Services · Building block

The AI knowledge base that answers with its source.

The AI knowledge base answers only from your approved sources, with document and page. If it finds nothing reliable, it says so.

Knowledge systemsigned in as Service
  • Maintenance manual VX-400.pdfVersion 03/2026Service
  • Case 813Ticketing, May 2026Service
  • Framework contract Hoffmann GmbH.pdfSales filesno read access
Gearbox lubrication interval, VX-400 series?

Every 2,000 operating hours, every 1,500 in dusty conditions. On this machine the interval was overrun once in May.

Manual p. 14Case 813
  • An answer, not a hit list

    Instead of forty hits in the file share, one answer with document and page. Anyone who wants to check is one click from the passage.

  • Only what the person may read

    Read permissions come from your sources. What sales files away, service does not see, not even through the chat.

  • Fewer questions for the experts

    New colleagues find the answer themselves. The experienced person stays free for the cases that really need them.

  • No second file share

    The search runs over the documents you already maintain. When one changes, the index follows and replaced versions drop out.

Where it gets stuck today.

  1. The colleague who knows everything.

    New staff ask the same experienced person several times a day. When she is on holiday, the work waits.

  2. 19%

    Two versions of one policy.

    The file search returns forty hits and the first is the version from two years ago. So people answer with whatever they remember.

    McKinsey Global Institute, 2012: share of knowledge workers' time spent searching for and gathering information
  3. The chat that knows too much.

    A chatbot wired up in a hurry quotes a document the person asking should never have seen. A day later it is switched off.

    OWASP Top 10 for LLM Applications 2025, LLM08: missing access control in the search index

From source to answer

Every answer shows where it came from.

We approve the maintained sources, split them into passages and keep version and read permissions. People ask in Teams, Outlook or their application, and the answer comes only from those passages.

Approved sourcesService
  1. SharePoint ServiceMaintenance manuals, 214 filescurrent
  2. Ticket systemCases since 2022current
  3. Sales filesFramework contractssales only
  4. Drive S: archiveManuals up to 2019stays out
Maintenance manual VX-400Section 7.3

7.3 Gearbox lubrication

Oil change every 2,000 operating hours, every 1,500 under dusty conditions.Passage 1

Oil grade CLP 220, filling quantity 4.2 litres.Passage 2

Version 3, valid from 1 March 2026, read access Servicekept

Knowledge systemTeams
  1. Gearbox lubrication interval, VX-400?

  2. Every 2,000 operating hours, every 1,500 under dusty conditions.

    Manual VX-400, p. 14

Limits

What it will not answer matters as much.

A knowledge system is only as good as its refusals. If it finds no solid passage, or the person may not read it, it says so and names who can help.

  • No source, no answer

    If it is not in the approved documents, the system does not write something plausible. The question goes to whoever maintains the source.

  • No read access, no hit

    Filtering happens before the answer, not after. A sales contract never shows up for service, neither as a quote nor as a hint.

  • Tested on real questions

    A set of real questions with confirmed answers runs again after every change. It also measures whether the system declined when it should.

Knowledge systemsigned in as Service
  1. Which warranty applies to Hoffmann?

  2. You have no source for that. Sales can help.

  3. And the torque for the gearbox cover?

  4. The VX-400 manual says nothing about it. I have passed the question to the person in charge.

HitsWarranty Hoffmann
  1. Framework contract HoffmannSales filesno read access
  2. Warranty policy v2replaced by v3outdated

Approaches

The usual routes to a knowledge system.

A knowledge system needs a search that finds both wording and meaning and respects read rights. There are a handful of proven ways to build one. Ours is PostgreSQL by default, and we say when another fits better.

  • PostgreSQL with pgvector

    Our default

    Our stack, run in Germany

    Fits when
    you want answers with a source from your own documents without data leaving your company or Germany. Search runs three ways: by meaning, by exact wording for part numbers and names, and by word stems in German and English; a self-hosted model reranks the hits. Every answer names chapter and page.
    Mind
    Read rights are enforced by the database itself: whoever may not see a document never gets it as a hit. Operation and updates are on us.
  • Azure AI Search

    Managed, Microsoft Azure

    Fits when
    Your data already lives in Azure and Microsoft 365. Hybrid search, full text and vector in one query, indexers for SharePoint and Azure storage.
    Mind
    Taking read rights straight from SharePoint is still in preview according to Microsoft; the cost depends on the tier you book.
  • Elasticsearch or OpenSearch

    Search engine, self-hosted or cloud

    Fits when
    A search already runs in-house. Both combine classic full text search with vector search in one query.
    Mind
    A cluster of its own with its own operating effort; worth it mainly where it already exists.
  • Qdrant, Weaviate or Pinecone

    Dedicated vector database

    Fits when
    Very many sections need to be searched fast and search should be a building block of its own.
    Mind
    One more system next to your data; rights and metadata have to be carried along. Pinecone is only available as a cloud service.
  • Microsoft Copilot with SharePoint

    Off the shelf, Microsoft 365

    Fits when
    The documents are kept tidy in SharePoint and Teams. Copilot answers from what the person may read there, without a project of its own.
    Mind
    Licence per user; overly broad sharing in SharePoint becomes visible, and sources outside Microsoft 365 need connectors.

Tasks and limits.

What the system takes on

  • Searches the approved sources with the permissions of the person asking, and no more
  • Answers only from the passages it found and names every source
  • Says when it finds nothing solid instead of writing something plausible
  • Is measured against a set of real questions, again after every change
  • Scores separately: right passage found, answer matches the passage, correctly declined
  • Collects unanswered questions for whoever maintains the source

What your team takes on

  • Decides which sources count and who maintains them
  • Decides which version is authoritative
  • Gives every binding answer to the outside world themselves
  • Writes down knowledge that lives only in heads; the system cannot read it
  • Fixes wrong policies; old sources give old answers, now with a citation
  • Analyses ERP figures; a query is the better tool for that

AI knowledge systems

A knowledge system answers with evidence. An agent acts. If the sources first need preparing, Data & documents comes before it.

Which one when?

  • AI agents

    Should something get done, not just answered? Then an AI agent that reads from the knowledge system.

  • Data & documents

    Are the sources scans, duplicates or old versions? Then start with Data & documents.

The way there.

From the first call to a system in operation. Each phase stands on its own; none ties you to the next.

  1. Intro call

    free

    30 minutes. We work out where AI fits for you and whether an AI strategy is worth it.

  2. AI strategy

    €1,990

    AI strategy for mid-sized companies: which of your processes suit AI, what the build costs and brings back, and which ones are not worth the effort.

  3. Pilot & implementation

    from €5,000

    AI pilot at a fixed price: one use case, fixed scope and acceptance criteria in writing up front. Built in your systems, tested with your data.

  4. Operation & improvement

    from 20% a year

    Running AI systems: hosting in Germany, monitoring, model and prompt maintenance, ongoing measurement, and the next use case on the same foundation.

Further reading

All posts

Frequent questions

Common questions about it.

The questions that almost always come up.

What is a RAG system?

RAG stands for retrieval-augmented generation. The system first searches your approved documents for the passages that fit and then lets the language model answer only from those passages, with the source. An AI knowledge base as we build it is such a RAG system, with read permissions and versions added.

How is an AI knowledge base different from ChatGPT?

ChatGPT answers from what is inside the model. A knowledge system searches your approved documents first and answers only from those, naming the passage it used. What is not in the sources does not get answered. That boundary is deliberate.

How do you stop an AI knowledge base inventing answers?

Three ways. It answers only from the passages it retrieved and names them. When nothing solid is found it says so instead of writing something plausible. And a set of real questions with known good answers is rerun after every change. A wrong answer can never be ruled out completely, which is why every statement carries its source, so it can be checked in seconds.

Which documents make good sources for an AI knowledge base?

Anything that is searched and read regularly: manuals, work instructions, policies, contracts, technical documentation, case histories from line-of-business systems. What matters is not volume but whether the documents are maintained.

Does everyone see the same answers?

No. Permissions come from your sources, and the system searches only what the person asking is allowed to read anyway. This is where such projects most often fail, so we test it with test accounts before go-live.

What happens in the knowledge base when a document changes?

The search index follows the change, replaced versions drop out, and the answer names the version it came from. There is no second filing system for anyone to maintain.

Does our data stay with us?

Yes. Processing in German data centres or inside your own infrastructure, under a data processing agreement under Article 28 GDPR before any data moves. We only use model access where the provider contractually commits not to train on your data.

What does an AI knowledge base cost?

The effort depends on the number and state of the sources, how deeply it is integrated, and the requirements around permissions and auditability. It is built as a fixed-price pilot; we narrow it down against your case in the free introductory call.

The next step

Let us find out where AI pays off for you.

30 minutes, free and without obligation.