A–C
- Agentic AI Architecture
- An AI system that can take sequences of actions autonomously to complete a goal — browsing the web, writing and executing code, calling APIs, or managing files — rather than simply responding to a single prompt. Agentic AI introduces significant governance considerations because it acts rather than advises. Example: An AI agent that receives "research and summarise our competitors" and independently searches, reads, and compiles a report without step-by-step human instruction.
- Alignment Safety
- The challenge of ensuring an AI system's goals, values, and behaviours match the intentions of its designers and the needs of its users. A misaligned AI may be technically capable but pursue objectives that are subtly or seriously wrong. Alignment research is one of the central concerns of AI safety. See also: RLHF, Constitutional AI
- Attention Mechanism Architecture
- The core innovation in transformer models — a mathematical process that allows the model to weigh the importance of different parts of the input when producing each part of the output. Attention lets models understand context and relationships across long sequences of text, rather than processing tokens one by one in isolation.
- Bias (AI) Ethics
- Systematic errors in AI output that result in unfair or inequitable treatment of different groups. Bias can originate in training data (if certain groups are overrepresented or misrepresented), in model architecture, or in how outputs are applied. AI bias is not always obvious — it can be statistical, cultural, or linguistic. See also: Fairness, Training Data
- Chain-of-Thought Prompting Prompting
- A prompting technique where the model is instructed — or shown by example — to work through a problem step by step before providing a final answer. Chain-of-thought significantly improves accuracy on reasoning, maths, and logic tasks. Example: "Think through this step by step, then give your final answer."
- Chunking RAG
- The process of splitting large documents into smaller segments before storing them in a vector database for use in RAG (Retrieval Augmented Generation) systems. Chunk size and strategy (fixed, sentence, paragraph, semantic) significantly affect retrieval quality. Too large loses precision; too small loses context. See also: RAG, Embeddings, Vector Database
- Constitutional AI Safety
- Anthropic's technique for training AI models to be helpful, harmless, and honest using a written set of principles (the "constitution"). Rather than only using human feedback, the model is also trained to critique and revise its own outputs according to these principles. Claude is trained using Constitutional AI.
- Context Window Models
- The maximum amount of text (measured in tokens) that a model can consider at one time — including both the input and its own output. A model cannot "remember" text outside its context window. Larger context windows allow longer documents, more history, and more complex tasks. Example: GPT-4o has a 128,000-token context window — roughly 100,000 words.
D–F
- Diffusion Model Image AI
- A type of generative AI model trained to create images by learning to reverse a process of adding noise. Starting from random noise, the model progressively refines toward a coherent image guided by a text prompt. Stable Diffusion and DALL-E use this approach. See also: DALL-E, Stable Diffusion
- Embedding RAG
- A numerical representation of text (or other data) as a vector — a list of floating-point numbers — that captures semantic meaning. Similar texts have similar embeddings; dissimilar texts have embeddings that are far apart in vector space. Embeddings are the foundation of semantic search and RAG systems. See also: Vector Database, RAG, Chunking
- Few-Shot Prompting Prompting
- A prompting technique where the model is given a small number of examples (typically 2–5) of the desired input/output pattern before being asked to complete a new instance. Fewer examples means fewer tokens consumed and faster responses. See also: Zero-Shot, One-Shot, Chain-of-Thought
- Fine-Tuning Training
- The process of taking a pre-trained foundation model and continuing its training on a smaller, task-specific dataset to specialise its behaviour. Fine-tuning can improve performance on narrow domains but requires significant data, compute, and care to avoid degrading general capability. Contrast with: RAG, which adds knowledge at inference time without retraining.
- Foundation Model Models
- A large AI model trained on massive, broad datasets that can be adapted (via fine-tuning or prompting) to a wide range of downstream tasks. Claude, GPT-4, Gemini, and Llama are all foundation models. The term emphasises that these models serve as a "foundation" for building more specialised applications.
- Function Calling / Tool Use Architecture
- A capability where an AI model can invoke structured external tools — APIs, database queries, calculators, code executors — as part of completing a task. The model outputs a structured call (e.g. a JSON object), the application executes it, and returns the result back to the model. Central to agentic AI systems. See also: Agentic AI, MCP
G–I
- Generative AI Category
- AI systems that produce new content — text, images, code, audio, video, or other data — rather than simply classifying or analysing existing content. Generative AI learns patterns from training data and generates novel outputs that match those patterns. ChatGPT, DALL-E, Stable Diffusion, and Claude are all generative AI systems.
- GPU (Graphics Processing Unit) Infrastructure
- Specialised hardware originally designed for rendering graphics but now the dominant compute substrate for training and running AI models. GPUs can perform thousands of parallel mathematical operations simultaneously, making them far more efficient than CPUs for the matrix calculations that power neural networks. Training frontier models requires thousands of high-end GPUs.
- Grounding Accuracy
- The process of connecting an AI model's outputs to verified, reliable sources of information — reducing hallucination and increasing factual accuracy. RAG is a grounding technique. Grounded AI systems cite their sources and can be verified; ungrounded systems generate plausible-sounding text that may have no factual basis. See also: RAG, Hallucination
- Guardrails Safety
- Constraints applied to an AI system to prevent it from producing harmful, inappropriate, or policy-violating outputs. Guardrails can be implemented via training (teaching the model what not to do), system prompts (instructing the model at runtime), or external filters (checking outputs before they reach users). See also: Alignment, Prompt Injection
- Hallucination Risk
- When an AI model generates factually incorrect, fabricated, or nonsensical content with apparent confidence — presenting fiction as fact. Hallucination occurs because models generate statistically probable continuations, not verified truths. It is a fundamental property of current LLMs, not a bug that will simply be "fixed." Mitigation strategies include RAG, chain-of-thought, and human review. See also: Grounding, RAG
- Human-in-the-Loop (HITL) Governance
- A design pattern where human judgment and approval is required at one or more stages of an AI-assisted process — particularly before consequential actions are taken. HITL systems keep humans as the ultimate decision-makers while using AI to augment capability and speed. The opposite of fully automated AI pipelines. See also: Agentic AI, Alignment
- Inference Models
- The process of running a trained AI model on new inputs to generate outputs. Inference is distinct from training — training builds the model, inference uses it. When you send a prompt to Claude or GPT, inference is happening. Inference can run on cloud infrastructure or locally (on-device inference). Contrast with: Training, Fine-Tuning
- In-Context Learning Prompting
- A model's ability to learn from examples and instructions provided within the prompt itself — without any weight updates or fine-tuning. Few-shot and chain-of-thought prompting are forms of in-context learning. Models with larger context windows can leverage more in-context examples. See also: Few-Shot, Context Window
J–L
- JSON Mode / Structured Output Integration
- A model capability (supported by OpenAI, Anthropic, Gemini) that constrains the model to produce output conforming to a specified JSON schema. Critical for reliable AI integration in applications — prevents freeform text that would break downstream parsing and enables schema validation. See also: Function Calling, Tool Use
- Knowledge Base RAG
- A curated collection of documents, facts, or structured data used to ground an AI system's responses. In RAG systems, documents are chunked, embedded, and stored in a vector database — then retrieved at query time to provide the model with relevant context it wasn't trained on. See also: RAG, Embeddings, Chunking
- Large Language Model (LLM) Models
- A type of AI model trained on massive text datasets to understand and generate human language. LLMs predict the most likely next token given a context, and this simple mechanism — scaled to billions of parameters — produces surprisingly capable language understanding and generation. Claude, GPT-4, Gemini, and Llama are all LLMs. See also: Foundation Model, Tokens, Transformer
- Latency Performance
- The time between sending a request to an AI model and receiving the first response. Latency is affected by model size, infrastructure, and token count. SSE streaming reduces perceived latency by delivering tokens as they are generated, rather than waiting for the full response. Critical for real-time AI applications.
- Llama Open Source
- Meta's family of open-source large language models, available in sizes from 1B to 405B parameters. Llama 3 is among the most capable open-weight models available, suitable for deployment on local hardware via Ollama or Hugging Face. Open-weight means the model weights are publicly available, though subject to a licence. See also: Ollama, Open-Weight Models
M–O
- MCP (Model Context Protocol) Architecture
- An open standard developed by Anthropic for connecting AI models to external tools, data sources, and services in a standardised way. MCP enables AI assistants to read files, query databases, call APIs, and take actions without each integration requiring custom code. Think of it as a universal plugin system for AI. See also: Agentic AI, Function Calling
- Multimodal AI Models
- AI models capable of processing and generating multiple types of data — text, images, audio, video — rather than a single modality. GPT-4o, Gemini, and Claude 3+ are multimodal: they can reason about images alongside text. Multimodal models enable use cases like document analysis (text + layout), visual Q&A, and image generation. See also: Vision Models
- Neural Network Fundamentals
- A computational system loosely inspired by the structure of the biological brain — interconnected layers of mathematical nodes (neurons) that transform inputs into outputs. Deep neural networks (many layers) are the foundation of modern AI. Training adjusts the connection weights between neurons to minimise prediction error on training data.
- Ollama Infrastructure
- An open-source tool that allows developers to run large language models locally on their own hardware — laptops, servers, or cloud instances — without sending data to an external API. Supports Llama, Mistral, Phi, Gemma, DeepSeek, and hundreds of other models. A critical enabler of data-sovereign AI deployments. See also: Local AI, Open-Weight Models
- One-Shot Prompting Prompting
- A prompting technique where a single example of the desired input/output pattern is provided before the model is asked to complete a new instance. Sits between zero-shot (no examples) and few-shot (several examples). See also: Zero-Shot, Few-Shot
- Open-Weight Model Models
- An AI model whose trained weights (parameters) are publicly available for download and local deployment, as opposed to closed/proprietary models accessible only via API. Open-weight models include Llama, Mistral, Phi, and Gemma. Note: "open weight" and "open source" are often used interchangeably but are technically distinct — some open-weight models have licence restrictions on commercial use.
P–R
- Parameters Models
- The numerical weights inside a neural network that are learned during training and determine the model's behaviour. "7B parameters" means 7 billion individual numbers that collectively encode the model's knowledge. More parameters generally means more capability but also more compute, memory, and inference cost. See also: Neural Network, Foundation Model
- Perplexity Evaluation
- A measure of how well a language model predicts a sample of text — essentially, how "surprised" the model is by the text. Lower perplexity means the model finds the text more predictable, indicating it has learned the patterns well. Used as an evaluation metric during model training.
- Prompt Prompting
- The input text given to an AI model — including instructions, context, examples, and the user's actual request. Prompt design (prompt engineering) is one of the most impactful ways to improve AI output quality without changing the model. A well-designed prompt is explicit, contextual, and specifies the desired output format. See also: System Prompt, Few-Shot, Chain-of-Thought
- Prompt Engineering Prompting
- The practice of crafting AI prompts to reliably produce high-quality, appropriate outputs. Includes techniques like role assignment, few-shot examples, chain-of-thought instructions, output format specification, and negative constraints. Prompt engineering is a core skill for anyone building AI-powered applications.
- Prompt Injection Security
- An attack where malicious instructions embedded in user input or external data attempt to override an AI model's system prompt and instructions — causing the model to behave outside its intended scope. The AI equivalent of SQL injection. Mitigated by input sanitisation, server-side system prompts, and output validation. See also: Guardrails, System Prompt
- RAG (Retrieval Augmented Generation) Architecture
- A technique that combines an AI model's generation capabilities with real-time retrieval of relevant information from a knowledge base. Instead of relying solely on training data, the model is given retrieved context (chunks of relevant documents) alongside the user's query — grounding its response in verified information and reducing hallucination. See also: Embeddings, Vector Database, Chunking, Grounding
- RLHF (Reinforcement Learning from Human Feedback) Training
- A training technique where human raters evaluate AI outputs and this feedback is used to train a reward model, which then guides further model training. RLHF is how most major chat models (ChatGPT, Claude, Gemini) are made helpful, safe, and aligned with human preferences rather than just statistically probable. See also: Alignment, Constitutional AI
S–U
- Semantic Search RAG
- Search that finds content based on meaning and intent rather than exact keyword matching. Powered by embeddings and cosine similarity: queries and documents are converted to vectors, and the most semantically similar documents are returned — regardless of whether they share exact words. Central to RAG retrieval. See also: Embeddings, RAG, Cosine Similarity
- SSE (Server-Sent Events) Integration
- A web standard for streaming data from server to browser over a single HTTP connection. Used in AI chat applications to stream model tokens as they are generated — rather than waiting for the complete response. SSE dramatically improves perceived performance and user experience for AI-powered interfaces. See also: Streaming, Latency
- Stable Diffusion Image AI
- An open-source image generation model developed by Stability AI, based on diffusion model architecture. Notable for being openly available — it can be run locally on consumer hardware. Widely used in image generation applications, often alongside commercial models like DALL-E. See also: Diffusion Model, DALL-E
- Streaming Integration
- In AI APIs, streaming refers to receiving model output token-by-token as it is generated, rather than waiting for the complete response. Implemented via SSE or WebSockets. Streaming makes AI interfaces feel responsive and allows users to read and act on output before it is complete. All major AI APIs support streaming. See also: SSE, Latency, Tokens
- System Prompt Prompting
- A special instruction given to an AI model before the user's message — typically set by the developer or application, not visible to end users — that defines the model's persona, capabilities, constraints, and behaviour. System prompts are how applications control and customise AI behaviour. In VoxMachina applications, system prompts are hardcoded server-side to prevent prompt injection. See also: Prompt, Prompt Injection, Guardrails
- Temperature Models
- A parameter that controls the randomness of an AI model's output. Low temperature (e.g. 0.1) makes outputs more deterministic and predictable — the model reliably picks the most probable token. High temperature (e.g. 1.0) introduces more randomness, producing more diverse, creative, but potentially less accurate outputs. For factual or structured tasks, lower temperature is preferred. See also: Top-P, Tokens
- Tokens Models
- The basic units an AI model reads and writes — roughly equivalent to parts of words. "ChatGPT" might be two tokens: "Chat" and "GPT". Most English words are 1–2 tokens; code and rare words may be more. AI API pricing is based on tokens consumed (input + output). Context window limits are also measured in tokens. 1,000 tokens ≈ 750 words.
- Transformer Architecture
- The neural network architecture introduced in the 2017 paper "Attention Is All You Need" that underpins virtually all modern large language models. The transformer's key innovation — the attention mechanism — allows models to relate any part of the input to any other part, enabling understanding of long-range context. GPT, Claude, Gemini, Llama, and Mistral are all transformer-based. See also: Attention Mechanism, LLM
- Transfer Learning Training
- The practice of taking a model trained on a large general dataset and adapting it (via fine-tuning) to a more specific task or domain. Transfer learning is why fine-tuning a foundation model is vastly more efficient than training from scratch — the model already "knows" language, structure, and general world knowledge. See also: Fine-Tuning, Foundation Model
- Top-P (Nucleus Sampling) Models
- A sampling parameter that controls the diversity of AI output by limiting the pool of candidate next tokens to the smallest set whose cumulative probability exceeds a threshold P. At Top-P = 0.9, the model considers only the tokens that together account for 90% of the probability mass — filtering out very unlikely tokens. Used alongside temperature to tune output style. See also: Temperature
- Tool Use
- See Function Calling.
- Unstructured Data Data
- Information that does not follow a predefined data model — PDFs, emails, images, audio recordings, social media posts, and free-form text. The majority of enterprise data is unstructured. AI (particularly LLMs and multimodal models) is uniquely capable of extracting meaning from unstructured data, making it one of the most impactful application areas for AI in business.
V–Z
- Vector Database RAG
- A database optimised for storing and querying high-dimensional vectors (embeddings). Rather than looking up rows by ID or matching keywords, vector databases find the most semantically similar entries using distance metrics like cosine similarity. Used in RAG systems to retrieve relevant document chunks. Examples: Pinecone, Weaviate, Chroma, pgvector. See also: Embeddings, RAG, Semantic Search
- Vision Model Models
- An AI model capable of understanding and reasoning about images as well as text. Vision models can describe images, extract text from documents, identify objects, read diagrams, and answer questions about visual content. Modern multimodal LLMs (Claude 3+, GPT-4o, Gemini) include vision capability. See also: Multimodal AI
- Weights Models
- The numerical parameters in a neural network that are learned during training and define the model's behaviour. A model's weights are what get downloaded when you use an open-weight model locally. Changing weights requires retraining or fine-tuning; prompting and RAG work at inference time without touching weights. See also: Parameters, Fine-Tuning, Open-Weight Model
- Zero-Shot Prompting Prompting
- Asking an AI model to perform a task with no examples provided — relying entirely on its training to interpret and complete the request. Modern frontier models are highly capable zero-shot performers for common tasks. For novel, specialised, or format-sensitive tasks, few-shot examples typically improve results. See also: Few-Shot, One-Shot, In-Context Learning
- Zero-Retention API Security
- An API configuration (available from Anthropic and OpenAI at enterprise tier) where prompt and response data is not stored by the AI provider after the request is processed. Critical for organisations processing sensitive, confidential, or personally identifiable information via AI APIs. Ask your AI provider explicitly whether zero-retention is available and active on your account. See also: Data Sovereignty, POPIA
info
This glossary is maintained by the VoxMachina team and updated as the field evolves. If you'd like to suggest a term, correction, or addition, contact us. For deeper explanations of any of these concepts, explore our AI Academy courses.