AI chatbot · RAG · Knowledge base
Answers from your own documents,
with sources and clear limits.
An AI chatbot answers customer questions on your website or, as an internal knowledge assistant, searches manuals, contracts and tickets. It answers from documents you have approved, shows where the answer came from and hands over to a person when it is unsure.
Discuss your projectFree initial conversation · Directly with Daniel Gläser
- 01Question
- 02Search + answer
- 03Check source
- 04Reply or hand-off
Use cases
A chatbot for customers or an assistant for your team
The technology is similar, the requirements are quite different. So we first establish who is asking and which documents are approved for it.
Website chat for customer enquiries
Answers recurring questions from prospects and customers using content approved for the public.
- Sources: website, product data, FAQ, delivery and service terms.
- Typical questions: services, delivery times, procedures, contacts, opening hours.
- Clearly labelled as AI, with a visible route to a person.
- Can record an enquiry with the conversation history as a ticket or email.
- No access to internal documents or customer data.
Internal knowledge assistant
Helps your team find knowledge in long-grown file stores instead of asking colleagues or digging through folders.
- Sources: manuals, work instructions, contracts, tickets, wiki, SharePoint or network drive.
- Typical questions: What is the procedure for a complaint? What notice period does the maintenance contract have?
- Document-level permissions: everyone only sees what they could open themselves.
- Available in the browser, in Microsoft Teams or in your line-of-business application.
- Answers link to the original document.
How it works
RAG: search first, then answer
A language model knows nothing about your business. Retrieval augmented generation, or RAG, solves this without training a model of your own: your documents are split into passages and stored in a search index, for example in PostgreSQL with pgvector or in Qdrant. For each passage the index keeps a kind of meaning fingerprint alongside the plain text.
When someone asks a question, the system first looks for the relevant passages. We combine search by meaning with classic keyword search so that item numbers and technical terms are reliably found. Only then does the language model receive the question together with those passages and the instruction to answer from them alone and cite the source.
If the search finds nothing suitable, the assistant says so rather than guessing. When a document changes, only the index is updated. The model does not need retraining, and your documents never become part of the model itself.
Control
Traceable, permission-aware, clearly labelled
Source references
Every answer points to the document and passage it is based on. That makes it quick to check whether the answer is right, and outdated documents come to light fast.
Document-level permissions
The assistant adopts the access rights from SharePoint, your network drive or your business application. Each question only searches passages the signed-in person is already allowed to open. Salary lists stay with the people entitled to see them.
Hand-off to people
For uncertain matches, complaints or when asked, the chat hands over to your team along with the conversation so far. Internally, the assistant points to the responsible person when a rule is missing or contradictory.
Labelled as AI
The chat identifies itself as an AI assistant in the window and in its first message. The EU AI Act requires that people are informed when they interact with an AI system, unless that is obvious. I add the label in every case.
Guide to the EU AI ActHosting & upkeep
Where the chatbot runs and how it stays good
Hosting, knowledge upkeep and quality measurement decide whether a chatbot is still in use three months later. I look after software and servers myself, so both are part of the plan from day one.
Language model in the EU cloud
Current models from large providers can be used with processing in the EU, for example Azure OpenAI with a deployment in the EU data zone, or the Mistral AI API, which hosts data in the EU by default. The upside is high answer quality without your own hardware. Data processing terms, retention periods and sub-processors are reviewed for the specific use.
Guide: AI and data protectionLanguage model on your own hardware
Open models, for example from the Llama, Mistral or Qwen families, run with tools such as Ollama or vLLM on a server on your premises. Documents and questions never leave your network. In return you need suitable hardware with a GPU and ongoing maintenance, and answer quality depends on model size.
Local AIMaintaining the knowledge base
Outdated price lists or duplicate versions of a manual lead to wrong answers. We decide which folders are included, who approves content and how often it is re-indexed. After that, new and changed documents reach the index automatically.
Measuring quality with test questions
Before launch we build a set of real questions with the expected answer and source. Every change to documents, model or settings is run against these questions. In operation we review ratings and unanswered questions and close specific gaps in the knowledge base.
Limits
What a chatbot cannot do
Language models can produce answers that sound plausible but are wrong, often called hallucinations. RAG, mandatory sources and the instruction not to invent anything when nothing is found reduce this risk considerably, but cannot rule it out. That is why a person checks the source for questions with consequences, and the chat never makes binding commitments on prices, deadlines or warranty.
An assistant is only as good as its documents. If two versions of a manual contradict each other, it cannot know which one applies. Scans without text recognition, tables with merged cells and knowledge that only exists in people's heads all need preparation. Calculations, stock queries or triggering transactions do not belong in a pure knowledge assistant; they belong in an agent with defined interfaces and permissions.
Cost depends on the number and condition of documents, the number of source systems, the permission model, the choice between cloud and your own hardware, and query volume. Upkeep and regular quality checks come on top.
Working together
A clear path to everyday use.
Collect questions and sources
We gather real questions from the inbox, phone or team and decide which documents and systems provide the answers.
Prototype with your documents
I build a prototype on a defined set of documents, with source references and your permissions.
Test and measure
The test questions run against the prototype. We review hits, errors and hand-offs and agree the quality level required for rollout.
Roll out and maintain
Integration into your website, Teams or business application, followed by ongoing upkeep of the knowledge base and review of feedback.
Before we begin
Good questions. Clear answers.
How much does an AI chatbot for a business cost?
It depends on the volume and condition of documents, the number of source systems, permissions, hosting and query volume. A website chat with a few public sources is much smaller than an internal assistant across SharePoint, a ticket system and a network drive. We start with a prototype whose effort is agreed up front, and I work through the running costs for model usage or servers with you beforehand.
Does the chatbot make up answers?
The risk can be reduced a great deal, but not eliminated. The assistant answers only from passages it has found, cites the source and says when it finds nothing. Test questions before launch and spot checks in operation show how reliable it is with your content.
Where is our data processed?
We decide that by type of data. Public website content is often fine with a model that processes data in the EU. For internal or confidential documents, the model can run on your own server on your premises. Data flows, retention and contracts are reviewed for the specific use.
Does a chatbot have to be labelled as AI?
The EU AI Act requires that people are told when they are interacting with an AI system, unless this is obvious from the circumstances. I therefore always label chats clearly in the window and in the first message. The guide to the EU AI Act covers the obligations in more detail.
Which documents can the knowledge assistant use?
Common formats such as PDF, Word, PowerPoint, Excel, HTML and plain text, plus content from SharePoint, network drives, a wiki or a ticket system. Scanned documents need text recognition first. We agree together which sources are included and who approves them.
How up to date are the answers?
As up to date as the index. New and changed documents are re-indexed on a fixed schedule or automatically when they change. Deleted or replaced documents are removed from the index so the assistant does not answer from old versions.