Use these terms to ask better questions when choosing training or commissioning work. Knowing the vocabulary helps you describe requirements; it is not proof that a system is reliable.
Model, prompt, context and token
A model is the system generating or analysing content. A prompt describes the task. Context is the material available to help answer it, such as your instructions and supplied documents. Tokens are units used to process text and other inputs; they do not correspond consistently to words across languages.
Hallucination and grounding
A hallucination is a plausible but unsupported or incorrect output. Grounding means tying the answer to relevant evidence. Ask for the source passage behind an important statement, then check that the passage actually supports it. Adding documents can help, but it does not eliminate the need for review.
Retrieval, knowledge base and RAG
A knowledge base is an organised collection of source material. Retrieval finds relevant parts for a task. RAG combines retrieved material with generation. For example, a support assistant may select policy sections before drafting an answer. Outdated or conflicting sources still need an owner to resolve them.
Workflow, agent and evaluation
A workflow follows defined steps; an agent may choose steps or tools within a task. Evaluation checks whether outputs and actions meet requirements using representative examples. A useful evaluation includes missing data, wrong assumptions and failure cases, together with a clear threshold for human handover.
Keep this checklist
- Important claims checked against sources
- Knowledge base has an owner
- Actions have defined limits
- Evaluation uses realistic examples
This is an original practical guide. Examples are illustrative; tool interfaces and available features may change.