🤖

Internal Knowledge AI

An assistant that answers staff from your own documents, shows the source and stays silent when it does not know.

Our own product

Danmark Data MCP

An MCP server with 13 read-only tools for Danish companies, addresses, vehicles and phone numbers. Add it to ChatGPT or Claude and ask in plain language; no sign-up, with fair-use limits.

In most companies the answer already exists somewhere: in a manual, an old e-mail, a price list, a shared folder. Finding it takes longer than asking a colleague, and the colleague is busy. Internal knowledge AI turns that search into a conversation. A member of staff asks in their own words, the system finds the passages that match, and a model writes a short answer with a link to where it came from.

The core is retrieval, not the model. We index your documents in Meilisearch with both keywords and vectors. For each document we generate, once, when it is saved, a set of sample questions in Danish, English and Ukrainian and index them together with the text; that is what lets a Danish question find an article written in Ukrainian. Questions that score below a threshold get no answer and go to a list of gaps for someone to fill. Results are filtered by company, team or role before the model sees them, so nobody gets an answer built from a document they are not allowed to open.

We first built this for the knowledge base behind our phone agent. On a fixed test set it found the right article for 13 of 13 questions that had one, and gave no invented answer to any of 5 questions that had none; before the change it found 10 of 13. The threshold, 0.2, came from the gap between the scores of relevant and irrelevant results, not from a guess. We also run Danmark Data MCP, a server that gives ChatGPT and Claude 13 read-only tools for Danish companies, addresses, vehicles and phone numbers. The same approach can put your internal systems in front of an AI assistant without giving it write access.

Costs come in three parts: building the index and the access rules; a small one-off model cost per document when it is indexed; and model usage per question. Embeddings are cheap, the answer step costs more and depends on the model. We log every call and can set a daily ceiling per team. Running the search engine on a server we manage is covered by project support from 600 DKK a month.

It is not worth it if your knowledge fits on one well-kept FAQ page: write the page instead. It is a poor fit for data that changes every minute, such as stock levels, which should be read live from the source system. And it is no substitute for a lawyer or an accountant. It can point to the clause, but someone qualified has to read it.

Key Features

Answers with sources

Every answer links to the document and passage it used. If the model cites something that was not in the material it was given, the answer is rejected rather than shown.

Search across languages

Ask in Danish and find text written in Ukrainian or English. Sample questions in all three languages are generated once per document and can be rewritten by hand.

Honest silence

Below a measured relevance score the assistant says it does not know and logs the question, instead of answering from the closest wrong document.

Access by role

Documents carry company, team and role labels. Search filters on them before any text reaches the model, so answers respect the same permissions as your files.

MCP connection

An MCP server lets ChatGPT, Claude and other assistants query your data through read-only tools, with request limits and a log of every call.

Keyword fallback

If the vector part of the search is unavailable, plain keyword search keeps answering, so the assistant gets less precise instead of going silent.

Why Choose This Service

⚡
Less time spent hunting for answers

Staff ask in their own words and get the passage, not a list of twenty files.

⚡
New colleagues find their feet faster

Routines and exceptions that experienced staff carry in their heads end up in documents the assistant can find.

⚡
You can see what is missing

The list of unanswered questions shows exactly which documents need to be written next.

⚡
Permissions stay intact

The assistant only answers from documents the person asking is allowed to open.

How We Deliver

Our proven process ensures quality at every step

1

Inventory the sources

We list where knowledge lives today, such as files, e-mail, CRM and intranet, who may see what, and which sources are reliable enough to answer from.

2

Index and label

Documents are split, labelled with access rights and indexed with keywords, vectors and sample questions. Stop words for Danish, English and Ukrainian are set before the first document goes in.

3

Measure the threshold

We run questions that must be answered and questions that must not, look at the spread of scores and set the cut-off in the gap between the two groups.

4

Connect the interface

The assistant goes where staff already work: a web page, Telegram, the CRM or an MCP connection to the AI assistant you already use.

5

Keep it current

New and changed documents are re-indexed automatically, and the gap list is reviewed regularly, so the base grows where the questions show it is thin.

Frequently Asked Questions

Do our documents leave the company?

The search index runs on a server we manage or on yours. The passages that answer a question are sent to the AI provider so it can write the answer, so we keep them to what is needed and tell you which provider receives them. Documents that must never leave can be excluded, or answered by search alone without a model.

Which file types work?

Text documents, PDFs with a text layer, web pages, spreadsheets and records from databases or a CRM. Scanned PDFs need text recognition first, which adds cost and errors, so we test a sample of your files before promising anything about them.

How do we know the answers are right?

We test on a fixed set of questions before launch and after every change, and every answer shows its source, so staff can check it with one click. When we built this for our phone agent, the right document came up for every question in the test set, and none of the off-topic questions got an invented answer.

What is MCP, and do we need it?

MCP is an open protocol that lets AI assistants such as ChatGPT and Claude call tools on a server. You need it if your staff already work in one of those assistants and should be able to look things up in your systems from there. Our Danmark Data MCP is an example: 13 read-only tools and no write access.

Can it update itself from our CRM or shared drive?

Yes, if the system has an API. New and changed records are synced on a schedule and deleted ones are removed from the index, so the assistant does not answer from documents that no longer exist. The sample questions are only regenerated for documents that actually changed.

Ready to Cross the Bridge?

Let's build your lightning bridge to digital transformation. From first consultation to live deployment — we move fast.