Reference
Enterprise AI glossary
Plain-language definitions of the terms that come up when enterprises build with AI — written by the team that builds these systems for a living.
- AI Agent
- A software system that uses a language model to choose and take actions toward a goal — calling tools, querying systems, and deciding its next step — rather than following a fixed script.
- Agentic AI
- AI systems built around agents that reason about a goal and act autonomously, adapting when reality does not match a predefined path. The engineering discipline behind them is agentic AI engineering.
- Generative AI
- AI that produces new content — text, code, images, audio — from a prompt, typically using large foundation models. In the enterprise it powers assistants, document intelligence, and content workflows.
- Large Language Model (LLM)
- A model trained on vast text data to predict and generate language. LLMs underpin most modern generative AI and AI agents.
- Retrieval-Augmented Generation (RAG)
- A technique that retrieves relevant information from your own data and supplies it to a model as context, so answers are grounded in your sources rather than the model’s general training.
- Context Engineering
- The discipline of deciding what information goes into a model’s limited context window — which knowledge to retrieve, how to compact history, which tools to expose — so it has exactly what it needs to be accurate.
- Prompt Engineering
- The practice of phrasing instructions to a language model to get better results. It is a subset of the broader work of context engineering.
- Multi-Agent System
- A system that splits a complex task across several specialised agents coordinated by an orchestrator, often with a verifier checking results before the final output.
- MLOps
- The set of practices for deploying, monitoring, and maintaining machine-learning systems reliably in production — covering evaluation, deployment, versioning, and observability.
- LLMOps
- Operational discipline applied specifically to large language model systems. It shares MLOps principles but adds challenges unique to generative AI, such as harder evaluation and cost and latency control.
- Fine-Tuning
- Further training a base model on domain-specific data to adjust its behaviour, style, or format. For changing facts, retrieval (RAG) is usually preferable; fine-tuning suits teaching style or narrow behaviours.
- Embeddings
- Numerical vector representations of text (or other data) that capture meaning, enabling similarity search — the foundation of retrieval in RAG systems.
- Vector Database
- A datastore optimised for storing and searching embeddings by similarity, used to retrieve the most relevant chunks of information at query time.
- Hallucination
- When a model generates confident but incorrect or unsupported information. It is largely an engineering problem addressed through grounding, context design, and evaluation — not an unavoidable property of models.
- AI Evaluation (Evals)
- A defined notion of correct behaviour plus a way to measure a model or agent against it continuously. Evaluation is what separates an AI system you can operate from one that merely demos.
- Tool Calling
- The mechanism by which an AI agent invokes external functions, APIs, or systems to take actions and retrieve information. Well-designed tools are the highest-leverage part of building reliable agents.
- AI Readiness
- A measure of whether an organisation is positioned to succeed with AI across dimensions such as strategy, data, talent, platform, governance, and operating model.
- AI Proof of Concept (PoC)
- A scoped build that tests whether an AI use case is worth investing in by confronting its riskiest assumptions — data quality, accuracy, integration — against real conditions, with success criteria set in advance.
- Agentic Coding
- Software development where an autonomous AI agent plans, writes, tests, and iterates on code in a loop with limited human input, while a developer reviews the resulting diff against acceptance criteria.
- Knowledge Graph
- A representation of data as entities (nodes) connected by labelled relationships (edges), enabling queries that follow relationships across multiple hops rather than matching isolated text.
- GraphRAG
- Retrieval-augmented generation that uses a knowledge graph, often alongside vector search, so a model can follow relationships and answer multi-hop and corpus-wide questions that plain vector RAG handles poorly.
- Hybrid Search
- A retrieval approach that combines lexical search (keyword matching, e.g. BM25) with semantic search (vector similarity), then fuses and reranks results to capture both exact terms and meaning.
- Reranking
- A second-stage step that re-orders an initial set of retrieved candidates by true relevance to the query, improving the precision of the final results passed to a model.
- KV Cache
- The memory a language model uses to store context it has already processed during generation. Its size is a major limit on context length and inference cost, making it a target for compression.
- Quantization
- Reducing the numerical precision of a model’s values (for example from 16-bit to a few bits) to cut memory and compute cost while aiming to preserve accuracy. Applied to weights or to the KV cache.
Want the deeper version?
Our guides go beyond definitions into how these systems are actually built and run.
Read the guides