--- title: "AI 核心概念 - AI·外脑 | 思行致远" canonical: "https://siyuanailab.com/knowledge/ai-core/" dateModified: "2026-09-14T19:22:38+08:00" --- 打开 AI·外脑目录 AI·外脑 [返回专栏首页](https://siyuanailab.com/knowledge/) × 站内搜索 ⌕ 主要栏目 - [AI 核心概念](https://siyuanailab.com/knowledge/ai-core/) - [工作流](https://siyuanailab.com/knowledge/workflows/) - [工具配置与选型](https://siyuanailab.com/knowledge/toolkits/) - [信源与标准](https://siyuanailab.com/knowledge/sources/) - [场景案例 待开发](https://siyuanailab.com/knowledge/scenarios/) - [职业与能力](https://siyuanailab.com/knowledge/career/) 专题 - [从这里开始](https://siyuanailab.com/knowledge/start/) - [获取与安装](https://siyuanailab.com/knowledge/install/) - [建设日志](https://siyuanailab.com/knowledge/ops/) [AI 核心概念子目录](https://siyuanailab.com/knowledge/ai-core/) [基础概念 **9**](https://siyuanailab.com/knowledge/ai-core/foundations/) - [01-01 人工智能](https://siyuanailab.com/knowledge/ai-core/foundations/artificial-intelligence/) - [01-02 生成式 AI](https://siyuanailab.com/knowledge/ai-core/foundations/generative-ai/) - [01-03 大语言模型](https://siyuanailab.com/knowledge/ai-core/foundations/large-language-model/) - [01-05 机器学习](https://siyuanailab.com/knowledge/ai-core/foundations/machine-learning/) - [01-06 神经网络](https://siyuanailab.com/knowledge/ai-core/foundations/neural-network/) - [01-07 Transformer 架构](https://siyuanailab.com/knowledge/ai-core/foundations/transformer-architecture/) - [01-08 参数](https://siyuanailab.com/knowledge/ai-core/foundations/parameters/) - [01-09 训练数据](https://siyuanailab.com/knowledge/ai-core/foundations/training-data/) - [01-10 预训练](https://siyuanailab.com/knowledge/ai-core/foundations/pre-training/) [模型与机制 **12**](https://siyuanailab.com/knowledge/ai-core/models/) - [02-01 Token](https://siyuanailab.com/knowledge/ai-core/models/token/) - [02-02 上下文窗口](https://siyuanailab.com/knowledge/ai-core/models/context-window/) - [02-03 注意力机制](https://siyuanailab.com/knowledge/ai-core/models/attention-mechanism/) - [02-04 嵌入/向量](https://siyuanailab.com/knowledge/ai-core/models/embedding/) - [02-05 检索增强生成](https://siyuanailab.com/knowledge/ai-core/models/retrieval-augmented-generation/) - [02-06 微调](https://siyuanailab.com/knowledge/ai-core/models/fine-tuning/) - [02-07 提示词工程](https://siyuanailab.com/knowledge/ai-core/models/prompt-engineering/) - [02-09 思维链](https://siyuanailab.com/knowledge/ai-core/models/chain-of-thought/) - [02-10 推理](https://siyuanailab.com/knowledge/ai-core/models/reasoning/) - [02-11 幻觉](https://siyuanailab.com/knowledge/ai-core/models/hallucination/) - [02-12 对齐](https://siyuanailab.com/knowledge/ai-core/models/alignment/) - [02-13 温度/采样参数](https://siyuanailab.com/knowledge/ai-core/models/temperature/) [Agent 与智能体系统 **8**](https://siyuanailab.com/knowledge/ai-core/agents/) - [03-01 智能体](https://siyuanailab.com/knowledge/ai-core/agents/ai-agent/) - [03-02 多智能体系统](https://siyuanailab.com/knowledge/ai-core/agents/multi-agent-system/) - [03-04 工具调用](https://siyuanailab.com/knowledge/ai-core/agents/tool-use-function-calling/) - [03-05 记忆系统](https://siyuanailab.com/knowledge/ai-core/agents/memory-system/) - [03-06 规划能力](https://siyuanailab.com/knowledge/ai-core/agents/planning/) - [03-07 委托机制](https://siyuanailab.com/knowledge/ai-core/agents/delegation/) - [03-08 模型上下文协议](https://siyuanailab.com/knowledge/ai-core/agents/model-context-protocol/) - [03-09 智能体工作流](https://siyuanailab.com/knowledge/ai-core/agents/agentic-workflow/) [工程与架构 **7**](https://siyuanailab.com/knowledge/ai-core/engineering/) - [04-01 AI 工程学](https://siyuanailab.com/knowledge/ai-core/engineering/ai-engineering/) - [04-02 LLM 网关](https://siyuanailab.com/knowledge/ai-core/engineering/llm-gateway/) - [04-04 测试时计算](https://siyuanailab.com/knowledge/ai-core/engineering/test-time-compute/) - [04-06 上下文工程](https://siyuanailab.com/knowledge/ai-core/engineering/context-engineering/) - [04-07 Vibe Coding](https://siyuanailab.com/knowledge/ai-core/engineering/vibe-coding/) - [04-08 人参与循环](https://siyuanailab.com/knowledge/ai-core/engineering/human-in-the-loop/) - [04-09 评估方法](https://siyuanailab.com/knowledge/ai-core/engineering/evaluation-methods/) [交互与提示 **4**](https://siyuanailab.com/knowledge/ai-core/interaction/) - [05-01 系统提示词](https://siyuanailab.com/knowledge/ai-core/interaction/system-prompt/) - [05-03 少样本提示](https://siyuanailab.com/knowledge/ai-core/interaction/few-shot-prompting/) - [05-04 角色提示](https://siyuanailab.com/knowledge/ai-core/interaction/role-prompting/) - [05-05 提示词注入](https://siyuanailab.com/knowledge/ai-core/interaction/prompt-injection/) [安全与治理 **4**](https://siyuanailab.com/knowledge/ai-core/safety/) - [06-01 AI 安全](https://siyuanailab.com/knowledge/ai-core/safety/ai-safety/) - [06-03 智能体治理](https://siyuanailab.com/knowledge/ai-core/safety/agent-governance/) - [06-04 模型偏见](https://siyuanailab.com/knowledge/ai-core/safety/model-bias/) - [06-05 数据隐私](https://siyuanailab.com/knowledge/ai-core/safety/data-privacy/) AI 核心概念 # 先把 44 个底层词讲清楚 这组母目录把 AI 的常用概念拆成 6 个子分类。它不是技术百科，而是给非技术创业者和 AI 助理共同调用的判断词表。 共 44 张词条，分为 6 个子目录 ## [基础概念](https://siyuanailab.com/knowledge/ai-core/foundations/) AI 领域最底层的通用概念 9 张 基础概念词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 01-01 | [人工智能](https://siyuanailab.com/knowledge/ai-core/foundations/artificial-intelligence/) | Artificial Intelligence | 让机器完成需要人类智能参与的任务，比如识别、判断、生成、规划和行动。 | AI人工智能基础概念 | | 01-02 | [生成式 AI](https://siyuanailab.com/knowledge/ai-core/foundations/generative-ai/) | Generative AI | 能根据提示生成文本、图片、音频、视频或代码的新内容的 AI。 | 生成式AI内容生成AI基础 | | 01-03 | [大语言模型](https://siyuanailab.com/knowledge/ai-core/foundations/large-language-model/) | Large Language Model | 在海量文本和代码上训练出来、擅长理解和生成语言的大型神经网络。 | LLM大语言模型模型基础 | | 01-05 | [机器学习](https://siyuanailab.com/knowledge/ai-core/foundations/machine-learning/) | Machine Learning | 让计算机从数据中学习规律，而不是只按人工写死的规则执行。 | 机器学习ML基础概念 | | 01-06 | [神经网络](https://siyuanailab.com/knowledge/ai-core/foundations/neural-network/) | Neural Network | 由多层计算单元组成的模型，用来从数据中学习复杂模式。 | 神经网络深度学习模型基础 | | 01-07 | [Transformer 架构](https://siyuanailab.com/knowledge/ai-core/foundations/transformer-architecture/) | Transformer Architecture | 一种基于注意力机制的神经网络架构，是现代大语言模型的核心底座。 | Transformer注意力机制LLM | | 01-08 | [参数](https://siyuanailab.com/knowledge/ai-core/foundations/parameters/) | Parameters | 模型在训练中学到的内部数值，决定它如何把输入映射成输出。 | 参数模型规模模型机制 | | 01-09 | [训练数据](https://siyuanailab.com/knowledge/ai-core/foundations/training-data/) | Training Data | 用来训练模型的样本集合，模型从中学习语言、图像、行为或任务规律。 | 训练数据数据质量模型训练 | | 01-10 | [预训练](https://siyuanailab.com/knowledge/ai-core/foundations/pre-training/) | Pre-training | 在大规模通用数据上先训练模型，让它获得基础语言和世界模式。 | 预训练模型训练LLM | ## [模型与机制](https://siyuanailab.com/knowledge/ai-core/models/) 理解模型如何学习、表示、生成和推理 12 张 模型与机制词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 02-01 | [Token](https://siyuanailab.com/knowledge/ai-core/models/token/) | Token | 模型处理文本时的基本片段，可能是一个字、词、词根、标点或代码片段。 | Token上下文成本 | | 02-02 | [上下文窗口](https://siyuanailab.com/knowledge/ai-core/models/context-window/) | Context Window | 模型一次对话或调用中能看到的输入与历史内容总容量。 | 上下文窗口ContextLLM | | 02-03 | [注意力机制](https://siyuanailab.com/knowledge/ai-core/models/attention-mechanism/) | Attention Mechanism | 让模型在处理某个位置时，动态关注输入中更相关部分的机制。 | 注意力机制Transformer模型机制 | | 02-04 | [嵌入/向量](https://siyuanailab.com/knowledge/ai-core/models/embedding/) | Embedding | 把文本、图片或其他对象转成数字向量，让机器能比较它们的语义距离。 | Embedding向量语义搜索 | | 02-05 | [检索增强生成](https://siyuanailab.com/knowledge/ai-core/models/retrieval-augmented-generation/) | Retrieval-Augmented Generation | 先从外部资料中检索相关内容，再把检索结果交给模型生成回答的方法。 | RAG检索增强生成知识库 | | 02-06 | [微调](https://siyuanailab.com/knowledge/ai-core/models/fine-tuning/) | Fine-tuning | 在已有模型基础上，用特定数据继续训练，让模型更适合某类任务或风格。 | 微调模型训练LLM | | 02-07 | [提示词工程](https://siyuanailab.com/knowledge/ai-core/models/prompt-engineering/) | Prompt Engineering | 通过设计输入、指令、示例和约束，让模型更稳定地完成任务的方法。 | Prompt Engineering提示词交互 | | 02-09 | [思维链](https://siyuanailab.com/knowledge/ai-core/models/chain-of-thought/) | Chain-of-Thought | 通过引导模型分步骤推理，提升复杂问题求解表现的方法。 | CoT思维链推理 | | 02-10 | [推理](https://siyuanailab.com/knowledge/ai-core/models/reasoning/) | Reasoning | 模型把已知信息连接起来，进行多步判断、归纳、演绎或规划的能力。 | 推理ReasoningLLM | | 02-11 | [幻觉](https://siyuanailab.com/knowledge/ai-core/models/hallucination/) | Hallucination | 模型生成了看起来可信、但事实不准确或没有依据的内容。 | 幻觉可靠性LLM风险 | | 02-12 | [对齐](https://siyuanailab.com/knowledge/ai-core/models/alignment/) | Alignment | 让模型行为更符合人类意图、价值、安全边界和任务要求的训练与约束过程。 | 对齐AlignmentAI安全 | | 02-13 | [温度/采样参数](https://siyuanailab.com/knowledge/ai-core/models/temperature/) | Temperature | 控制模型生成随机性和多样性的参数，温度越高通常越发散。 | Temperature采样参数生成控制 | ## [Agent 与智能体系统](https://siyuanailab.com/knowledge/ai-core/agents/) 从工具调用到多智能体协作的行动系统 8 张 Agent 与智能体系统词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 03-01 | [智能体](https://siyuanailab.com/knowledge/ai-core/agents/ai-agent/) | AI Agent | 以模型为核心，能接收目标、规划步骤、调用工具并根据反馈持续行动的软件实体。 | AI Agent智能体工具调用 | | 03-02 | [多智能体系统](https://siyuanailab.com/knowledge/ai-core/agents/multi-agent-system/) | Multi-Agent System | 多个智能体按角色分工协作，共同完成单个智能体难以处理的任务。 | 多智能体Agent协作 | | 03-04 | [工具调用](https://siyuanailab.com/knowledge/ai-core/agents/tool-use-function-calling/) | Tool Use / Function Calling | 模型通过标准接口调用搜索、计算、数据库、代码执行或业务 API 等外部工具。 | 工具调用Function CallingAgent | | 03-05 | [记忆系统](https://siyuanailab.com/knowledge/ai-core/agents/memory-system/) | Memory System | 让智能体在短期上下文之外保存、检索和更新信息的机制。 | 记忆系统Agent Memory上下文 | | 03-06 | [规划能力](https://siyuanailab.com/knowledge/ai-core/agents/planning/) | Planning | 智能体把目标拆成步骤、安排顺序，并根据反馈调整路径的能力。 | Planning规划Agent | | 03-07 | [委托机制](https://siyuanailab.com/knowledge/ai-core/agents/delegation/) | Delegation | 把任务从主控智能体分派给子智能体、工具或人类，并回收结果的机制。 | Delegation委托Subagent | | 03-08 | [模型上下文协议](https://siyuanailab.com/knowledge/ai-core/agents/model-context-protocol/) | Model Context Protocol | 一种让 AI 应用以标准方式连接工具、数据源和上下文服务的开放协议。 | MCPModel Context Protocol工具协议 | | 03-09 | [智能体工作流](https://siyuanailab.com/knowledge/ai-core/agents/agentic-workflow/) | Agentic Workflow | 把模型、工具、人工检查和状态流组合成可重复运行的任务流程。 | Agentic Workflow工作流AI工程 | ## [工程与架构](https://siyuanailab.com/knowledge/ai-core/engineering/) 把 AI 能力放进真实业务流程的工程方法 7 张 工程与架构词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 04-01 | [AI 工程学](https://siyuanailab.com/knowledge/ai-core/engineering/ai-engineering/) | AI Engineering | 把大模型能力变成可靠、可控、可维护产品和流程的工程方法。 | AI工程学AI Engineering工程化 | | 04-02 | [LLM 网关](https://siyuanailab.com/knowledge/ai-core/engineering/llm-gateway/) | LLM Gateway | 位于应用和多个模型供应商之间的统一入口，用于管理调用、路由、限流、日志和成本。 | LLM网关模型管理AI工程 | | 04-04 | [测试时计算](https://siyuanailab.com/knowledge/ai-core/engineering/test-time-compute/) | Test-time Compute | 在模型回答阶段投入更多计算，如多步推理、多次采样、搜索或自检，以换取更好结果。 | Test-time Compute推理成本 | | 04-06 | [上下文工程](https://siyuanailab.com/knowledge/ai-core/engineering/context-engineering/) | Context Engineering | 系统性选择、组织、压缩和注入模型所需上下文的工程方法。 | 上下文工程ContextAI工程 | | 04-07 | [Vibe Coding](https://siyuanailab.com/knowledge/ai-core/engineering/vibe-coding/) | Vibe Coding | 用自然语言和 AI 编程工具快速表达意图、生成代码、边跑边改的开发方式。 | Vibe CodingAI编程原型 | | 04-08 | [人参与循环](https://siyuanailab.com/knowledge/ai-core/engineering/human-in-the-loop/) | Human-in-the-loop | 在 AI 流程中保留人类审核、纠偏或最终决策的机制。 | Human-in-the-loop人机协同治理 | | 04-09 | [评估方法](https://siyuanailab.com/knowledge/ai-core/engineering/evaluation-methods/) | Evaluation Methods | 用测试集、基准、人工审核、自动指标或实战任务来判断 AI 系统表现的方法。 | 评估Evals质量 | ## [交互与提示](https://siyuanailab.com/knowledge/ai-core/interaction/) 人与模型协作时的提示、上下文和交互结构 4 张 交互与提示词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 05-01 | [系统提示词](https://siyuanailab.com/knowledge/ai-core/interaction/system-prompt/) | System Prompt | 由应用或开发者设置的高优先级指令，用来规定模型角色、边界和行为规则。 | System Prompt系统提示词Prompt | | 05-03 | [少样本提示](https://siyuanailab.com/knowledge/ai-core/interaction/few-shot-prompting/) | Few-shot Prompting | 在提示中提供几个输入输出示例，让模型模仿格式、风格或判断标准。 | Few-shot PromptingPrompt示例 | | 05-04 | [角色提示](https://siyuanailab.com/knowledge/ai-core/interaction/role-prompting/) | Role Prompting | 通过指定模型扮演的专业角色、视角或职责，影响其回答方式。 | Role PromptingPrompt交互 | | 05-05 | [提示词注入](https://siyuanailab.com/knowledge/ai-core/interaction/prompt-injection/) | Prompt Injection | 攻击者通过输入诱导模型忽略原有规则、泄露信息或执行不该执行的动作。 | Prompt Injection安全LLM风险 | ## [安全与治理](https://siyuanailab.com/knowledge/ai-core/safety/) AI 系统的风险、边界、评估和治理机制 4 张 安全与治理词条列表 | 编号 | 词条 | 英文名 | 一句话摘要 | 标签 | | --- | --- | --- | --- | --- | | 06-01 | [AI 安全](https://siyuanailab.com/knowledge/ai-core/safety/ai-safety/) | AI Safety | 研究和实践如何让 AI 系统可靠、可控、不会造成不可接受风险的领域。 | AI安全治理风险 | | 06-03 | [智能体治理](https://siyuanailab.com/knowledge/ai-core/safety/agent-governance/) | Agent Governance | 对智能体的权限、目标、日志、审核、责任和生命周期进行管理的机制。 | Agent Governance智能体治理AI安全 | | 06-04 | [模型偏见](https://siyuanailab.com/knowledge/ai-core/safety/model-bias/) | Model Bias | 模型因训练数据、目标函数或使用方式，对某些群体、观点或场景产生系统性不公平或失真。 | 模型偏见公平性治理 | | 06-05 | [数据隐私](https://siyuanailab.com/knowledge/ai-core/safety/data-privacy/) | Data Privacy | 在 AI 系统中保护个人、客户和组织敏感数据不被不当收集、使用、泄露或再训练。 | 数据隐私隐私安全 |