Vibulary
Menu
Appearance and language
Terms by topic · 63 guides

AI

Learn how AI gets instructions, finds facts, uses tools, gives answers, and gets checked.

01 / Atlas

All AI terms

Open a term for a simple meaning, an example, when to use it, how official it is, and its code name.

  1. 01 · More detail

    Agent & Tool Calling — Learning Bundle

    This learning bundle separates two AI system ideas. An agent works toward a goal across steps. Tool calling lets a model ask for a limited outside action. The bundle is not one standard agent protocol.

  2. 02 · More detail

    Agent Loop

    An Agent Loop is a controlled cycle. The agent checks state, acts with an allowed tool, reads the result, then stops or continues.

  3. 03 · More detail

    Agent Memory

    Agent memory is saved information that an AI agent can use after the current message. Clear rules say what is saved, for how long, and how it can change.

  4. 04 · More detail

    Agent Orchestration

    Agent orchestration is the plan that coordinates AI agents, tools, handoffs, shared information, and stopping rules. It keeps a many-step job under control.

  5. 05 · More detail

    Agent Sandbox

    An agent sandbox is a limited place where an AI agent can run code or tools. It separates the task from the rest of the computer, but protection depends on its design.

  6. 06 · More detail

    Agent Skill

    An Agent Skill teaches one repeatable kind of work. It contains focused rules and may include scripts or resources.

  7. 07 · More detail

    Agent2Agent (A2A) Protocol

    Agent2Agent (A2A) is an open, versioned protocol for AI agents to discover skills, exchange messages, and manage tasks. Each implementation must name a supported protocol version.

  9. 09 · More detail

    AI Benchmark

    An AI benchmark is a fixed and written-down set of tasks, data, measures, and test conditions. It lets people compare systems using the same rules.

  10. 10 · Core

    AI Evaluation

    AI evaluation is a repeatable way to test an AI system. It uses clear tasks, test data, scoring rules, and safety limits.

  11. 11 · More detail

    AI Grounding

    AI grounding links generated claims or actions to evidence, found sources, or trusted system data. Methods and scores differ. There is no single grounding protocol.

  12. 12 · More detail

    AI Guardrail

    AI guardrail is a broad industry name for safety controls. These controls can find, limit, block, change, or flag unsafe AI inputs, outputs, or actions. It is not one standard tool.

  13. 13 · Core

    AI Hallucination

    AI hallucination is a common name for an answer with unsupported or false claims. NIST uses the more exact risk term confabulation.

  14. 14 · More detail

    AI Tracing

    AI tracing records the linked steps in one AI run, such as model calls, search, tool use, and agent work. The record helps people follow the run from start to finish.

  15. 15 · More detail

    Computer Use

    Computer use lets a model view a screen and return actions such as clicking, typing, or scrolling. An application harness checks and performs each allowed action.

  16. 16 · More detail

    Context Compaction

    Context compaction replaces older conversation state with a smaller item so a long task can continue. The item may not be a readable summary, and details may be lost.

  17. 17 · More detail

    Context Engineering

    Context Engineering selects and orders what a model receives. This can include facts, instructions, tools, state, and evidence.

  18. 18 · Core

    Context Window

    A context window is the number of tokens one model request can hold for both input and generated output. The exact counting rules depend on the model and API.

  19. 19 · Core

    Conversation History

    Conversation History is earlier chat and useful state passed into later turns. It is often filtered or summarized.

  20. 20 · Core

    Document Chunking

    Document chunking splits content into small parts that a search system can find. Clear rules set each part's size, overlap, and stopping point before indexing.

  21. 21 · Core

    Embedding

    An embedding is a list of numbers a model learns for an item. A distance or similarity rule can then compare patterns between items in that number space.

  22. 22 · More detail

    Embedding & Vector Database — Learning Bundle

    This learning bundle joins two retrieval ideas. An embedding turns content into number vectors. A vector database stores, indexes, filters, and searches those vectors. The bundle is not one standard datastore type.

  23. 23 · More detail

    Fine-tuning

    Fine-tuning keeps training a ready-made model with carefully chosen examples from one task or field. This changes the model's learned weights and behavior.

  24. 24 · More detail

    Groundedness

    Groundedness asks how well the claims in an answer are backed by the sources given to the model. It is one test area, not one worldwide standard score.

  25. 25 · More detail

    Guardrail, Human-in-the-Loop & Eval — Learning Bundle

    This learning bundle joins three different control methods. Guardrails limit behavior or send it for more review. Humans check high-impact cases. Evals measure quality on realistic tasks. The bundle is not one standard safety method.

  26. 26 · Core

    Hallucination & Grounding — Learning Bundle

    This learning bundle compares a failure with a way to reduce it. Hallucination is model output with no support. Grounding ties claims to evidence that's given, found, or checked. The bundle is not one standard quality metric.

  27. 27 · More detail

    Harness Engineering

    Harness Engineering builds the setup around coding agents. It includes the repository, instructions, tools, safety, environment, and feedback.

  28. 28 · More detail

    Human in the Loop

    Human in the loop is a planned review step. A named person gets the facts, power, and duty to act. They can change, approve, reject, or stop AI-assisted work.

  29. 29 · More detail

    Hybrid Search

    Hybrid search uses both exact-word search and meaning-based vector search for one question. It then joins or reranks both sets of results.

  30. 30 · More detail

    Knowledge Distillation

    Knowledge distillation trains a small student model to copy useful signals from a larger teacher model. Training often uses both teacher scores and known correct labels.

  31. 31 · Core

    Language Model

    A language model is a statistical model that predicts and creates sequences. It can often follow instructions for text and other kinds of data, called modalities.

  32. 32 · More detail

    Model Context Protocol

    Model Context Protocol connects AI applications to outside tools and data. Hosts use clients to talk with servers offering tools, resources, and prompts. It is open, and both sides first agree on available capabilities.

  33. 33 · More detail

    Model Grader

    A model grader uses an AI model to score or label an output with a written rule. It can be wrong, so people must check it against known examples.

  34. 34 · Core

    Model Inference

    Inference is when a trained model uses new input to make a prediction, give a score, or create an answer. Its learned weights do not change.

  35. 35 · More detail

    Model Quantization

    Model quantization stores weights or activations with less exact number formats. This can lower memory use, data movement, and computing cost.

  36. 36 · More detail

    Model Routing

    Model routing chooses which model or model path handles each request. The choice follows clear rules for quality, skill, cost, delay, or safety policy.

  37. 37 · Core

    Model Token

    A model token is a unit from one specific tokenizer. The tokenizer turns text or bytes into number IDs for a model. A token is not always one word or character. Its split and count depend on the tokenizer version.

  38. 38 · More detail

    Multimodal AI

    Multimodal AI can use or create several data types. These may include text, images, audio, video, or sensor data.

  39. 39 · More detail

    OpenAI Reasoning Effort

    reasoning.effort is an OpenAI API setting for how much model work a request should use before answering. Supported levels and defaults vary by model, and higher effort does not promise a correct answer.

  40. 40 · More detail

    Output Tokens per Second

    Output tokens per second measures how fast a model creates output tokens after decoding starts. The report must name the tokenizer, time range, grouping method, and number of requests running together.

  41. 41 · Core

    Prompt

    A prompt is everything given to a model for one task. It can include instructions, context, examples, and data.

  42. 42 · More detail

    Prompt Caching

    Prompt caching lets a model service reuse work for the same beginning of an input. It may reduce repeat processing, but it does not reuse the old answer.

  43. 43 · More detail

    Prompt Injection

    Prompt injection is harmful or unexpected text. It tries to make an AI ignore the app's trusted rules and follow other instructions.

  44. 44 · More detail

    Prompt Versioning

    Prompt versioning keeps named, fixed revisions of a production prompt. Teams can review, test, release, and restore an exact revision.

  45. 45 · More detail

    ReAct

    ReAct means Reason and Act. An agent acts, reads the result, and uses that fact for its next step. It is not React.

  46. 46 · More detail

    Reranking

    Reranking uses a second model to score search results again. It compares the question and each found document more closely, then changes their order.

  47. 47 · Core

    Retrieval Relevance

    Retrieval relevance tells how well a search result or answer meets the user's stated need under clear scoring rules.

  48. 48 · More detail

    Retrieval-Augmented Generation

    Retrieval-Augmented Generation, or RAG, finds outside evidence at run time. It gives that evidence to a model before writing.

  49. 49 · More detail

    Sampling

    Sampling means choosing the next text piece (token) from a list of possible choices and their chances. The model does not always pick the most likely one.

  50. 50 · More detail

    Stateless Request

    A Stateless Request uses only data sent in that call. It does not assume the model remembers earlier calls.

  51. 51 · More detail

    Streaming Response

    A Streaming Response arrives in small events or chunks. Useful output can appear before the job finishes.

  52. 52 · Core

    Structured Output

    Structured output makes a model follow a machine-readable schema instead of writing any form of prose.

  53. 53 · More detail

    Sub-agent

    A Sub-agent is a separate agent with one clear, limited task. It has its own context and a result to return.

  54. 54 · More detail

    System Prompt

    A system prompt is a set of high-priority instructions. It sets an AI product’s role, limits, tools, and rules for replies.

  55. 55 · Core

    Temperature

    Temperature is a setting used before sampling. It changes whether token chances are packed around a few choices or spread across more choices.

  56. 56 · More detail

    Time to First Token

    Time to first token (TTFT) measures an early wait. It starts at a clearly named request point. It ends when the client or server sees the first generated token. Always state both points.

  58. 58 · Core

    Token & Context Window — Learning Bundle

    This learning bundle joins two different model runtime ideas. A token is a model-specific piece of input or output. A context window limits the sequence available for one generation. The bundle is not one standard metric.

  59. 59 · More detail

    Tool Calling

    Tool calling lets a model request a named outside tool in a set format. The request includes arguments. The host app decides whether and how to run it.

  60. 60 · More detail

    Top-p Sampling

    Top-p sampling sorts possible tokens by chance, then keeps the smallest group whose total chance reaches p. It picks the next token from that group.

  61. 61 · Core

    Transformer

    A Transformer is a kind of neural network. It uses attention to learn how parts of a sequence relate and can work on many positions at once.

  62. 62 · More detail

    Vector Database

    A vector database is an industry name for a data store that saves vectors and searches for similar ones. It often filters other details and uses fast approximate indexes, but there is no single universal standard.

  63. 63 · Core

    Vibe Coding

    Vibe Coding builds software through conversation. A person sets and checks the goal while an AI agent writes much of the code.