AI & Prompt Engineering Glossary

All terms A–Z

• Agent — An AI agent is a system that uses a language model to pursue a goal over multiple steps, deciding which actions or tools to invoke, observing the results, and adapting its plan. Unlike a single prompt-and-response call, an agent runs an iterative loop of reasoning, acting, and updating until the task is done.
• Benchmark — A benchmark is a standardized dataset and scoring method used to measure and compare model capabilities on a defined task, such as reasoning, coding, or question answering. Shared benchmarks make progress comparable across models, though results can be inflated by data contamination or narrow test design.
• Chunking — Chunking splits large documents into smaller passages before they are embedded and indexed for retrieval. Because embeddings and context windows work best over focused units, chunk size and boundaries strongly affect retrieval quality: chunks must be small enough to be specific yet large enough to stay self-contained.
• Context Engineering — Context engineering is the discipline of designing everything a model sees at inference time — system instructions, retrieved documents, tool definitions, examples, memory, and conversation history — so the right information reaches the model in the right form. It broadens prompt engineering to the full working context around a request.
• Context Window — The context window is the maximum number of tokens a language model can consider at once, spanning both the input prompt and the generated output. Everything the model can "see" for a given request — instructions, retrieved documents, conversation history, and its own response — must fit within this fixed token budget.
• Elo Rating — An Elo rating is a relative skill score derived from head-to-head comparisons, borrowed from chess and applied to models. In systems like Chatbot Arena, human voters pick the better of two model responses, and Elo converts these pairwise wins and losses into a ranking that reflects overall preference.
• Embedding — An embedding is a dense vector of numbers that represents text, images, or other data in a continuous space where semantic similarity corresponds to geometric closeness. Items with related meaning map to nearby vectors, letting systems compare, cluster, and retrieve content by measuring distances between embeddings.
• Evaluation (Evals) — Evaluation, or "evals," is the systematic measurement of a model or AI application's quality against defined criteria and test cases. Evals quantify accuracy, safety, format compliance, and task success, turning subjective impressions into repeatable metrics that guide prompt, model, and system changes.
• Few-Shot Learning — Few-shot learning gives a model a small number of worked examples in the prompt so it infers the pattern and applies it to a new input. Also called in-context learning, it steers format and behavior at inference time without updating any weights, relying on demonstrations rather than fine-tuning.
• Fine-Tuning — Fine-tuning adapts a pre-trained model to a specific task or style by continuing training on a smaller, curated dataset, updating the model's weights. Unlike prompting, which steers a frozen model at inference time, fine-tuning changes the model itself so the new behavior is baked in.
• Function Calling — Function calling lets a model request execution of a developer-defined function by emitting a structured, schema-conforming set of arguments instead of free text. The application runs the function and returns the result, enabling models to fetch live data or take actions while the developer keeps control of execution.
• Grounding — Grounding is the practice of tying a model's responses to verifiable external information — retrieved documents, tool outputs, or a knowledge base — so answers reflect real, current data rather than only the model's internal parameters. Grounding reduces hallucination and supports source attribution and citation.
• Guardrails — Guardrails are the controls placed around a model to keep its inputs and outputs within safe, policy-compliant, and on-topic bounds. They include content moderation, input and output validation, allowed-topic restrictions, schema enforcement, and human review, forming a defensive layer independent of the model's own training.
• Hallucination — A hallucination is output from a language model that is fluent and confident but factually wrong, fabricated, or unsupported by the provided sources. Because models generate the statistically likely next token rather than verified facts, they can invent citations, names, or details that never existed.
• Inference — Inference is the process of running a trained model to produce output from new input — as opposed to training, which learns the model's weights. For language models, inference means generating tokens one step at a time, and its speed and cost dominate the economics of deploying models in production.
• Jailbreak — A jailbreak is a prompt crafted to bypass a model's safety training and content policies, coaxing it to produce output it is designed to refuse. Techniques include role-play framing, obfuscation, and multi-step manipulation that reframe a disallowed request as an acceptable one.
• Knowledge Distillation — Knowledge distillation trains a smaller "student" model to reproduce the behavior of a larger "teacher" model, learning from the teacher's outputs rather than only from labels. The result is a compact model that retains much of the teacher's capability while being cheaper and faster to run.
• Latency — Latency is the time delay between sending a request to a model and receiving output. For streaming language models it is often split into time-to-first-token — how long until output begins — and per-token latency, which together determine how responsive an application feels to users.
• LoRA (Low-Rank Adaptation) — LoRA is a parameter-efficient fine-tuning method that freezes a model's original weights and injects small trainable low-rank matrices into its layers. Only these compact matrices are trained, drastically cutting memory and storage while approaching the quality of full fine-tuning for many adaptation tasks.
• Mixture of Experts (MoE) — A mixture of experts is a model architecture with many specialized sub-networks, or "experts," and a gating router that activates only a few per input. This sparse activation lets total parameter count grow large while the compute used for any single token stays modest, improving efficiency at scale.
• Model Context Protocol (MCP) — The Model Context Protocol is an open standard for connecting AI models to external tools, data sources, and systems through a common interface. It lets any compliant client and server exchange context, tools, and resources, so integrations are reusable across applications rather than rebuilt for each model or vendor.
• Multimodal — A multimodal model can process or generate more than one type of data — such as text together with images, audio, or video — within a single system. Multimodality lets a model reason across modalities, for example answering questions about a picture or describing a chart alongside text.
• Prompt Injection — Prompt injection is an attack in which crafted input causes a model to ignore its original instructions and follow the attacker's instead. Because models cannot reliably separate trusted instructions from untrusted data in the same context, malicious text — even hidden in a retrieved document — can hijack behavior.
• Prompt Template — A prompt template is a reusable prompt structure with placeholders that are filled with variable inputs at runtime. It standardizes instructions, format, and examples across many requests, so applications produce consistent prompts programmatically instead of hand-writing each one.
• Quantization — Quantization reduces the numerical precision of a model's weights and activations — for example from 16-bit floating point to 8-bit or 4-bit integers — to shrink memory use and speed up inference. It trades a small, often negligible, loss in accuracy for substantially lower cost and hardware requirements.
• Reranking — Reranking is a second-stage retrieval step that reorders an initial set of candidate documents by relevance to the query, usually with a more precise but costlier model. It refines the fast first-pass results so the most relevant passages rise to the top before being passed to a generator.
• Retrieval-Augmented Generation (RAG) — Retrieval-augmented generation combines a language model with an external retrieval step. Before generating, the system fetches relevant documents from a knowledge source and supplies them as context, so answers are grounded in specific, up-to-date data rather than relying solely on the model's parametric memory.
• RLHF (Reinforcement Learning from Human Feedback) — RLHF is a training method that aligns a model with human preferences by learning a reward model from human comparisons of outputs, then optimizing the model against that reward with reinforcement learning. It is a core technique for making models more helpful, honest, and harmless than pre-training alone produces.
• Role Prompting — Role prompting assigns the model a persona or expertise — for example an editor or a security reviewer — so it answers from that perspective. Framing the task with a role primes relevant vocabulary, standards, and tone, often improving relevance and consistency without changing the underlying request.
• Semantic Search — Semantic search retrieves content by meaning rather than exact keyword match, comparing embedding vectors of the query and documents to find conceptually related results. It surfaces relevant material even when wording differs, complementing traditional lexical search that depends on shared terms.
• Structured Output — Structured output constrains a model to return data in a specified machine-readable format, such as JSON matching a schema, rather than free-form prose. This makes responses reliably parseable by downstream code and is essential for function calling, data extraction, and integrating models into software pipelines.
• System Prompt — A system prompt is a top-level instruction that sets a model's role, persona, rules, and constraints for a conversation before any user message. It establishes standing behavior — tone, boundaries, output format, and priorities — that the model should honor across all subsequent turns.
• Temperature — Temperature is a sampling parameter that controls the randomness of a language model's output. It scales the token probability distribution before sampling: values near zero make output nearly deterministic and focused, while higher values flatten the distribution, producing more diverse and creative but less predictable text.
• Throughput — Throughput is the volume of work a model-serving system completes per unit of time, commonly measured in tokens per second or requests per second across all concurrent users. It reflects overall serving capacity and cost efficiency, in contrast to latency, which measures the delay of a single request.
• Token — A token is the unit of text a language model reads and generates. Tokenizers split text into tokens — often subword fragments rather than whole words — and each is mapped to an integer ID. Model context limits, pricing, and throughput are all counted in tokens, not characters or words.
• Tool Use — Tool use is a model's ability to call external functions, APIs, or services during a response — such as search, a calculator, a database, or code execution — instead of relying only on its own knowledge. The model decides when a tool is needed, invokes it, and incorporates the returned result.
• Top-p (Nucleus Sampling) — Top-p, or nucleus sampling, restricts token selection to the smallest set of most-probable tokens whose cumulative probability reaches a threshold p. The model then samples only from this "nucleus." Unlike a fixed top-k cutoff, the set size adapts to how confident the model is at each step.
• Vector Database — A vector database stores embeddings and retrieves them by similarity, using approximate nearest-neighbor search to find the vectors closest to a query vector quickly. It is the storage and search layer behind semantic search and retrieval-augmented generation, scaling similarity lookups to millions or billions of items.
• Vision-Language Model (VLM) — A vision-language model jointly processes images and text, letting it answer questions about pictures, read documents, describe scenes, and connect visual content with language. It typically pairs a vision encoder with a language model so visual features and text tokens are reasoned over together in one context.
• Zero-Shot Learning — Zero-shot learning asks a model to perform a task from an instruction alone, with no worked examples provided. It relies entirely on knowledge acquired during pre-training and instruction tuning, making it the simplest and cheapest baseline before adding examples or heavier prompting scaffolds.