Associations, professional societies and public bodies have thousands of pages - and visitors who cannot find the right one. The navigator understands the intent and routes to the right source, the right chapter, the current version. It formulates solely on the basis of sources it has found and then leads to the authoritative original passage.
The assistant in the bottom right of this site runs on the same machinery, indexed on our own corpus. Below is its unedited output for a real question - nothing about it has been touched up.
What the visitor typed
„We want to migrate from VMware to Proxmox and secure operations.“
What the system does with it
Classification
WZ-IT reads your project as a strategic migration from a VMware environment to Proxmox VE, to reduce licence costs and regain control over virtualisation. This is not only about the technical move but also about securing operations through high availability, backup and disaster recovery.
The main solution combines the migration itself with building a stable operational base. Proxmox VE handles virtualisation, while backup, monitoring and network infrastructure secure stability. The migration runs in stages: analysis of the current environment, test migration, cutover and final operational hardening.
WZ-IT delivers planning and architecture, including risk analysis and rollback strategy. In operations we take on monitoring, CVE watching, backups and updates.
Matching building blocks
VMware to Proxmox migration
Direct fit for migrating VMware ESXi/vSphere to Proxmox VE, focused on planning, test migration, cutover and operational hardening.
Proxmox & private cloud
Supports building a Proxmox-based private cloud with high availability, backup and disaster recovery as the target architecture.
Managed operations
Securing operations after the migration through monitoring, backups and updates is central to the project.
So we can start where it matters
Sources on wz-it.com: 8
The assistant handles these projects too - ask it yourself:
The assistant is in beta and can classify a request wrongly, which is why every classification carries a route to a human. For knowledge corpora, the sourced passage with chapter and version replaces the offer building blocks.
Not access to a platform, but a system that belongs to you and whose parts you can name.
Your corpora, connected through a CMS interface, sitemap or file storage - documented as to which areas are included and which are deliberately not.
Split by chapter, with title, section, date and audience in the metadata. That is the basis of every sourced answer.
Which requests lead to which areas, and where the navigator hands over instead of answering.
Share of sourced answers, handovers, unanswered topics - measured against test questions from your real enquiry traffic.
Integrated into your site, in your visual identity, without rebuilding your CMS.
How re-indexing works, where the configuration lives, what you get out if you switch providers.
Every answer shows source, chapter link and version. What is not in the corpus is never invented - the curated catalogue routes, the content provides evidence.
Sitemap-based indexing with change detection: CMS content, PDFs, chapter-wise chunking, archived versions stay out. Updates run automatically, critical content after review.
Configurable guardrails: defined topic boundaries, rejected query types, input protection against personal data - enforced before the model call, not requested in the prompt.
The common pattern: the answer is already on your site - nobody finds it because the question is phrased differently from the heading.
Visitor asks: I need something that meets a specific requirement - which of these fits?
Navigator: The navigator clarifies the decisive attributes in one or two follow-up questions, routes to the matching catalogue entries and links the data sheet and the contact.
Visitor asks: Is there any funding for our project?
Navigator: It narrows down by region, project type and eligibility and routes to the matching programme - with deadline, version and a link to the application. It never promises an approval.
Visitor asks: What does the current version say on this topic?
Navigator: It routes to the right chapter of the valid version, names the version date and flags superseded editions. Judging the individual case stays with the professional.
Visitor asks: Why does this not work for me?
Navigator: It routes to the matching help article or how-to - and where nothing fits, straight into the contact form instead of sending the visitor further through the archive.
Visitor asks: How is this handled here?
Navigator: It routes to the valid policy or form template including its version, instead of an outdated PDF that is still linked somewhere.
Visitor asks: What applies to my case?
Navigator: It distinguishes audiences - member, professional, press, general public - and routes each into the right area, including the contact and booking path.
The effort is not in the model, it is in the structure: what belongs to the corpus, what is archive, and which intent leads where?
Sitemap, CMS API and document stores are opened up. First we clarify what belongs to the valid corpus, what is archive and who owns which area editorially.
Content is split by chapter and enriched with metadata: area, language, date, version. Archived editions are excluded, not merely ranked lower.
A curated catalogue defines which intent leads where, where a follow-up question is asked and where a human takes over. That is editorial work, not model selection.
We test with real questions from your actual intake and check every answer against the cited source. Wrong paths are corrected in the catalogue, not talked away in the prompt.
A daily diff keeps the index current. The unanswered questions are the most valuable output: they show where the corpus is missing content.
Before we write a quote, we analyse the corpus. Page count is not what decides; what decides is how much of it is actually machine-readable. We demonstrate the procedure on our own corpus.
How many addresses, how they spread across areas, where duplicates and archived versions sit. On wz-it.com that is 119 knowledge articles, 109 blog posts and around 190 service and landing pages per language - over 400 units per language, a good 800 in total.
A dozen pages per type: does the source return text or only navigation? How long are the sections, how cleanly can they be split by chapter? Experience says a share will not be directly processable - the point is to know that beforehand.
Archived versions are excluded, at any path position, not just under one directory. File views get their own handling and never run through the text path.
With the number of processable documents, the outliers, and what should be tidied up in your corpus first.
Along the way this regularly surfaces problems that have nothing to do with AI: broken sitemap references, unreachable files, duplicates between language versions. You get told about those whether or not the project happens.
The most common practical failure in specialist corpora is not the wrong answer but the outdated one. Four rules, anchored in the index rather than in the prompt:
What currently applies is found first. Anything else needs a reason.
A superseded version that only sits further down the result list will resurface on the next question. It belongs out of the search space.
The reader sees how old the basis is without having to open the source.
If a version has been superseded, the navigator names both and says which one applies. This is the case where staying silent costs more than one extra sentence.
Without measurement, any statement about quality is a claim. We test against questions from your real enquiry traffic and correct in the catalogue, not in the prompt.
How often the navigator could tie its statement to a source - and how often it could not.
How often it passed the visitor to a human. A high figure is not a failure but a finding about the corpus.
The actual editorial backlog: what people ask about that has no content.
Which content actually carries. Often not the pages at the top of the menu.
In this audience, function alone does not decide. We clarify these points before the quote, because otherwise they surface during procurement.
The widget has to be operable, including by keyboard and screen reader. Since 28 June 2025 the German accessibility act applies to many digital services.
Separate corpora per language rather than machine translation at runtime - otherwise the navigator cites a source that does not exist in that language.
Members, professionals, press and the general public need different slices of the same corpus.
The data processing agreement and the description of measures exist before the project starts, not after.
Requests and analytics stay with you. You decide who may see them.
If you switch, you get catalogue, configuration and index out, in a documented format.
A single area can be disabled without stopping the system - important while a corpus is being revised.
With public expert knowledge, a wrong answer costs more than a missing one. That is why the limits are architecture, not phrasing.
| Guardrail | How it is enforced |
|---|---|
| Answers only from cited sources | Without a hit in the corpus there is no answer, only a handover. Free model knowledge is not an admissible source of answers. |
| Citation is mandatory | Every statement carries a link, chapter and version date. An answer without evidence is discarded rather than delivered. |
| Escalation to a human | If nothing fits or the question becomes case-specific, the navigator visibly routes to the contact form or booking page - not into an endless loop. |
| No legal, medical or financial advice | Case-specific assessments are rejected before the model call. The navigator names the authoritative source and the way to the responsible professional. |
| Logging | Query type, chosen path, source and result are logged - without storing personal input as raw text. The log lives in your system. |
| Switchable off at any time | The navigator is a component of your site, not a replacement for it. It can be disabled per area or entirely without making content unreachable. |
A guided navigator is deliberately narrower than a chatbot. That narrowness is exactly why you can point it at public expert knowledge.
It takes the always-identical paths off your desk and hands over the rest cleanly. Anyone who needs a real answer reaches a human faster - not less often.
It does not chat about everything. Topics outside the corpus are declined rather than answered creatively. That is what separates it from a generic website bot.
It shows what your sources say and who is responsible. Assessing a concrete case stays with you.
The navigator works on released content. For internal stores with permissions, the internal assistant is the right build - it enforces permissions before retrieval.
See the internal assistant →Honestly upfront: page count alone does not justify a navigator. What matters is whether the corpus is versioned, whether different audiences need different parts of it, and what a wrong answer costs. Three hundred heavily versioned guidelines are a case for it; two thousand flat pages often are not, and there a good search is cheaper.
Where we stand: the navigator is currently being built in a pilot project on a large medical knowledge corpus. We will publish figures and results once the project is finished and cleared - before that they would be nothing but a claim.
Runs where you want
We name the price after the corpus analysis, because without it we would be guessing. These six points determine it:
How the corpus analysis is charged is settled in the first conversation. We tell you about the problems we find along the way; how we charge overall is settled in the first conversation.
When internal documents need searching - with permissions that apply before retrieval.
When visitors should be routed to the right offer rather than the right source.
When something should be done, not just found: tickets, workflows, systems.
How answers from cited sources actually come about - without the sales lens.
Answers to the most important questions
Search finds pages containing words. The navigator understands the intent (even colloquially phrased), maps it to the corpus and routes to the right source including chapter and version - with a clarifying question only on genuine ambiguity.
No - that is the core architectural decision: the navigator routes to existing content and cites it with a link. A curated catalogue defines the targets, the index provides evidence. Anything outside the corpus leads to an honest handover to humans.
Via your sitemap and - where available - the CMS API: content is split by chapter, enriched with metadata (area, language, date, version) and embedded. A daily diff detects changes; archived versions are excluded, not merely downranked.
Yes, with hard guardrails: the navigator answers no case-specific questions but routes to the authoritative source with its version date. Inputs are checked for personal data before any model call; critical query types are blocked and logged.
Then there is no answer but a handover: the navigator states that the question is outside the corpus and routes to the contact form or booking page. These cases are evaluated and show where content is missing.
The catalogue is editorial work and stays with you - we build it, document the structure and train your editors. Changing paths or topic boundaries is therefore a task in your system, not a ticket to us.
Yes. The stack runs on European infrastructure with a European model (Mistral), self-hosted on your side if preferred. Queries are not stored as raw text, and the processing is documented cleanly.
Fixed-price setup after a corpus analysis (scope depends on source count and structure) plus ongoing operations with index maintenance and reporting. In the intro call we analyse your sitemap and give you a reliable figure - no strings attached.
Page count alone does not decide. Three things do: is the corpus versioned, so that valid and superseded versions exist? Do different audiences need different slices? And what does a wrong answer cost? Three hundred heavily versioned guidelines are a clear case; two thousand flat pages often are not, and there a good full-text search is cheaper and faster.
A PDF with an outline and embedded text can be split by chapter and is processable. A scanned PDF without a text layer needs OCR first, and that costs extra. The corpus analysis before the quote establishes what the share is - which is why it comes at the start and not at the end.
Through separate corpora per language. Machine translation at runtime leads to the navigator citing a source that does not exist in the language asked about. If content exists only in German, the navigator says so and links the German version instead of inventing an English one.
It has to be operable by keyboard and screen reader - since 28 June 2025 the German accessibility act applies to many digital services. We check focus order, labels, contrast and the announcement of status messages, and document the state rather than claiming accessibility in the abstract.
We name the price after the corpus analysis, because before that we would be guessing. It depends on the number and type of sources, the PDF share, existing versioning, the number of audiences and languages, the type of connection and the operating model. How the analysis is charged is settled in the first conversation.
On request entirely in your environment, or on European infrastructure we operate. Requests and analytics stay in your system - you decide who may see them. The data processing agreement and the description of technical measures exist before the project starts.
A full-text search returns a result list and leaves the judgement to the reader. The navigator understands the request, picks the right area, formulates a short classification solely from sources it has found, and then leads to the authoritative passage - with chapter and date. Where a search shows twelve hits, it names one and says why.
A support chatbot answers recurring questions on a manageable, slow-changing set of topics. The navigator works on a large, versioned specialist corpus in which telling the valid version from the superseded one matters for safety. That is a different class of problem - and it is the one we build for.
It says so and hands over to a human or the contact form. An answer without evidence is the most expensive mistake a system on a specialist corpus can make. How often that happens is in the quality report - and the topic list from it doubles as your editorial backlog.
Catalogue, configuration and index in a documented format, plus the operations documentation. The corpus itself was always yours - the navigator lays a layer over it; it does not pull your content into a foreign system it cannot leave.
No risk: worst case, you leave with a clearer understanding of your project than before.


“WZ-IT's advice on our Azure migration was technically sound and completely non-binding right from the intro call - we took away a great deal.”
ml&s speaks about integrating a local AI solution. The other voices cover architecture, data sovereignty and operations - exactly the maturity an AI project needs to reach production.
Whether a specific IT challenge or just an idea - we look forward to the exchange. In a brief conversation, we'll evaluate together if and how your project fits with WZ-IT.
Timo Wevelsiep & Robin Zins
Managing Directors of WZ-IT
