AI Glossary
In plain words.Every technical term in skill descriptions, written as a glossary entry you can finish
Each entry stands alone and covers one concept: what it is, why it exists, and where you will run into it while using AI tools. read on and you fill in a whole area; or filter by field and level first and pick the few you can finish today.
Some fields on this page have no English version yet — they are shown in the original Chinese.
嵌入(Embedding)
把一段文字换算成一串数字,让电脑能比较哪两句话意思更接近。搜索和知识库能在换了说法时也找对答案,靠的就是它。
MCP
Model Context Protocol
A standard interface that lets AI connect to outside tools and data in one uniform way. Build a tool to it once and any AI assistant can call it, saving you from writing a separate integration for each tool.
上下文窗口(Context Window)
AI 在一次对话里能同时读进去的文字总量。这个量有上限,聊得越久,开头的内容越容易被挤出去。
多智能体(Multi-Agent)
让多个 AI 分角色合作完成一件事:有的查资料,有的写,有的专门挑错。适合把一个大任务拆给一支小队,而不是一个 AI 从头扛到尾。
RAG(检索增强生成)
Retrieval-Augmented Generation
让模型在回答前先翻一遍你的资料,搜出相关段落再照着这些内容作答。资料可以随时替换,不必重新训练模型,也是压制它胡编的常用手段。
Token(词元)
模型读写文字的最小单位。它不按字算,而是把文字切成小块来处理,一次能读多少、要收多少钱,都拿这个数量来量。
推理(Inference)
模型训练完成之后,每次回答你的那个计算过程就叫推理。决定响应快不快、运行贵不贵的,是它,而不是训练。
本地部署(Self-hosted)
把模型或软件装在自己的电脑、公司服务器上运行,输入和输出都不离开这台机器。换来的是隐私和可控,代价是机器和维护都得自己扛。
LangChain(AI 应用开发框架)
一个开源的开发框架,帮程序员把大模型和文档、数据库、外部工具接在一起。它提供现成的零件,省掉大量重复的对接工作。
Function Calling
Letting a model do more than answer: it can ask an outside program to act — check live weather, do exact arithmetic, send email, read a file. It's the step that turns an AI from something that only talks into something that gets things done.
微调(Fine-tuning)
拿你自己的数据在已有模型上再训练一轮,让它更贴你的行业说法和表达习惯。要它记住资料该用知识库,要改它的行为习惯才用它。
幻觉(Hallucination)
模型一本正经地说出并不存在的事实——虚构的书名、打不开的链接、记错的日期。它语气越笃定,越容易让人信以为真。
AGENTS.md(项目约定文件)
放在项目里写给 AI 看的一份说明,交代这个项目的技术栈、代码规范和注意事项。AI 一进项目读到它,就知道该按什么方式干活。
Dify(AI 应用搭建平台)
一个开源的 AI 应用搭建平台,用拖拽的方式把模型、知识库和流程拼成问答机器人或客服助手,不需要从零写代码。
Every concept here comes from real skill descriptions — to see which skills they show up in, browse Skill indexby category.
Terms all make sense now — what next?
Head back to the index and pick a few skills to install in your AI client — most take a single command.