Glossary: the terms behind the AI knowledge base
Every technical term in 60 words at most, understandable without technical knowledge. For looking up, passing on and quoting — every entry has its own anchor link.
AI knowledge base
An AI knowledge base is a central, searchable store for company documents, combined with an AI that answers questions about them in plain language. The answers are based on the company’s own documents and name the source. Whether the web may also be searched is decided by the company itself. It is intended for documented know-how – instructions, reports, solutions to problems – not for records that stay in the ERP.
RAG (retrieval-augmented generation)
RAG is a method in which an AI, before answering, first retrieves the relevant passages from a collection of documents and bases its answer on them. This is meant to reduce invented answers (hallucinations) and makes source references possible against which an answer can be checked. RAG is the technical basis of most AI knowledge bases.
LLM (large language model)
An LLM is a large AI language model that understands and writes text – well-known examples are the models behind ChatGPT or Claude. An LLM on its own does not know your company documents; only the combination with a document base (see RAG) turns it into a tool for internal company knowledge.
OCR (text recognition)
OCR (optical character recognition) converts images of text – scans, photos, old PDFs – into searchable text. For companies, this means that documents from the paper archive can also be included in the search. How well recognition works depends on the original – print quality, capture, age; handwriting is the hardest case.
Semantic search
Semantic search finds content by meaning rather than exact wording. A question about “heating not getting warm” can thus also bring up the report that says “flow temperature too low”. Classic full-text search, by contrast, needs the keyword as it appears in the text. Semantic search is a core component of most AI knowledge bases.
Embedding
An embedding is the translation of a text into a series of numbers that represents its meaning. Similar content gets similar series of numbers – this lets a system calculate how close in meaning texts are. Embeddings make semantic search technically possible and are stored in a vector database.
Vector database
A vector database stores embeddings (series of numbers representing meaning) and retrieves the passages that are closest in content to a question. It is the “memory” of an AI knowledge base: with every query it supplies the relevant document excerpts from which the AI builds its answer.
Hallucination
A hallucination is a convincingly worded but factually wrong AI answer. It arises when a language model answers without reliable sources. The remedy in business use: base the AI on your own documents (RAG) and name the source for every answer, so it can be checked against the original.
Evidence for the answer (grounding)
Grounding means that an AI answer is traced back to specific sources in real documents, for example via a numbered reference for each statement. For companies, this is the difference between “sounds plausible” and “can be verified”: via the source you open the original document and read it yourself.
Tagging
Tagging assigns attributes to each document – customer, project, installation, date. The AI suggests them when the document is captured; the document type is confirmed by a person, because it decides visibility. Tagging makes large collections filterable and helps assign answers to the right customer or project.
DMS (document management system)
A DMS stores documents in a structured way and makes them findable again via folders, metadata and keywords. Anyone who wants to search has to know the filing logic. An AI knowledge base goes one step further: you ask a question in plain language and get an answer with its source.
DPA (data processing agreement)
A data processing agreement under Art. 28 GDPR governs how a service provider processes personal data on behalf of a company – for example when hosting a knowledge base. It is required by law and sets out the purpose, security measures and sub-processors. It is worth requesting and reading it before signing a contract.
The European AI Act
The EU AI Act is the European law on artificial intelligence. It divides AI applications into risk classes and applies in stages; transparency obligations (Art. 50) since 2 August 2026, high-risk obligations from December 2027 or August 2028. Companies that use AI must take measures to promote the AI literacy of their staff (Art. 4).
Experiential knowledge (tacit knowledge)
Experiential knowledge is the know-how in the heads of long-serving employees: how a system is set up, how a problem was solved at a customer’s site, what to watch out for on a project. It is lost when people retire or leave – unless it is documented and made searchable.
Fine-tuning
Fine-tuning means further training a language model on your own texts. For company knowledge, this is rarely the right route: the knowledge merges with the model, can no longer be deleted individually, knows no permissions and names no source. A knowledge base using the RAG method keeps the documents separate — changeable, deletable, with evidence.
Prompt
A prompt is the input to an AI system — in a knowledge base, simply the question someone types or dictates. Good questions name the specifics: customer, installation, project number or time period. No special jargon is required; you ask the way people talk at your company.
Token
A token is the smallest unit of text into which a language model breaks down a text — roughly part of a word. The number of tokens determines how much text a model can process at once and what processing costs. For a company, this matters only indirectly: it explains why a system retrieves the relevant passages specifically instead of reading entire folders.
Context window
The context window is the amount of text a language model can take into account at the same time for a single answer. It is limited, which is why no model can simply read the entire document collection. Instead, the system first retrieves the relevant passages and puts only these into the context window — that is the core of the RAG method.
Chunking (text sections)
Chunking is splitting long documents into smaller sections before they are made searchable. The section boundaries help determine answer quality: sections that are too small lose context, sections that are too large dilute the source. That is why, during set-up, it is decided how contracts, reports or drawing labels are sensibly divided.
Metadata
Metadata are the attributes of a document: customer, project, installation, date, document type. They make large collections filterable and ensure that an answer is assigned to the right case. The system suggests most values based on the content when the document is captured; the document type is confirmed by a person.
Document type
The document type says what something is: quotation, inspection report, payslip, contract. It is the most important classification, because visibility depends on it — anyone allowed to see a document type sees all documents of that type. That is exactly why the system only suggests the type; it is confirmed by a person.
Role and permission concept
The permission concept determines who may see and query which documents. It is controlled via roles and document types, not individual files — otherwise it could not be maintained day to day. A cut-off date is also possible: only from a certain date, or only older documents up to then. If a role lacks the permission, the system says so explicitly.
Full-text search
Full-text search finds documents in which the words entered appear literally. It is fast and precise, but depends on the choice of words: a search for “maintenance contract” does not return a “service contract”. A knowledge base therefore combines full-text search with semantic search, which looks for meaning rather than spelling.
Indexing
Indexing is the process by which documents are made searchable: recognising text, splitting it into sections, assigning attributes and adding it to the search index. It runs in the background as soon as a document is added. Only after indexing can a question name this document as a source.
Interface (API)
An interface is a route provided by the vendor through which another program may read data. Whether a program has one, and whether it is enabled and reachable on the network, depends on version, licence and set-up. This is clarified before every connection — sometimes a scheduled export is the more robust route.
Database export
A database export is a scheduled, recurring extract from a business application: the program writes files at fixed intervals to a directory from which they are read. This route is used when no interface is available. It has one advantage over a live connection: nothing changes in the running system.
Run in-house (on-premise)
Run in-house means that a program runs on a server at your own company instead of at the provider. This applies to many long-established industry programs. For connecting to a knowledge base, it means: files are encrypted and mirrored only via outbound connections, and no access to the server from outside is set up.
Separate instance
A separate instance is a distinct installation per company — its own database, its own search index, its own logins. There is no shared system in which several customers’ documents sit side by side. This also simplifies the end: when the contract ends, the documents are exported and the instance is deleted.
ERP
An ERP is a company’s commercial software: orders, purchases, stock, invoices. It knows WHAT was sold and invoiced. A knowledge base is for HOW something was done — instructions, reports, solutions to problems, correspondence. It does not replace an ERP; where a connection is possible, the ERP is only read, nothing is changed there.
Controller and processor
Under data protection law, the company is the controller: the data belongs to it, and it decides how it is used. The provider of the system is the processor and may only do what has been agreed by contract. This is governed by the data processing agreement (DPA), which sets out the instructions, the security measures and deletion.
Deletion concept
A deletion concept governs what disappears and when. In practice this means two things: deleted items are removed from the search index so they no longer appear in answers, and they are permanently deleted after a set period — immediately on request. Without the first step, a deleted document could keep appearing in answers even though nobody can open it any more.
Knowledge management
Knowledge management is how a company handles what its people know: documenting knowledge, keeping it findable and passing it on. The difficult part is rarely writing it down, but finding it again years later. That is exactly where a knowledge base comes in: instead of demanding a new filing system, it takes questions on the existing documents.
Shadow IT
Shadow IT means tools that employees use without approval because the official one is not enough — for example company documents that end up in a private AI account. The risk lies not with the people but in the gap: anyone without an approved route will find one. An approved system with permissions can close this gap.
AI literacy of employees
Since Regulation (EU) 2026/1744, the European AI Act requires companies that use AI to take measures to promote the AI literacy of their staff (Art. 4) — small ones too. A specific level of competence does not have to be guaranteed; what is meant is not a certificate but a basic understanding that can be documented: what the system can do, what it cannot, and what does not belong in it. A usage policy and documented training are the usual evidence for this.
Terms clear, benefit unclear?
In a thirty-minute video demo we show you, on examples from your industry, what the answers look like, with sources.