Technical terms, briefly explained

Language model (LLM)

A program that has learned which words are likely to follow one another in context, trained on very large amounts of text. From that comes the ability to understand, summarise, translate or generate text. "LLM" stands for large language model. A language model knows nothing about your company unless you give it access – that is what retrieval-augmented generation is for.

Open weights

The trained numerical values of a model are freely downloadable, so you can run the model yourself instead of using it as a service. Not the same as open source: the training data and training code usually remain undisclosed. Where data must not leave the building, this is often the decisive point.

Context window

The amount of text a model can take into account at once for a single request, measured in tokens – parts of words. If the window is too small for your documents, a selection has to be made beforehand. That is exactly what retrieval-augmented generation does.

Prompt and prompt template

The prompt is the instruction given to the model. A prompt template is a reusable pattern for it, with the concrete data inserted at runtime – comparable to a mail merge. Templates keep the instruction reproducible instead of rewriting it on every call.

Retrieval-augmented generation (RAG)

A technique for letting a model work with your own documents without retraining it. For each request, the relevant passages are retrieved from your documents and passed to the model; only then does it formulate an answer. The benefit: the answer rests on your sources and you can show where it came from. New documents are usable immediately.

Embedding

The conversion of a passage of text into a sequence of numbers that represents its meaning. Texts with similar meaning produce similar sequences, even when they share no words. That is how a search for "holiday request" also finds a document that only mentions "absence notification".

Vector database

A database built for embeddings. It does not answer "which entry contains this word" but "which entries are closest in meaning". It is the storage layer behind a RAG setup. Qdrant is one example.

Ingestion (ingestion pipeline)

"Ingestion" literally means "intake" or "loading" – an ingestion pipeline is the chain of automated steps that takes a raw document from its source all the way into the fully searchable vector database: reading it in, splitting it into chunks, converting those into embeddings, and storing them.

Function calling

Instead of only producing text, the model returns which function of your application should be called with which values. Your application performs the call, not the model – so control over what may actually happen stays with you. This is the technical basis for AI agents.

AI agent (agentic)

A language model that does not just answer but is allowed to operate tools: send an email, create a record, run a calculation. It decides for itself which tool is needed in which order. Because it has real effects, limits and approval points belong in the design from the start.

Skill

A reusable, self-contained set of working instructions for a model or an agent – comparable to a documented procedure in quality management. It defines three things:

• When it applies – how the model recognises that this is the task at hand.
• How to proceed – the steps in order, along with proven practice and typical pitfalls.
• What it works with – bundled scripts, templates, examples and reference files.

The difference from a prompt template: a template fills in a single instruction, whereas a skill bundles an entire procedure together with its tools. The practical benefit is that knowledge which otherwise sits only in individual people’s heads is written down once and is then available in the same form on every run.

Multi-agent system

Several AI agents with different jobs work together on one process – for example one that classifies a request, one that gathers documents, and one that drafts the answer. Useful for multi-step processes, because each step can be checked and improved on its own. More effort than a single agent, and not always necessary.