath. Readers who would rather look for themselves can switch the answer off.
Stage four
A withdrawn page stops answering
Most tools are fed a page once and never told when it is withdrawn. Withdraw a page and it drops out of every answer, summary and result built on it.
An answer, not a reading list.
A grounded answer is not automatically a useful one. The reader still has to be able to act on it.
What most tools return
Grounded, and no use to anybody
Based on the retrieved documents, it appears that extensions of time may be available in certain circumstances. The documentation indicates that applications are processed in accordance with published service standards. Please refer to the source materials for further information.
What we return
Written the way one of your team would say it
You can ask for more time if the reason was outside your control.1 Send the request with your evidence and you will get a decision within 15 working days.2
1 Applying for an extension of time2 Service standards 2026
Both answers are grounded in the same two documents. Only one of them tells somebody what to do next. The difference is not a better instruction to the model. It comes from knowing what was actually asked, and from the eligibility rules leaving only current documents to answer from.
What a technical review will ask.
Eligibility is decided at indexing time. Content that is withdrawn, superseded, embargoed or outside its approved audience never reaches the knowledge base, so it cannot reach an answer.
Permissions resolve before retrieval, not after it. Access rules decide what can be fetched at all, and different sources can carry different rules.
The model runs on our own serving stack, vLLM. Retrieval and generation both happen inside our UK estate, so a question never travels to get answered.
Every answer is kept with its evidence. The question, the documents retrieved and the answer given are stored together, so an auditor can see months later how a claim was reached.
However people want to ask.
People do not all arrive the same way. One governed knowledge base serves all of them.
Site search
Straight into your site search
Axistwelve RAG replaces the search box you already have. It does not sit beside it. The results follow the intent of the question, with a cited answer above them and filters built from that same question.
Chat
A conversational assistant
Keep the chatbot you have bought. Copilot Studio was never the difficult part. This is the layer behind it, deciding what the assistant is allowed to read when somebody asks a follow-up.
Staff
An internal knowledge base
Your own team gets cited answers across decisions, policies and previous cases, under permissions that differ from the public view of the same material.
Embedded
Inside a service you already run
It sits behind your own product, form or portal, so guidance reaches people at the point they need it instead of somewhere they have to go and look.
See it on your content.
Send us your web address. We index a sample and send you a link where your own content is searchable, with answers you can check.
- Back within five working days
- No call required to see it
- We delete the sample index within 30 days unless you ask us not to
- The architecture and security pack comes on request, in the same form
We use these details to build your sample and contact you about it. See our privacy notice.
Included as standard.
Every deployment gets all of it.
Frequently asked questions.
How is it priced?
You pay a fixed managed service fee, set by the size of your content estate and your expected volume, not by the token or by the seat. Per-token pricing moves with query length, retrieval volume and the provider's own price changes, and historic search volumes predict none of it. Because we serve the models ourselves, we pick one for the quality of its answers, and your cost does not move when someone asks a long question.
How does content get in, and how fast does a change take effect?
We index your published content directly, so it does not matter which system produced it. We connect by feed, API or crawl, whichever your systems already support.
A withdrawal claim is only worth making if it is immediate. Unpublish a page and it leaves answers, results and facet counts within minutes, not at the next reindex.
My metadata is inconsistent. Is this really a metadata project?
It depends on how far you want the eligibility rules to reach. Basic retrieval and cited answers work on your content as it stands. Audience scoping and document-type filters are the parts that need those fields to be reliable.
We start with what you have and show you where the gaps actually affect answers. You decide whether to fix them. What we will not do is quote you for a search product and hand you a remediation programme in month two.
Is it accessible?
Search, facets and the generated answer are built to WCAG 2.2 AA and tested with a screen reader before handover. The generated answer is a labelled region that can be skipped. The reader can switch it off, and you can switch it off site-wide without a release.
Where does it run?
Your content, your users' queries and the responses all stay inside Axistwelve's UK infrastructure, served from our own vLLM stack. That infrastructure is private cloud in two former military facilities in Kent and Newbury. They are rated at Impact Levels 3 and 4 and held at List X and List N, backed by a formal Risk Management and Accreditation Document Set. Axistwelve is one of ten suppliers appointed to Cloud Secure+, Lot 4 of the Government Commercial Agency's Cloud Compute 2 framework. That standard comes from our government hosting work, not from this product. Axistwelve RAG runs nowhere else.
What is it allowed to know?
Content has to be published and current, appropriate to the intended audience and use case, and within an approved document type. It cannot be archived, withdrawn, superseded or under legal embargo, and it has to carry usable metadata such as document type, date and audience. Permissions are applied before anything reaches a model.
What happens when the service is abused?
Public AI endpoints attract automated abuse, and a search box that invites a real question invites a long one. The service runs within an agreed usage envelope, with bot filtering at the edge before requests reach a model, and a review if verified human traffic and billed traffic diverge. We handle that in the architecture, so the cost of abuse does not land on your invoice.
Is it certified?
Axistwelve has held ISO 27001 since 2013, certificate 598644, recertified to the 2022 revision with BSI in 2023, alongside ISO 27017, ISO 27018, ISO 9001, ISO 14001, Cyber Essentials Plus, PCI DSS with the BSI Kitemark and Green Hosting. ISO 27001, 27017, 27018 and 9001, Cyber Essentials Plus and Green Hosting, with scopes, are on our certifications page. ISO 14001, PCI DSS, the BSI Kitemark and the hosting accreditations are on the secure hosting page. Those cover the infrastructure and the operating business. The AI layer is evidenced by architecture and by live client work, and does not yet carry independent certification.
Infrastructure and operating certifications.
ISO 9001
ISO 27001 since 2013
ISO 27017
ISO 27018
Cyber Essentials Plus
Green Hosting
Certified through BSI, apart from Green Hosting, which is certified by the Green Web Foundation.
Try it on your own site.
Send us your web address and we will come back with a searchable sample within five working days.