第1章 大模型——Agent的大脑 Ch01 LLM — Agent's Brain
从预测下一个词到涌现推理能力
From next-token prediction to emergent reasoning
从预测下一个词到涌现推理能力
From next-token prediction to emergent reasoning
Agent = Model + Harness,ReAct循环与三年进化
Agent = Model + Harness, ReAct loop and 3-year evolution
从Prompt到Context到Harness,三年三次范式演进
From Prompt to Context to Harness, three paradigm shifts
ReAct循环、有向图、多Agent协作
ReAct loop, directed graphs, multi-agent collaboration
Function Calling、MCP协议与工具设计
Function Calling, MCP protocol and tool design
短期/长期记忆、RAG与知识编译
Short/long-term memory, RAG and knowledge compilation
代码沙箱、Docker容器与E2B云沙箱
Code sandboxes, Docker and E2B cloud sandboxes
状态管理、检查点与持久化
State management, checkpoints and persistence
三层安全防线与人类审批
Three-layer defense and human approval
用自然语言构建AI产品
Building AI products with natural language
LangGraph/CrewAI/AutoGen/Dify对比
Comparison of LangGraph/CrewAI/AutoGen/Dify
可观测性、评估与成本优化
Observability, evaluation and cost optimization
自进化Agent、具身智能与行业渗透
Self-evolving Agents, embodied intelligence and verticalization
四家 harness 对照,以及灵犀(Lumina)产品结构与落地
Four harnesses compared, plus Lumina (灵犀) product structure
用有向无环图(DAG)表示和调度复杂任务里的子任务和依赖关系
Using Directed Acyclic Graphs to represent and orchestrate complex task dependencies
LLMOps管大模型从开发到上线的全生命周期,RAI确保安全、公平、可信、合规
LLMOps for full LLM lifecycle management, RAI for safety, fairness, transparency and compliance
基于渣打银行 RAI 框架和 LLMOps 生产栈的 RAG 项目深度分析
Deep dive into RAG project mapped to SCB RAI framework and LLMOps production stack
在复杂 Text-to-SQL 场景下,图搜索或树搜索比 RAG 效果更好、更精准
Graph and tree search outperform RAG in complex Text-to-SQL scenarios
从基础原理到企业级架构的 Text-to-SQL 深度拆解
Deep analysis of Text-to-SQL from fundamentals to enterprise architecture