MCPA shared way for AI applications to connect tools
The Model Context Protocol is an open protocol for communication between AI applications and external tools or information sources.
Read explanation ↗AI dictionary
Unfamiliar AI term? Look it up and understand it through an everyday example.
Search by English term, abbreviation or Chinese name.
10 terms
SkillA reusable task instruction package
Instructions and resources an assistant can load for a task. The Agent Skills format uses a SKILL.md file and can include scripts and templates.
Read explanation ↗AgentAn assistant that chooses steps toward a goal
An AI system that selects actions and uses available tools toward a goal, often checking results before deciding what to do next.
Read explanation ↗Prompt
The request and relevant material you give an AI. A prompt can include instructions, context, examples, images and an expected output format.
Read explanation ↗TokenA unit a model processes
Models split content into small units called tokens. A text token might represent a character, part of a word or another fragment.
Read explanation ↗Context windowHow much context fits at once
The amount of information a model can handle in one processing context, usually measured in tokens. Think of a work surface with limited space.
Read explanation ↗AIArtificial intelligence
A broad family of technologies for tasks such as recognizing images, processing language and making predictions. A chat assistant is one application of AI.
Read explanation ↗LLMLarge language model
A model trained on large amounts of material to learn language patterns and process or generate text. It supplies language capabilities behind many assistants.
Read explanation ↗HallucinationPlausible output without a sound basis
An AI produces false, unsupported or source-inconsistent content in fluent language. Confident wording can still be wrong.
Read explanation ↗MultimodalWorking with more than one content format
Processing formats such as text, images, audio or video. Supported inputs and outputs depend on the model and product.
Read explanation ↗APIAn interface for software to work together
An application programming interface specifies how one program can request information or actions from another.
Read explanation ↗API keyA credential for calling a service
A credential issued to an application or project to identify requests, sometimes also controlling access and billing.
Read explanation ↗RAGRetrieve material, then write an answer
Retrieval-augmented generation finds relevant material and supplies it to a model before generating an answer, like consulting the right pages in an open-book task.
Read explanation ↗EmbeddingRepresenting content with comparable numbers
An embedding represents features of text or other content as numbers so software can compare related material.
Read explanation ↗Fine-tuningFurther training for a particular task
Continue training an existing model on task-specific data, updating some or all parameters to adapt its behavior.
Read explanation ↗InferenceRunning a trained model to produce output
Inference is the process of applying a trained model to new input and computing a result.
Read explanation ↗ReasoningWorking through a problem in more steps
In AI products, reasoning usually describes multi-step analysis, such as breaking down a problem and checking constraints. Some models use extra computation before answering.
Read explanation ↗WorkflowA sequence of connected work steps
A workflow passes work through steps such as input, processing, checking and output. It can combine AI, ordinary software and human review.
Read explanation ↗Tool callingUsing external tools to get work done
A model requests a tool and supplies arguments; the application executes the request and returns the result for the model to use.
Read explanation ↗CLIAn interface controlled with typed commands
A command-line interface lets you enter commands in a terminal and receive results. Some AI development tools provide this entry point.
Read explanation ↗IDEA workspace that combines development tools
An integrated development environment brings tools for editing, running and debugging code into one workspace.
Read explanation ↗Open weightsMaking trained model parameters available
A publisher provides the learned parameters that influence model outputs. Compatible files and software can allow self-hosting or research under the license terms.
Read explanation ↗Local modelA model running on your own device
The model files and runtime execute on your computer or device. Available memory and processing hardware affect model size and speed.
Read explanation ↗TemperatureA setting that affects sampling randomness
For models that support it, temperature changes how the next output is sampled. Lower settings generally concentrate choices; higher settings allow more variation.
Read explanation ↗Rate limitHow much a service allows in a time period
A service caps requests or token usage over time. Limits can differ by account, project, model and service tier.
Read explanation ↗BenchmarkA task set and scoring method for comparison
A benchmark compares performance on defined tasks under specified conditions. Like an exam, it provides evidence about a limited slice of work.
Read explanation ↗AIGCAI-generated content
Text, images or code created by AI in response to instructions. AIGC describes the output, rather than the name of a particular app.
Read explanation ↗Foundation model
A model trained broadly enough to be adapted to many tasks. Different chat and coding products can build on the same underlying model.
Read explanation ↗No matching term yet
Try a full name, abbreviation or Chinese name, or clear the filters to browse every term.
Show all termsExplanations draw on public documentation. Examples illustrate concepts; check your product’s documentation for its actual behavior.
New entries reviewed on 2026-10-15. Existing entries retain their own review dates.