GLOSSARY

A language model trained on very large datasets and parameterised to process or generate language and related sequences.

LLMs can support drafting, classification, extraction, search and software tasks, but value depends on workflow design, evaluation, data governance and human oversight.

A support assistant retrieves approved product documentation, sends relevant passages to an LLM and requires cited answers with escalation for low-confidence cases.

LLMs can produce plausible errors, expose sensitive data, reproduce bias and behave unpredictably across prompts. They do not inherently verify truth, authorisation or current context.

Many current LLMs use transformer architectures and self-supervised pre-training to estimate token probabilities from context, followed by adaptation such as instruction tuning or fine-tuning. Applications wrap the model with prompts, retrieval, tools and controls.