Glossary

AI terms for people who build websites. Every entry lists its sources and was checked against them.

Fundamentals

Context window

Also: working memory

The amount of text a language model can look back on and reference when generating new text, functioning as its working memory rather than the training corpus. A larger context window enables processing longer prompts and maintaining coherence over extended conversations; a smaller one may limit these abilities.

Embeddings

Also: vector embeddings, text embeddings

A text embedding is a vector (list) of floating-point numbers that represents a text string. The distance between two vectors measures their relatedness, with small distances indicating high relatedness. Embeddings are used for search, clustering, recommendations, anomaly detection, diversity measurement, and classification.

Fine-tuning

Also: model fine-tuning

Further training a pretrained language model on additional data to adapt it to a specific domain, task, or writing style.

Hallucination

Also: confabulation

Hallucination is when a language model generates text that is factually incorrect or inconsistent with the given context. It can undermine the reliability of AI-driven solutions. Techniques like allowing uncertainty, grounding responses in direct quotes, and verifying claims with citations can reduce hallucinations.

Inference

Also: AI inference

Inference is the process of making predictions by applying a trained model to unlabeled examples. It can be done statically (offline/batch), where predictions are cached, or dynamically (online/real-time), where predictions are made on demand.

Large Language Model (LLM)

Also: LLM

Large language models (LLMs) are AI language models with many parameters, trained on vast amounts of text data, capable of generating human-like text, answering questions, summarizing information, and more.

See also: Fine-tuning, Token, Context window.

Multimodal model

Also: multimodal AI, multimodal large language model

A model that can process and generate multiple data types, such as text and images. For example, recent language models can analyze image inputs (vision) and generate images from text prompts.

Open-weight model

Also: open weights, open-weights model

An open-weight model is an AI model whose trained weights and biases are publicly available, allowing others to fine-tune, adapt, or deploy it. However, it typically does not include training code, the training dataset, or comprehensive data transparency, so it falls short of open source AI.

Prompt

A prompt is a natural language instruction that tells a generative AI model what task to perform, guiding the structure and content of its output. It can include variables and context data to provide dynamic information at runtime.

Reasoning model

Reasoning models use internal reasoning tokens before producing a response, helping the model plan, use tools effectively, inspect alternatives, recover from ambiguity, and solve harder multi-step tasks. They are especially suited for complex problem solving, coding, scientific reasoning, and multi-step agentic workflows.

Retrieval-Augmented Generation (RAG)

Also: RAG

Retrieval augmented generation (RAG) is a technique that combines information retrieval with language model generation, retrieving relevant information from an external knowledge base at query time and passing it to the model along with the original query to improve accuracy and ground responses in evidence.

System prompt

Also: system message, system instruction

A system prompt is an instruction provided to an AI model at the start of a conversation to set its behavior, role, or tone. Even a single sentence can make a difference in focusing the model's responses for a specific use case.

Temperature

Temperature is a parameter that controls the randomness of a model's predictions during text generation. Higher temperatures lead to more creative and diverse outputs, while lower temperatures result in more conservative and deterministic outputs.

Token

Also: tokens

Tokens are the smallest individual units a language model processes, which can correspond to words, subwords, characters, or bytes. For Claude, a token approximately represents 3.5 English characters, and text is encoded into tokens before model processing.

Vector database

Also: vector DB

A vector database indexes and stores vector embeddings for fast retrieval and similarity search, offering CRUD operations, metadata filtering, horizontal scaling, and serverless.

Agents

Agent2Agent (A2A) protocol

Also: A2A

An open standard for communication between AI agents, giving them a common language regardless of framework or vendor.

Agentic workflow

Also: workflow

An agentic workflow orchestrates LLMs and tools through predefined code paths, offering predictability and consistency for well-defined tasks. It contrasts with agents, where LLMs dynamically direct their own processes and tool usage.

AI agent

Also: LLM agent, autonomous agent

An AI agent is a system where a large language model dynamically directs its own processes and tool usage, planning and operating independently rather than following predefined code paths.

MCP client

MCP clients are instantiated by host applications to communicate with particular MCP servers. Each client handles one direct communication with one server, while the host application coordinates multiple clients.

MCP host

An MCP host is an AI application that coordinates and manages one or more MCP clients. It establishes connections to MCP servers by creating one MCP client for each server. Examples include Claude Code, Claude Desktop, and Visual Studio Code.

MCP server

Also: Model Context Protocol server

An MCP server is a program that exposes specific capabilities—tools, resources, and prompts—to AI applications through the Model Context Protocol's standardized interfaces.

Model Context Protocol (MCP)

Also: MCP

MCP (Model Context Protocol) is an open-source standard for connecting AI applications like Claude or ChatGPT to external systems such as data sources, tools, and workflows. It uses a client-server architecture where an MCP host creates MCP clients to maintain connections to MCP servers, enabling AI apps to access key information and perform tasks.

Structured outputs

Also: structured output, JSON Schema output

A feature that ensures model responses adhere to a supplied JSON Schema, guaranteeing valid typing, required keys, and allowed enum values. It removes the need to validate or retry incorrectly formatted responses.

Related tool: Structured Data Validator. See also: Grounding.

Tool use (function calling)

Also: function calling, tool calling

Tool use (function calling) lets a model call external functions or tools you define or a provider hosts. The model decides when to invoke a tool based on the user request and tool description, then returns a structured call that your application or the provider executes before continuing.

AI search

AI crawler

Also: AI web crawler, AI robot, AI spider

An AI crawler is a web crawler used by AI products to automatically discover and scan websites. It can serve search features, ad safety checks, or training generative AI models, and is typically managed via robots.txt.

AI Mode

An AI-powered Google Search feature that provides a comprehensive response with links to supporting websites for nuanced questions requiring exploration, reasoning, or complex comparisons.

AI Overviews

AI Overviews are a Google Search feature that provides a concise AI-generated summary for complex queries, with links to supporting websites. They are shown only when deemed additive to classic search results.

Guide: read the guide. See also: Generative Engine Optimization (GEO), Grounding.

Answer Engine Optimization (AEO)

Answer engine optimization (AEO) refers to work specifically focused on improving visibility in AI search experiences. Google considers optimizing for generative AI search as part of SEO.

Guide: read the guide. See also: Generative Engine Optimization (GEO), AI Overviews.

Generative Engine Optimization (GEO)

Also: GEO

Generative Engine Optimization (GEO) is the practice of improving content visibility in responses generated by AI-powered search engines.

Guide: read the guide. See also: AI Overviews, AI crawler.

Google-Extended

Google-Extended is a robots.txt user-agent token that lets site owners control whether Google-crawled content may be used to train Gemini models or for grounding in Gemini Apps and Vertex AI. It does not affect a site's inclusion or ranking in Google Search.

Related tool: AI Crawler Access Checker. See also: AI crawler, Training vs. retrieval crawlers.

Grounding

Also: grounding with Google Search, google search grounding

Grounding connects a language model to real-time web content through tools like Google Search, reducing hallucinations by basing responses on verifiable sources and providing citations.

Related tool: Structured Data Validator. See also: Structured outputs, AI Overviews.

LLM citation

Also: citations, source citations

A feature that lets an LLM return the exact passage supporting each claim, along with a pointer to its location in a source document. This enables verification and can surface sources to users.

llms.txt

An /llms.txt file is a markdown file placed on a website to provide LLM-friendly content, including a concise project summary and curated links to detailed markdown pages, so agents can efficiently find and fetch relevant information.

Related tool: AI Crawler Access Checker. See also: AI crawler.

Query fan-out

Also: query fanout

A technique used by AI search features to perform multiple related searches across subtopics and data sources, helping generate more comprehensive responses.

Guide: read the guide.

Training vs. retrieval crawlers

Training crawlers, like GPTBot, crawl web content that may be used to train generative AI foundation models. Retrieval crawlers, like OAI-SearchBot, fetch pages to surface them in AI search features; site owners can allow or block each independently via robots.txt.

Related tool: AI Crawler Access Checker. See also: AI crawler, Google-Extended.

Infrastructure

Batch API

Also: Batches API, Message Batches API

An API endpoint that accepts a file of multiple requests and processes them asynchronously, typically offering lower costs and higher rate limits than synchronous calls, with results returned within 24 hours.

Model routing

Also: model router, intelligent prompt routing

Model routing is the process of using a trained model or lightweight ML model to analyze each prompt in real time and route it to the most suitable large language model based on complexity, reasoning, task type, and other attributes. It optimizes costs and latencies while maintaining quality by selecting cheaper models for simple tasks and more powerful models for complex tasks.

Prompt caching

Also: context caching

Prompt caching stores reusable prompt prefixes so subsequent requests with the same prefix can skip reprocessing, reducing latency and cost. Cached prefixes have a time-to-live (commonly 5 minutes or 1 hour) and can be controlled automatically or with explicit breakpoints.

Quantization

Also: 8-bit quantization, integer quantization, low-precision quantization

Quantization reduces the computational and memory costs of running AI models by representing weights and activations with low-precision data types, such as 8-bit integers instead of 32-bit floats.

Rate limit

Also: rate limiting

Rate limits set the maximum number of API requests an organization can make over a defined period of time. They are enforced using the token bucket algorithm, with capacity continuously replenished up to the maximum limit. Exceeding a rate limit returns HTTP 429 with a retry-after header.

Tokens per minute (TPM)

Also: TPM

Tokens per minute (TPM) is a rate limit that measures combined input and output token usage over one minute. Many API providers use a single TPM limit that may include all tokens, including cached ones; Claude instead tracks input tokens per minute (ITPM) and output tokens per minute (OTPM) separately.

Security

Data poisoning

Also: data and model poisoning, model poisoning

Data poisoning is an integrity attack where pre-training, fine-tuning, or embedding data is manipulated to introduce vulnerabilities, backdoors, or biases, compromising model security, performance, or ethical behavior.

Expired domain abuse

Expired domain abuse is when an expired domain name is purchased and repurposed primarily to manipulate search rankings by hosting content that provides little to no value to users. Examples include affiliate content on a site previously used by a government agency or casino-related content on a former elementary school site.

Indirect prompt injection

Indirect prompt injection occurs when an LLM processes input from external sources such as websites or files, and embedded content in that data alters the model's behavior in unintended or unexpected ways. The injection can be intentional or unintentional.

Jailbreak

Also: jailbreaking

Jailbreaking is a form of prompt injection where an attacker crafts inputs that cause a language model to disregard its safety protocols or guidelines.

Prompt injection

Also: prompt injection attack, LLM prompt injection

A prompt injection vulnerability occurs when user prompts alter an LLM’s behavior or output in unintended ways. Inputs can be direct or indirect, and may be imperceptible to humans.

Scaled content abuse

Scaled content abuse is the practice of generating many pages or pieces of content at scale primarily to manipulate search rankings, rather than providing original value to users.