AI & LLM Glossary

Agent-to-Agent protocol for agents to discover, negotiate, and delegate tasks.

Generating a summary using new phrasing rather than extracting sentences.

The proportion of correct predictions out of total predictions.

A mathematical function applied to a neuron's output, introducing non-linearity so the network can learn complex patterns.

A small neural module inserted into a pre-trained model and trained for a specific task.

An input designed to trick a model into making a mistake.

An AI agent is a system that uses a language model to perceive its environment, make decisions, and take actions to reach a goal. Unlike a simple chatbot that only responds to prompts, an agent can loop: observe state, plan next steps, call tools, review results, and adapt until the task is done. Agents are built from several components. A planner breaks a goal into subtasks. A memory module stores conversation history and working context. A tool interface lets the agent call APIs, run code, query databases, or interact with other systems. A feedback loop checks whether each step moved the agent closer to the goal. Simple agents might answer a question by searching the web. Complex agents can write and test code, file pull requests, or coordinate with other agents. The more autonomy an agent has, the more important safety guardrails become, such as human approval for destructive actions and clear logging for every decision.

A manifest describing an A2A agent's capabilities, endpoint, and authentication requirements.

A tool that helps write, review, or debug code using AI.

The process of ensuring a model behaves in ways consistent with human values and intentions.

The process of adding labels or metadata to training data.

Application Programming Interface, a set of rules for software components to communicate.

The cost structure for using a model or service via API, usually per input and output tokens.

Hypothetical AI that can understand, learn, and perform any intellectual task a human can do across any domain.

The broad field of creating machines that can perform tasks requiring human-like intelligence, such as reasoning, learning, and perception.

An AI application that helps users complete tasks through conversation.

A mechanism that lets a model focus on relevant parts of input when producing each output token.

The degree to which an agent can operate without human intervention.

Generating output one token at a time, using previously generated tokens as context for the next.

The algorithm used to train neural networks by propagating error gradients backward and updating weights.

A subset of training data processed together in one forward and backward pass.

Sending groups of requests together, often at a lower price but with higher latency.

The number of training examples used in a single update step.

A standardized dataset and task used to compare models.

A learnable value added to a neuron's input before activation, helping the model fit data better.

A metric that evaluates the quality of generated text by comparing it to reference text.

A ranking function used in keyword search to estimate document relevance.

Byte Pair Encoding, a subword tokenization algorithm used by many language models.

Storing and reusing previous results to reduce latency and cost.

Content Delivery Network, a geographically distributed network that speeds up content delivery.

Prompting a model to show its reasoning step by step before giving a final answer.

A conversational interface that uses AI to interact with users.

Splitting documents into smaller pieces before embedding and storing them.

Continuous Integration and Continuous Deployment, automated pipelines for building and releasing software.

When some classes or outcomes are far more common than others in training data.

Assigning input data to predefined categories.

An agentic terminal coding tool powered by Claude.

Grouping data points into clusters based on similarity.

Running code generated by a model, usually in a controlled environment.

Difficulty in making predictions for new users or items with little historical data.

Making recommendations based on patterns across many users.

The text generated by a model in response to a prompt.

A training approach where models critique and revise their own outputs according to a set of principles.

A system that blocks or flags disallowed content.

Making recommendations based on attributes of items a user has liked.

The context window is the maximum number of tokens a model can consider in a single forward pass. It includes the system prompt, user messages, retrieved documents, and the model's own generated output. If the total exceeds the window, the oldest tokens are dropped or the request fails. Context windows vary widely. Small models may handle 4,000 tokens, while frontier models can process 128,000, 1,000,000, or even 10,000,000 tokens. Long context is useful for summarizing books, analyzing large codebases, and holding extended conversations without losing earlier details. A larger window does not always mean better results. Very long inputs can dilute attention, making the model miss important details. Techniques like RAG, selective summarization, and hierarchical chunking help fit the most relevant information into the window without exceeding the limit.

Further pre-training a model on additional domain-specific data before fine-tuning.

AI systems designed for natural language dialogue.

An AI assistant embedded in a workflow to augment human work.

A common pricing unit for API-based language models.

Training a model on easier examples first and gradually increasing difficulty.

An AI-native code editor built on VS Code with strong agentic features.

Creating additional training examples from existing data to improve model robustness.

When information from outside the training set inappropriately influences model training.

The process of finding and fixing errors in software.

A model component that generates output sequences from an encoded representation.

A branch of machine learning based on multi-layer neural networks that can learn complex patterns from large amounts of data.

A model where all parameters are active during every forward pass.

Practices that combine software development and IT operations to shorten delivery cycles.

A generative model that learns to reverse a noise-adding process to create images, audio, or video.

Adapting a model to perform better on data from a specific domain or industry.

A regularization technique that randomly disables neurons during training to prevent co-adaptation.

Running models close to end users to reduce latency.

An embedding is a list of numbers, usually called a vector, that represents the meaning of a piece of data. Semantically similar items end up close together in this numeric space, which lets a computer compare meaning using distance rather than exact keyword matches. Embedding models are trained to produce these vectors. For text, the model reads a sentence or document and outputs a dense vector, often with hundreds or thousands of dimensions. You can then measure similarity with cosine similarity or Euclidean distance. Two sentences about payment processing will have embeddings closer to each other than a sentence about baseball, even if they share no common words. Embeddings power search, recommendations, clustering, and RAG. A typical RAG system stores document chunks as vectors in a vector database and retrieves the nearest neighbors to a user's query embedding. Embeddings can also represent images, audio, and other modalities when the model is trained on multimodal data.

A model that converts data into dense numerical vectors that capture semantic meaning.

A model component that processes input into a dense internal representation.

An architecture that first encodes input into a representation and then decodes it into output.

A distinct object or concept represented in a knowledge graph, such as a person, place, or product.

One complete pass through the entire training dataset during model training.

The process of measuring a model's performance on tasks or benchmarks.

The degree to which a model's decisions can be understood by humans.

Creating a summary by selecting existing sentences or phrases from the source.

The harmonic mean of precision and recall.

The degree to which generated content is factually correct.

The property of a model treating different groups equitably.

An individual measurable property or characteristic of data used by a model.

Transforming raw data into features that improve model performance.

Learning a task from only a few examples, often by including them in the prompt.

Including examples of desired input-output pairs in the prompt to guide the model.

Fine-tuning is the process of further training a pre-trained model on a smaller, task-specific dataset so it becomes better at a particular job. The base model already knows grammar, facts, and reasoning from pre-training; fine-tuning teaches it the style, format, or domain you care about. Common reasons to fine-tune include matching a brand voice, classifying support tickets, extracting structured fields from documents, and improving performance on low-resource languages. You typically need hundreds to thousands of high-quality examples. Each example pairs an input with the desired output, and the model's weights are updated to reduce the error on those examples. Fine-tuning is not always the right first step. Prompt engineering, retrieval augmentation, and few-shot examples are faster and cheaper to iterate. Fine-tuning becomes worthwhile when the behavior you want is hard to describe in a prompt, must be consistent at scale, or needs to run without sending long examples every request. Techniques like LoRA and QLoRA make fine-tuning feasible on consumer hardware by updating only a small subset of weights.

The time until the first token of a response is received.

A large model trained on broad data that can be adapted to many downstream tasks.

A model capability to generate calls to external functions with structured arguments.

A model's ability to perform well on new, unseen data rather than only memorizing training examples.

AI systems that create new content such as text, images, audio, or code based on learned patterns.

A binary format for storing quantized models for efficient local inference.

An AI pair programmer from GitHub that provides code suggestions across editors.

An optimization algorithm that iteratively adjusts parameters to minimize a model's loss function.

A query language for APIs that allows clients to request exactly the data they need.

The accurate reference answer used to evaluate model predictions.

A high-performance RPC framework that uses protocol buffers for service definitions.

Controls that constrain model behavior to stay within acceptable boundaries.

When a model generates plausible-sounding but false or unsupported information.

A design where humans review or approve agent actions at key decision points.

Combining vector similarity with keyword or structured filtering for retrieval.

A configuration value set before training begins, such as learning rate or batch size.

Integrated Development Environment, a software application that provides tools for coding.

A model's ability to learn a task from examples embedded directly in the prompt.

The process of running a trained model on new input data to produce an output or prediction.

The process of loading, chunking, embedding, and storing documents for retrieval.

A token counted from the prompt sent to a model.

Fine-tuning a model on instruction-following examples to improve its ability to respond to user requests.

The study of understanding how models represent and process information internally.

A prompt crafted to bypass a model's safety guidelines or restrictions.

A model configuration that constrains output to valid JSON.

JSON Web Token, a compact format for securely transmitting claims between parties.

Retrieving documents based on exact or approximate word matches.

A structured repository of information an AI system can query or retrieve from.

A network of entities and relationships used to represent structured knowledge.

A cache of key and value tensors used to speed up autoregressive generation by avoiding redundant computation.

The correct output associated with a training example.

A neural network trained on vast text data to understand and generate human language.

The delay between a request and the start of a response.

A hyperparameter that controls how much model weights are updated during each training step.

A C++ implementation for running Llama models efficiently on consumer hardware.

Distributing requests across multiple servers to improve reliability and performance.

A model that runs on local hardware without requiring cloud API calls.

Recording events and messages from software for analysis.

The ability of a model to process very large context windows, often hundreds of thousands of tokens.

Low-Rank Adaptation, a parameter-efficient fine-tuning method that updates small adapter matrices instead of all weights.

A function that measures how far a model's predictions are from the correct answers during training.

A subset of AI where systems improve at tasks through experience and data without being explicitly programmed.

A technique that decomposes user-item interaction matrices into latent factors.

The maximum number of tokens a model is allowed to generate in a response.

MCP stands for Model Context Protocol. It is an open standard that lets AI clients connect to external tools, data sources, and prompts through a single, consistent interface. Anthropic introduced MCP in late 2024, and it has since been adopted by Claude Desktop, Cursor, Cline, VS Code, Windsurf, and a growing list of community clients. An MCP server is a small program that exposes three things: tools the model can call, resources the client can read, and prompts that help users accomplish common tasks. An MCP client discovers those capabilities and decides when to invoke them. Transport is usually stdio for local servers or Server-Sent Events for remote ones. For developers, MCP removes the need to build a custom integration for every API. You write one server, and any compatible client can use it. For users, it means AI assistants can securely access files, databases, SaaS tools, and web services without each client reinventing the wheel.

An application that connects to MCP servers and uses their capabilities.

A program that exposes tools, resources, and prompts via the Model Context Protocol.

A research area that reverse-engineers neural networks to understand their internal circuits.

A quantitative measure of model performance, such as accuracy, F1, or BLEU.

An architecture where only a subset of specialized sub-networks is activated per input, improving efficiency.

Practices for deploying and maintaining machine learning models in production.

An open standard that lets AI assistants connect to external data sources and tools through a common interface.

Filtering or flagging content that violates safety policies.

Tracking system health, performance, and behavior over time.

A system where multiple agents collaborate, compete, or delegate tasks to achieve complex goals.

Running multiple attention mechanisms in parallel to capture different kinds of relationships between tokens.

A model that can process and generate multiple types of input, such as text, images, and audio.

Identifying and classifying named entities such as people, organizations, and locations in text.

The field focused on enabling computers to understand, interpret, and generate human language.

Finding the closest data points to a query in a vector space.

A computational model inspired by biological neurons, organized in layers that process input data to produce outputs.

Scaling data to a standard range or distribution.

An authorization framework for delegated access to resources.

The ability to understand internal system state from external outputs.

A tool for running open-source models locally with simple commands.

Running a model locally on a user's device rather than on a remote server.

Performing a task after seeing a single example.

An open format for representing machine learning models, enabling cross-framework deployment.

A specification format for describing HTTP APIs.

Coordinating multiple tools, agents, or services to complete a workflow.

Extracting structured data from model outputs, often using schemas or regular expressions.

A token generated by the model in its response.

When a model learns training data too closely and performs poorly on unseen data.

Shared text between adjacent chunks to preserve context across chunk boundaries.

A configurable internal variable in a neural network that is learned during training and determines model behavior.

Methods that adapt a pre-trained model to new tasks while updating only a small fraction of parameters.

A measure of how well a probability model predicts a sample; lower is better.

A defined identity or character assigned to a model in a prompt.

Tailoring outputs or recommendations to individual users.

A linear sequence of data processing or model steps.

The process of deciding which actions to take and in what order to achieve a goal.

An add-on module that extends a system's capabilities.

Training a model on a large corpus to learn general language patterns before task-specific adaptation.

The proportion of predicted positives that are actually correct.

A model that learns to rank outputs based on which ones humans prefer.

Cleaning and transforming raw data before it is fed to a model.

The input text given to a language model to elicit a desired response.

Breaking a complex task into a sequence of prompts where each step uses the previous output.

Prompt engineering is the practice of crafting inputs to a language model so it produces better outputs without changing the model's weights. It covers word choice, structure, examples, constraints, and the order in which information appears. A well-engineered prompt can turn a mediocre response into a precise, actionable one. Effective prompts are usually clear, specific, and formatted. They state the task, define the audience, set the output format, and include any constraints. Adding examples, known as few-shot prompting, helps the model understand patterns that are hard to describe in words. Breaking complex tasks into steps, called chain-of-thought prompting, improves reasoning and arithmetic. Prompt engineering is iterative. You write a prompt, test it on diverse inputs, measure the results, and refine. Tools like the VePrompts Prompt Optimizer can surface issues such as ambiguity, missing constraints, or conflicting instructions. Good prompt engineering is often the fastest way to improve an AI feature before investing in fine-tuning or custom infrastructure.

An attack where malicious input overrides or leaks system instructions.

A reusable prompt pattern provided by an MCP server for common tasks.

A language-neutral binary serialization format developed by Google.

Processes designed to ensure products meet quality standards.

Reducing the precision of model weights to decrease memory usage and increase inference speed.

Transforming a user query to improve retrieval, such as expanding acronyms or adding synonyms.

A maximum allowance of usage for an account or key.

RAG stands for Retrieval-Augmented Generation. It is a pattern that gives a language model access to information outside its training data by fetching relevant documents at query time and including them in the prompt. Instead of memorizing facts, the model reasons over retrieved snippets, which makes answers more accurate, current, and traceable. A typical RAG pipeline has four stages. First, documents are split into chunks and converted into embeddings using an embedding model. Second, those embeddings are stored in a vector database. Third, when a user asks a question, the system embeds the query and searches the database for the closest chunks. Finally, the retrieved chunks are added to the prompt as context, and the model generates an answer grounded in that evidence. RAG is especially useful when answers depend on private data, such as internal wikis, support tickets, or product documentation. It also reduces hallucination because the model can cite the retrieved text. Teams often tune RAG by changing chunk size, overlap, reranking algorithms, and query rewriting strategies.

A cap on the number of requests or tokens allowed in a time window.

A second-stage model that scores and reorders retrieved documents for better relevance.

A pattern where an agent Reasons and Acts in alternating steps to solve tasks.

A model optimized for step-by-step logical reasoning and complex problem solving.

The proportion of actual positives that were correctly identified.

Suggesting items or actions to users based on data.

Attempting to find vulnerabilities, biases, or harmful behaviors in a model.

An agent evaluating its own output and revising it based on critique.

Techniques used to reduce overfitting by discouraging overly complex models.

Learning by interacting with an environment and receiving rewards or penalties.

Training a model using human preference signals to make outputs more helpful and harmless.

A connection between two entities in a knowledge graph.

Read-only data exposed by an MCP server that a client can pull into context.

Representational State Transfer, an architectural style for designing networked APIs.

Generating responses grounded in retrieved external documents to improve accuracy and recency.

A model trained to score outputs according to human preferences, used in RLHF.

Short for Reinforcement Learning from Human Feedback.

A model's ability to maintain performance under noisy or adversarial inputs.

Asking the model to assume a specific role or persona to shape its responses.

A set of metrics for evaluating automatic summarization by comparing overlap with reference summaries.

Practices that reduce harmful, unethical, or dangerous model outputs.

An isolated execution environment that limits what code can access.

Adjusting compute resources to handle varying workloads.

Attention applied within a single sequence, allowing each token to relate to every other token.

An agent identifying and fixing its own mistakes.

Searching by meaning rather than exact keyword matches, often using embeddings.

Determining the emotional tone or opinion expressed in text.

A model architecture that maps an input sequence to an output sequence.

Finding items with embeddings close to a query embedding, usually by cosine or Euclidean distance.

A specific capability advertised by an agent or service.

A model where most parameters are inactive for any given input, reducing compute per forward pass.

Server-Sent Events, an HTTP-based transport for streaming messages from server to client.

Transforming data to have zero mean and unit variance.

A transport that uses standard input and output for local MCP server communication.

A string that signals the model to stop generating further tokens.

Requiring the model to produce output conforming to a defined schema.

Producing a shorter version of a longer text while preserving key information.

Training a model on labeled input-output pairs.

Data generated by models or simulations rather than collected from real-world sources.

A system prompt is the high-level instruction that sets the model's role, tone, constraints, and behavior for a conversation. It is sent once at the start of the context and influences every response that follows. While users see the assistant's reply, they usually do not see the system prompt unless the application exposes it. A good system prompt is specific and scoped. Instead of saying you are helpful, it might say you are a senior React reviewer who gives concise feedback in bullet points, flags security issues, and never writes full code replacements. This reduces ambiguity and makes the model's output more consistent across sessions. System prompts are also the first line of defense for safety and product requirements. You can use them to enforce output formats, reject off-topic requests, require citations, or ask the model to disclose uncertainty. Because they carry so much influence, small changes to a system prompt often produce larger improvements than adding more examples to user messages.

Temperature is a sampling parameter that controls how random a language model's outputs are. It scales the logits, or raw scores, that the model assigns to each possible next token before a token is chosen. A lower temperature makes the model more conservative and deterministic; a higher temperature makes it more creative and varied. At temperature zero, the model almost always picks the highest-scoring token, which is ideal for tasks like code generation, factual answers, and structured output where consistency matters. At temperature one or above, the model is more willing to sample lower-scoring tokens, which can produce surprising phrasing, creative writing, and diverse brainstorming ideas. There is no universal best setting. Coding and data extraction usually benefit from low temperatures around 0.1 to 0.3. Marketing copy, fiction, and idea generation often feel better at 0.7 to 1.0. If outputs are too repetitive, raise the temperature. If they become erratic or off-topic, lower it.

Spending more computation during inference to improve output quality, such as through reasoning or search.

Evaluating software to ensure it behaves as expected.

Slowing or limiting requests to enforce rate limits.

The number of tokens or requests processed per unit of time.

A token is the basic unit a language model reads and writes. It can be a whole word, part of a word, or even a single punctuation mark. Models do not see raw text; they see a sequence of token IDs produced by a tokenizer. Different models use different tokenizers, so the same sentence may cost a different number of tokens on GPT-4o than on Claude or Gemini. As a rough guide, one token is about four English characters or three quarters of a word. A 500-word article might be 700 to 900 tokens, depending on the tokenizer. Tokens matter for two reasons. First, pricing is per token, so longer prompts and outputs cost more. Second, models have a context window measured in tokens; if your input exceeds that limit, the model cannot process it. Tools like the VePrompts tokenizer show you exactly how a specific model splits your text.

Splitting text into tokens for model processing.

A tool that converts text into tokens for model processing.

The rate at which a model generates tokens after the first one.

An external function an agent can call to perform an action or retrieve data.

The ability of a model to invoke external tools or APIs to complete tasks.

A sampling parameter that limits token selection to the k highest-probability candidates.

Nucleus sampling parameter that limits token selection to the smallest set whose cumulative probability exceeds p.

The presence of harmful, offensive, or abusive content in model outputs.

Recording the path of a request through a system to diagnose issues.

Leveraging knowledge learned from one task to improve performance on a related task.

A neural network architecture using self-attention to process sequences in parallel, forming the basis of modern LLMs.

Converting text from one language to another.

The communication channel between an MCP client and server, such as stdio or SSE.

A subject-predicate-object statement representing a fact in a knowledge graph.

Confidence that a model will behave reliably, safely, and as intended.

When a model is too simple to capture patterns in the training data, leading to poor performance overall.

Training a model on data without explicit labels to find hidden patterns.

A database optimized for storing and searching high-dimensional embeddings.

A model that understands both images and text, often enabling image captioning or visual question answering.

The set of tokens a model recognizes.

A popular source-code editor developed by Microsoft.

An HTTP callback that triggers an action when an event occurs.

A numerical value in a neural network that scales the strength of connections between neurons.

A predefined sequence of steps that may include models, tools, and conditional logic.

Performing a task without any task-specific training examples.