Enterprise AI Glossary

A software architecture that uses a Large Language Model as its reasoning engine. Unlike a chatbot that only generates text, an agent can plan, call tools and act toward a goal, often with some form of memory between steps or runs.

The mechanism that allows two software components to communicate. In the context of AI, agents use APIs as their "hands" to interact with the world—such as writing a record to a CRM or sending an email through a mail service.

The process of taking a pre-trained foundation model and training it further on a smaller, specific dataset. This produces new (or adapter) weights that adopt a specific tone, format or domain behaviour. It teaches style and structure well; it is a poor way to add facts that change.

A massive AI model trained on a vast quantity of unlabeled data at scale (e.g., the GPT, Claude, Gemini and Llama families). These models are highly versatile and serve as the "foundation" upon which specific applications and agents are built.

When an LLM generates information that is factually incorrect, nonsensical, or ungrounded in the provided data. Retrieval and grounding checks reduce, but do not remove, hallucination; what they add is the ability to trace each claim to a source or flag it as unsupported.

An architectural safety mechanism where an autonomous agent prepares an action but pauses execution until a human operator reviews and explicitly approves it.

A type of artificial intelligence trained on very large amounts of text to model and generate human language. LLMs form the cognitive core of autonomous agents.

A middleware layer that dynamically routes queries to different LLMs based on cost, speed, or reasoning requirements, reducing vendor dependence and improving availability through fallback.

An architecture in which a retrieval step finds passages relevant to the query from a corpus you control and places them in the prompt, so the model can answer from current, verifiable material. The model is instructed, not compelled, to use them — which is why grounding checks exist.

The basic unit of data processed by an LLM. A token is typically a piece of a word (for English prose, roughly 100 tokens to 75 words; the ratio is worse for code and non-Latin scripts). API costs are calculated based on the number of tokens sent in the prompt and generated in the response.

A database (a dedicated product, or an extension such as pgvector on Postgres) that stores data as multi-dimensional vectors (embeddings). This allows an AI system to perform 'semantic search'—finding documents that mean the same thing, even if they don't contain the exact keywords.