A glossary of AI terms for B2B companies

A model trained on a large body of text that predicts the continuation of an input. It knows language and general knowledge from training; it does not know your company data unless you supply it.

An architecture where the system retrieves relevant passages from your documents before answering and attaches them to the question. The model then answers from your data and can cite the source.

A system that plans and carries out several steps on its own to reach a goal, including calling external tools and writing to company systems. Unlike a chatbot it does not just answer, it acts.

An interface that answers questions within a single conversation and takes no action outside it. Suited to informational questions, not to getting work done.

Further training of a model on your own data, which changes its behaviour or style. For adding factual knowledge that changes often, RAG is the better fit.

The conversion of text into a numeric vector that captures its meaning. Two texts with similar meaning have nearby vectors, which is what allows search by sense rather than by word match.

A database optimised for retrieval by semantic similarity between vectors. At smaller volumes an extension to your existing database, such as pgvector in PostgreSQL, is enough.

Splitting documents into smaller pieces before indexing. Pieces that are too large or too small both reduce retrieval precision; a badly split table is a common cause of nonsense answers.

An architecture where one infrastructure serves several clients while each client's data stays separate. The separation is enforced by metadata filtering at the retrieval layer.

Restricting retrieval by a document's tags, such as client, department or access level. It has to happen before retrieval, not after it.

A factually wrong answer delivered with high confidence. The risk drops when the answer rests on a specific source document that can be checked.

An attack where instructions are planted in input data, such as an email or a document, to push the system off its task. Defended against by separating instructions from data and limiting permissions.

The unit text is split into for the model to process, roughly a fragment of a word. API running costs are billed per token, so they rise with how much the system is used.

Running the system on your own infrastructure or within the EU instead of a public cloud API. Mainly relevant for healthcare, legal and finance, where data must not leave the controlled environment.

European regulation of AI systems introducing transparency and risk-management requirements according to a system's risk category. Some company deployments in sensitive sectors can fall into a higher category.