What it is
Built directly on udX.
After udX has analysed a document’s entities, its content is automatically grounded in a vector database: stored so it stays precise, fast and searchable without losing the connection to the original document.
On top of that, udX Chat lets you talk to your documents through a language model. Two mechanisms make sure every answer is correctly framed: a three-stage search that always looks in the right place, and a graded evidence system that shows how well each answer is backed.
- No separate onboardingEvery document analysed by udX is automatically chat-ready.
- Evidence required by defaultBacked answers come exclusively from your documents, enforced in the backend rather than in the prompt. The optional research stage is kept strictly separate and visibly marked.
- Research in secondsQuestions that used to take hours of manual searching are answered with evidence attached.
Three-stage search
Always the right search depth: automatically.
Each query is answered at the appropriate level. The user never has to pick the search depth; the system escalates on its own, so no relevant information is missed.
Within the current document
The narrowest, most precise level: used when you chat from inside a document view.
Within a document cluster
Related documents, such as all annexes of a contract or all policies of one customer. Clusters are built automatically from references and shared entities.
Across all documents
The full knowledge base, when the answer spans multiple documents: searched by an agentic retrieval step that selects the relevant passages itself.
Graded evidence
Every answer shows how well it is backed.
Answers are not delivered as a single block of text but structured by their level of evidence: the hardest, best-backed level is always tried first. Each grade is technically validated in the backend before anything reaches the user.
The answer rests directly on entities and the exact wording of the documents. The backend verifies that every quoted passage really exists, character by character: only then is the source highlighted in the original.
Information is derived from the document content but remains directly covered by the text: verified against the source passage after normalisation.
The answer draws a conclusion that is not spelled out word for word: a calculation or logical inference. Every underlying fact must still be quoted verbatim and validated.
If no level produces a validated answer, udX Chat says so explicitly: "This information is not contained in the available documents." Nothing unverified is ever delivered.
Research stage
When your own documents fall short.
Some questions simply cannot be answered from your own documents. By default udX Chat stops at abstention. If you would rather have an answer, you can enable the research stage: it answers such questions from general knowledge or from a web search.
The separation is what matters. Answers from the research stage are unmistakably marked as such and are never mixed with backed statements. The evidence requirement for answers drawn from your document base is untouched by it.
- Optional and switchableThe stage is off by default. The application points it out before use, and it can be disabled entirely.
- Web search via an EU providerThe web search works with an EU provider. Providers outside the EU are blocked by default and usable only after explicit approval.
- Never mixedBacked and researched statements stay apart. You can always see what an answer rests on.
Formats
What udX Chat can read.
Layout awareness instead of a wall of text: tables, columns and page structures are preserved as such. Scanned contracts and photographed pages become searchable through text recognition including descriptions of figures. Complex workbooks are condensed into dependable key statements rather than dumped cell by cell, with the time reference and origin of every figure staying traceable.
Connecting sources
Collected where the documents already live.
Network drives and SharePoint
Both can be attached as connected sources. OneDrive and Google Drive are prepared.
Pre-analysis is mandatory
Before a source is connected you see the document count per file type, the page volume and a cost estimate. Only then do you confirm.
Deliberately manual sync
One click identifies new and changed files and analyses only those. Nothing runs unnoticed in the background racking up cost.
Cost protection built in
A token budget per chat, page and usage quotas per group, plus a pause switch for the entire processing pipeline.
Core functions
What udX Chat does.
Automatic grounding
Documents are placed in the vector database right after udX analysis: text, entities and even figures from charts and images become searchable.
Source reference per answer
Inline markers link every statement to the exact page, clause and highlighted passage in the original document: clickable in the viewer.
Anti-hallucination gate
Quotes that cannot be technically verified are discarded; answers without a single valid quote are never shown. Enforced in the backend, not just prompted.
Ask in any language
Questions are normalised and translated automatically; the answer comes back in your language while quotes stay in the original.
Follow-up aware
References like "and in euros?" are resolved against the conversation, so research works like a dialogue.
Live search status
While searching, the current stage is shown live: from the single document up to the full base.
Security & data sovereignty
Built for production, not for demos.
Operated in Germany
Runs in an EU region, Germany West Central for new installations. Your documents do not leave the EU.
No model training on your content
Your documents are processed to answer your questions, not to train models.
Virus scan on every upload
Every file is checked before it is read, analysed or indexed. Files identified as malicious are rejected, removed and logged.
Defence against manipulated documents
External content is handed to the language model explicitly marked as data. Hidden instructions inside a document cannot steer the application.
Fail-closed, no fallback path
Identity is verified cryptographically on the server. If the check is missing or fails, the request is rejected. There is no path that grants access when in doubt.
No keys in circulation
All service access uses managed identities with least-privilege roles. Secrets live exclusively in a managed key vault.
Deployment models
Three ways to run udX Chat.
An isolated environment per customer is the norm in production: no shared database, no application-level tenant separation that could be misconfigured.
Shared test environment
EvaluationFor trying it out without standing up infrastructure of your own.
- Operated by koeniq, shared
- Sign-in via koeniq
- Test data only
Per-customer deployment
ProductionThe norm: a dedicated, isolated environment for your company alone.
- Own isolated environment per customer
- Sign-in via your corporate account
- Data fully separated
Your infrastructure
Highest requirementsFor organisations whose data must not leave their own network.
- Runs in your data centre or your cloud
- Sign-in via your directory
- Data never leaves your premises
Typical scenarios
Where udX Chat earns its keep.
udX Chat answers open questions across your entire document base. Its sibling udX Audit does the opposite: it checks specific target documents against their governing sources, continuously.
Discover udX AuditSee udX Chat on your own documents.
Book a free 30-minute call. We show you the product live and tell you honestly whether it fits your document base.
Book a demo