提示词Prompt
你交给模型的话、素材和要求。它是起点，不是咒语；清楚的意图比“技巧”更重要。
What you hand the model: words, material, requests. A starting point, not a spell; clear intent beats tricks.
再深一点Deeper
好的提示通常说清：你要什么、给谁、什么感觉、哪些不要。
Good prompts usually say what you want, for whom, in what feeling, and what to avoid.
上下文Context
模型在这一次回应时“摆在眼前”的全部内容：你的话、之前的对话、附上的文件与检索到的片段。容量有限。
Everything in front of the model for this one reply: your words, earlier turns, attached files and retrieved passages. Its room is limited.
再深一点Deeper
上下文不保证永久保存；开一个新对话，可能就没有了。
Context is not guaranteed to persist; in a new conversation it may be gone.
大语言模型（LLM）LLM
用大量文本训练出来的模型，擅长生成与改写语言。它根据上下文估计下一个词元的概率分布，再从中选出一个，而不是去“查”一个答案库。
A model trained on large amounts of text, good at generating and rewriting language. It estimates a probability distribution over the next token and picks from it, rather than looking an answer up in a database.
再深一点Deeper
许多产品把模型与搜索、工具、记忆组合在一起，所以体验不只来自模型本身。
Many products combine a model with search, tools and memory, so the experience is not only the model.
多模态Multimodal
能处理不止文字的东西，如图像、声音。“看懂一张图”与“画出一张图”是两种不同的能力。
Handles more than text, such as images or sound. "Understanding an image" and "making an image" are different abilities.
再深一点Deeper
是否支持某种输入或输出，取决于具体的模型与产品。
Whether an input or output is supported depends on the specific model and product.
推理Reasoning
拆解步骤、比较方案、检查矛盾的能力表现。产品可能只显示答案或摘要，而不展示完整的内部思路。
The capability of splitting steps, comparing options and checking contradictions. A product may show only the answer or a summary, not the full internal thinking.
再深一点Deeper
判断好坏，看假设是否说清、结论能否验证。
Judge by whether assumptions are stated and conclusions can be verified.
幻觉Hallucination
模型生成了读起来合理、实际上没有依据或是错的内容。流畅不是证据。
The model produces content that reads plausibly but is unfounded or wrong. Fluency is not evidence.
再深一点Deeper
缺少来源时更常见。附上来源、让它标出不确定处，并自己核对。
More common without sources. Attach sources, ask it to mark uncertainty, and check yourself.
API
程序与服务之间约定好的接口，让一个应用请求另一个服务做事。是否需要密钥或计费，因服务而异。
An agreed interface between programs and services, letting one app ask another service to do something. Whether it needs a key or is billed depends on the service.
再深一点Deeper
本页没有调用任何 API；所有演示都是在你的浏览器里本地运行的固定规则。
This page calls no API; every demo is a fixed rule running locally in your browser.
OAuth
一种授权方式：不交出密码，也能允许某个应用在限定范围内访问你的账户数据。
A way of authorizing: let an app access some of your account data within a limited scope without handing over your password.
再深一点Deeper
你应能看到授权的范围，并能随时撤销。
You should be able to see the scope granted and revoke it at any time.
MCP
Model Context Protocol（模型上下文协议）：让 AI 应用以统一方式连接外部工具与数据的开放协议。
Model Context Protocol: an open protocol that lets AI apps connect to external tools and data in a common way.
再深一点Deeper
MCP 是连接方式，不是模型；能否使用、如何授权取决于具体平台，需要查证。
MCP is a way of connecting, not a model; whether and how you can use it depends on the platform, so check.
代理Agent
围绕目标反复循环的系统：规划、调用工具、查看结果、调整，直到停止。需要边界与人的复核。
A system that loops around a goal: plan, call tools, read results, adjust, until it stops. It needs boundaries and human review.
再深一点Deeper
代理会放大错误，所以权限、预算与停止条件要由人设定。
Agents can amplify mistakes, so permissions, budgets and stop conditions should be set by people.
检索 / RAGRetrieval / RAG
先从指定资料里找出相关片段，再放进上下文，让模型据此回答并给出引用。
First find relevant passages in chosen sources, place them into context, and let the model answer from them with citations.
再深一点Deeper
效果取决于资料质量与切分方式；回答仍要核对引用。
Results depend on source quality and how it is split up; still check the citations.
迭代Iteration
看结果、调整想法、再试一次的循环。每一轮里“保留什么”是你的判断。
The loop of looking at results, adjusting the idea and trying again. In each round, "what to keep" is your judgment.
再深一点Deeper
每轮只改一件事，更容易看清是什么带来了变化。
Change one thing per round, and it is easier to see what made the difference.
词元Tokens
模型处理文本时的小片段。它一次生成一个词元。
The small pieces of text a model works with. It generates one token at a time.
再深一点Deeper
词元是处理文本的单位，与下面的“额度”是两回事。
Tokens are a unit of processing text — a different thing from the "credits" below.
词元 vs 额度Tokens vs credits
词元是模型处理文本的片段；额度（credits）是某个服务用来计量使用量的单位。二者不是同一回事。
Tokens are the pieces a model processes; credits are a unit a given service uses to meter usage. They are not the same.
再深一点Deeper
具体怎样计量因服务而异，请以官方说明为准；本页不列任何价格。
How usage is metered varies by service; follow official statements. This page lists no prices.
Uploading is not model training in itself, and context is not guaranteed permanent memory; products may index and retain files differently, so check their documentation.