Services
AI & Software
Solutions
Training
Tools
Insights
References
Contact
Free initial consultation Ask a first question by email
About Deutsch
← AI & Software

AI assistant with sources (knowledge vault)

Company knowledge as a versioned vault: answers only with sources, “unknown” instead of a guess, reading and control separated. With a maintenance process.

In the Teams call the assistant names a deadline, everyone nods, and two days later it turns out the deadline came from a version that was superseded in spring. “Where does that come from?” is the most common follow-up to any answer. An assistant that writes fluently while hiding its source saves time on the first read and costs double on the check. Once you cannot trust it, you ask the colleague after all.

How I work: with source · with approval · in your system.

Starting point

Company knowledge is spread out: approvals in folders, project status in the project system, rules in people’s heads, decisions in emails. A general language model knows none of it. Without its own basis, it invents a plausible answer, which is the most expensive failure in business use. Then there is the second question people rarely ask: who checked this document last, and is it still valid?

What I do

I build company knowledge as a versioned vault. Every note carries a source, a review date and a checklist that shows its state. Search runs lexically and needs no model; for imprecise questions a hybrid search with a reranker can be switched on. The assistant answers with source IDs, so every statement traces back to a note. If there is no reliable basis, it answers “unknown”.

Reading and control are separated. The assistant may answer questions and read data. Writing actions take their own path: preview, then a signed confirmation token, then the write with a read-back of the result. Anything that cannot be undone stays locked.

The assistant runs locally on its own hardware through a private connection. A public model API is not required for that. Retention of conversation history is configurable. That keeps the question “where does the data sit?” answerable.

Result

You get a running assistant, a permission concept and a maintenance process. Add the answer rules, the source checklist with review dates and a defined retention period. Without that list, the vault is just a folder.

Example

The knowledge assistant is my own product and in internal use as a module in the Projectile Mobile marketplace. Projectile Mobile is my own development in front of Projectile, not a product of the vendor. It runs as a standalone package on its own GPU, privately connected, with no public model endpoint. Search finds the vault offline, and hybrid search with a reranker can be switched on. Answers carry source IDs, the checklist shows the state of each note, retention is adjustable. Writes go through preview, a signed confirmation token and a read-back. If the basis is missing, it answers “unknown” instead of guessing.

A manufacturer engagement shows the same construction in production: vault, reranker, binding answer rules and a test catalogue of domain questions. The engagement is anonymous; names and numbers do not belong on this page.

Limit

The maintenance effort is real. A vault without review dates ages, and an aged vault answers confidently wrong. That is why a person belongs to it, approving sources and retiring old versions.

The second limit is scope. What is not in the vault, the assistant will not answer, and that is intended. Too large a vault dilutes the passages, too small a one answers only half the questions. That call needs the business side, not IT alone.

How it runs

  1. Review sources, clarify rights, create the vault
  2. Define answer rules and how passages are weighted
  3. Connect the assistant, locally or in the EU
  4. Test questions from daily work, review with the domain teams
  5. Hand over the maintenance process, set the review rhythm

Common questions

What happens when the answer is not in the sources?

The assistant answers “unknown” and names the missing basis. It does not fill the gap from general model knowledge when the answer must be evidenced. If the model is down, lexical search still returns passages, just without a written answer.

Can the assistant change documents or data?

Not in the default setup. Writing actions are separated and run through preview, a signed confirmation token and a read-back. Anything irreversible stays locked, even with confirmation.

How much maintenance does the vault need?

That depends on how fast the knowledge moves. Stable rule sets need a quarterly review. In project-driven companies the effort is higher.

Is this permissible under data protection law?

Local operation makes the question easier but does not answer it alone. Processing records, retention and access belong in the documentation. I provide the technical basis; the legal review stays with you.

The next step costs nothing.

30 to 60 minutes, no strings attached: where is the biggest friction, what is the most effective lever, and is a collaboration a fit?