<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>AI 学习手记</title><description>从原始资料出发，用笔记理解，用实验验证，沿着路径系统学习 AI。</description><link>https://blog.foov.cc/</link><language>zh-CN</language><item><title>Agent 上下文预算 Skills 与前缀缓存</title><link>https://blog.foov.cc/notes/agent-context-skills-cache/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-context-skills-cache/</guid><description>区分上下文、运行状态、技能说明和缓存，按需加载知识并验证压缩后的决策信息。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Agent 评估指标 轨迹边界与协作成本</title><link>https://blog.foov.cc/notes/agent-evaluation-coordination/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-evaluation-coordination/</guid><description>区分探索成功与重复可靠性，用任务和轨迹前缀验证行为，并比较多 Agent 的信息收益、冲突与预算。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>事件驱动 Agent 的取消与交互状态</title><link>https://blog.foov.cc/notes/agent-event-interaction/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-event-interaction/</guid><description>将外部消息、迟到工具结果和用户修订建模为事件，控制任务归属、抢占、重复动作及界面状态验证。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Harness 与 Coding Agent 的执行和验收</title><link>https://blog.foov.cc/notes/agent-harness-coding/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-harness-coding/</guid><description>把模型建议、工具执行和任务完成分开，用工作区隔离、补丁、验证器、预算与恢复构成运行环境。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>结构化记忆与知识索引的来源和权限</title><link>https://blog.foov.cc/notes/agent-structured-memory/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-structured-memory/</guid><description>用主体、时间、来源和修订表达记忆，比较文本、关系、摘要树与 GraphRAG 的检索任务和验证方式。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Agent 轨迹学习 后训练与持续改进</title><link>https://blog.foov.cc/notes/agent-trajectory-learning/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-trajectory-learning/</guid><description>从可验证失败定位改动，区分知识、指令、程序和参数更新，构造训练任务并使用独立回归与发布门禁。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI Agent 教材十章对照与知识补充</title><link>https://blog.foov.cc/notes/ai-agent-book-comparison/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-agent-book-comparison/</guid><description>将 ai-agent-book 的十章与现有知识库逐项对照，补充六个工程专题，明确学习顺序与实验边界。</description><pubDate>Sat, 10 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 应用实战路线：从开发到架构与面试</title><link>https://blog.foov.cc/notes/ai-application-roadmap/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-application-roadmap/</guid><description>衔接现有基础、后端、RAG/agent 综合项目、实测与架构讨论，按可交付成果学习。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 系统大纲：核心、支线与完成标准</title><link>https://blog.foov.cc/notes/ai-curriculum/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-curriculum/</guid><description>对照 GitHub 教材与开放课程，组织 61 篇笔记的前置顺序、知识清单、支线和综合任务。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 知识地图：八个模块与学习顺序</title><link>https://blog.foov.cc/notes/ai-knowledge-map/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-knowledge-map/</guid><description>把常见 AI 方向组织成 8 个模块、61 篇笔记与 10 条路径，按前置知识、练习和原始来源学习。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>智能体项目：状态、轨迹与失败验证</title><link>https://blog.foov.cc/notes/agent-project-validation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-project-validation/</guid><description>从一个明确任务出发，用固定工作流作基线，验证状态、工具成功、预算和恢复。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 容量与成本：并发、存储和请求预算</title><link>https://blog.foov.cc/notes/ai-capacity-cost/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-capacity-cost/</guid><description>用明确单位估算流量、在途请求、向量存储和模型费用，再用真实负载修正假设。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 开发与架构面试：六类案例和回答证据</title><link>https://blog.foov.cc/notes/ai-interview-casebook/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-interview-casebook/</guid><description>用项目、容量、权限、可靠性与工具故障练习技术讨论，区分理解、实现和生产经验。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 项目评估报告：基线、对照与失败证据</title><link>https://blog.foov.cc/notes/ai-project-evaluation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-project-evaluation/</guid><description>把模型效果、检索质量、权限、工具和服务表现写成可复核报告，明确样本、分母与改动归因。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 服务后端：流式响应、取消与重试</title><link>https://blog.foov.cc/notes/ai-service-backend/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-service-backend/</guid><description>把模型应用接入服务工程：定义请求契约、流式事件、截止时间、幂等和身份传播。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 服务可靠性：SLI、SLO 与错误预算</title><link>https://blog.foov.cc/notes/ai-slo-reliability/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-slo-reliability/</guid><description>把用户体验转成可测量的服务目标，区分流式完成、答案质量、告警和故障处置。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 系统设计：边界、数据流与架构取舍</title><link>https://blog.foov.cc/notes/ai-system-design/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-system-design/</guid><description>从业务约束设计知识助手的数据、检索、工具与运行边界，说明扩展和一致性的理由。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>RAG 综合项目：任务拆解与验收清单</title><link>https://blog.foov.cc/notes/rag-capstone-spec/</link><guid isPermaLink="true">https://blog.foov.cc/notes/rag-capstone-spec/</guid><description>围绕自制知识库完成检索、生成、权限、更新、评估和交付，明确每阶段应提供的证据。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>权限与缓存实验：同问不同人、撤权与版本</title><link>https://blog.foov.cc/notes/retrieval-access-cache-lab/</link><guid isPermaLink="true">https://blog.foov.cc/notes/retrieval-access-cache-lab/</guid><description>用 3 份自制文档和 10 个顺序场景复现缓存越权与过期，记录基线、改进和版本漏更新的边界。</description><pubDate>Thu, 08 Oct 2026 00:00:00 GMT</pubDate></item><item><title>上下文与记忆：选择、压缩、隔离与遗忘</title><link>https://blog.foov.cc/notes/agent-context-memory/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-context-memory/</guid><description>区分当前输入、任务状态、长期记忆和知识库，建立来源、时间、用户边界与删除机制。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>智能体执行循环：模型、工具、状态与停止</title><link>https://blog.foov.cc/notes/agent-control-loop/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-control-loop/</guid><description>将 agent 看作受约束的执行系统，理解工具选择、结果观察、状态持久化和可验证终止。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>工作流与多智能体：拆解、路由、协作和交接</title><link>https://blog.foov.cc/notes/agent-workflows-orchestration/</link><guid isPermaLink="true">https://blog.foov.cc/notes/agent-workflows-orchestration/</guid><description>按任务依赖选择固定流程、路由或多智能体，明确状态所有权、合并规则与失败恢复。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 数学基础：张量、梯度与概率</title><link>https://blog.foov.cc/notes/ai-mathematics/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-mathematics/</guid><description>把线性代数、微积分和概率连接到真实模型操作：看懂形状、梯度更新与不确定性。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 应用安全：数据、权限、注入与执行</title><link>https://blog.foov.cc/notes/ai-security-boundaries/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ai-security-boundaries/</guid><description>把安全落实到认证身份、资料可见性与工具执行程序，理解提示注入、泄漏与数据生命周期。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>语音 AI：音频数据、ASR、TTS 与评估</title><link>https://blog.foov.cc/notes/audio-asr-tts/</link><guid isPermaLink="true">https://blog.foov.cc/notes/audio-asr-tts/</guid><description>从采样率和频谱进入转写与语音合成，区分 WER/CER、自然度、流式延迟和说话人边界。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>概率校准与不确定性：分数能否支持决策</title><link>https://blog.foov.cc/notes/calibration-uncertainty/</link><guid isPermaLink="true">https://blog.foov.cc/notes/calibration-uncertainty/</guid><description>区分分类准确率、排序、校准和不确定性，避免把高分或 token 概率当事实置信度。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>数据漂移与监控：分布变化不等于质量下降</title><link>https://blog.foov.cc/notes/data-drift-monitoring/</link><guid isPermaLink="true">https://blog.foov.cc/notes/data-drift-monitoring/</guid><description>区分输入、标签先验和概念漂移，结合数据质量、迟到标签及告警调查维护模型。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>数据与特征管线：先划分，再拟合</title><link>https://blog.foov.cc/notes/data-feature-pipelines/</link><guid isPermaLink="true">https://blog.foov.cc/notes/data-feature-pipelines/</guid><description>把数据探索、缺失处理、类别编码、标准化和特征选择放进可复用的训练管线，避免泄漏。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>深度学习训练：计算图、损失与优化</title><link>https://blog.foov.cc/notes/deep-learning-training/</link><guid isPermaLink="true">https://blog.foov.cc/notes/deep-learning-training/</guid><description>理解神经网络训练闭环、过拟合与复现记录，区分训练目标改善和实际能力改善。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>扩散与图像生成：噪声、条件、调度器与复现</title><link>https://blog.foov.cc/notes/diffusion-generation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/diffusion-generation/</guid><description>理解迭代去噪、潜空间与生成管线，区分模型、scheduler、引导强度与可重复运行条件。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Embedding 与向量检索：表示、相似度和索引</title><link>https://blog.foov.cc/notes/embeddings-vector-search/</link><guid isPermaLink="true">https://blog.foov.cc/notes/embeddings-vector-search/</guid><description>理解文本向量、距离、归一化与近似检索，避免把相似度分数当作事实置信度。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>实验统计：配对比较、区间与显著性</title><link>https://blog.foov.cc/notes/experiment-statistics/</link><guid isPermaLink="true">https://blog.foov.cc/notes/experiment-statistics/</guid><description>用独立单位、配对重采样和置换检验分析差异，区分离线评估与线上因果实验。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>公平性与可解释性：分组评估和解释边界</title><link>https://blog.foov.cc/notes/fairness-explainability/</link><guid isPermaLink="true">https://blog.foov.cc/notes/fairness-explainability/</guid><description>检查数据与错误的群体差异，理解公平指标取舍和 permutation importance 的局限。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>LLM 评估：样本、评分、评判模型与回归</title><link>https://blog.foov.cc/notes/llm-evaluation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/llm-evaluation/</guid><description>把格式、任务成功、事实支持和执行成本分别测量，建立独立测试集与可追溯错误分类。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>LLM 训练与适配：预训练、SFT、LoRA 与对齐</title><link>https://blog.foov.cc/notes/llm-training-adaptation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/llm-training-adaptation/</guid><description>把知识输入、任务学习和偏好优化分开，理解聊天模板、数据质量与微调评估。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>机器学习工作流：任务、数据与泛化</title><link>https://blog.foov.cc/notes/machine-learning-workflow/</link><guid isPermaLink="true">https://blog.foov.cc/notes/machine-learning-workflow/</guid><description>从回归、分类和聚类开始，建立数据划分、基线、特征处理与评估的完整顺序。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>MCP：协议、工具发现、版本与授权边界</title><link>https://blog.foov.cc/notes/mcp-tool-boundaries/</link><guid isPermaLink="true">https://blog.foov.cc/notes/mcp-tool-boundaries/</guid><description>以 2026-07-28 官方规范理解 MCP，区分连接、身份、能力与工具执行权限。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>机器学习评估：指标、阈值与交叉验证</title><link>https://blog.foov.cc/notes/ml-metrics-validation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/ml-metrics-validation/</guid><description>手算混淆矩阵和回归误差，理解 ROC/PR、分组与时间划分、嵌套验证和阈值选择。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>MLOps 生命周期：版本、测试、发布与回滚</title><link>https://blog.foov.cc/notes/mlops-lifecycle/</link><guid isPermaLink="true">https://blog.foov.cc/notes/mlops-lifecycle/</guid><description>把数据、模型、代码和配置形成可追溯版本，明确数据测试、行为测试与模型发布门槛。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 模型选型：先看任务，再比较模型</title><link>https://blog.foov.cc/notes/model-selection/</link><guid isPermaLink="true">https://blog.foov.cc/notes/model-selection/</guid><description>区分产品、模型、权重和服务；用任务、版本、数据边界与固定样本建立可复查的选型表。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>NLP 任务建模：分类、标注、问答与生成</title><link>https://blog.foov.cc/notes/nlp-task-modeling/</link><guid isPermaLink="true">https://blog.foov.cc/notes/nlp-task-modeling/</guid><description>按任务区分文本表示、标签对齐和评价口径，不把所有语言任务都简化为聊天生成。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>优化与正则化：学习率、泛化与训练曲线</title><link>https://blog.foov.cc/notes/optimization-regularization/</link><guid isPermaLink="true">https://blog.foov.cc/notes/optimization-regularization/</guid><description>区分优化算法和模型容量，学习 SGD、动量、Adam、正则化、初始化与早停的检查顺序。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>提示设计：任务、上下文、示例与输出约束</title><link>https://blog.foov.cc/notes/prompt-design/</link><guid isPermaLink="true">https://blog.foov.cc/notes/prompt-design/</guid><description>用可测试的任务说明替代模糊要求，区分零样本、少样本、任务拆解与证据边界。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>PyTorch 训练实践：数据、梯度与 checkpoint</title><link>https://blog.foov.cc/notes/pytorch-training-practice/</link><guid isPermaLink="true">https://blog.foov.cc/notes/pytorch-training-practice/</guid><description>把 Dataset、DataLoader、张量形状、训练/推理模式和保存恢复连成可核查的训练流程。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>RAG 评估：召回、排序、忠实度与正确性</title><link>https://blog.foov.cc/notes/rag-evaluation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/rag-evaluation/</guid><description>把检索和回答拆开评价，明确 context recall、context precision、faithfulness 与事实正确性的参照物。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>RAG 文档管线：解析、清洗、切片与版本</title><link>https://blog.foov.cc/notes/rag-ingestion-chunking/</link><guid isPermaLink="true">https://blog.foov.cc/notes/rag-ingestion-chunking/</guid><description>保留文档结构、来源位置和可见范围，让入库、更新、删除与引用形成闭环。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>RAG 检索：召回、混合搜索与重排</title><link>https://blog.foov.cc/notes/rag-retrieval-reranking/</link><guid isPermaLink="true">https://blog.foov.cc/notes/rag-retrieval-reranking/</guid><description>拆开候选召回和精排，理解关键词、向量、过滤、查询改写与 top-k 的代价。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>推荐系统：召回、排序、反馈与冷启动</title><link>https://blog.foov.cc/notes/recommendation-ranking/</link><guid isPermaLink="true">https://blog.foov.cc/notes/recommendation-ranking/</guid><description>理解内容推荐、协同过滤、矩阵分解与多阶段排序，处理未曝光反馈和离线评价的边界。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>强化学习：奖励、价值与策略</title><link>https://blog.foov.cc/notes/reinforcement-learning/</link><guid isPermaLink="true">https://blog.foov.cc/notes/reinforcement-learning/</guid><description>建立状态、动作、奖励与回报的概念，理解 Q-learning、策略梯度和 PPO 的学习目标。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>AI 推理工程：延迟、缓存、吞吐与复现</title><link>https://blog.foov.cc/notes/serving-performance/</link><guid isPermaLink="true">https://blog.foov.cc/notes/serving-performance/</guid><description>区分 prefill、decode 和端到端延迟，理解 KV/prefix cache、批处理与量化的工程取舍。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>结构化输出与工具调用：格式正确之后还要验证</title><link>https://blog.foov.cc/notes/structured-output-tools/</link><guid isPermaLink="true">https://blog.foov.cc/notes/structured-output-tools/</guid><description>区分 JSON、schema、工具提议和实际执行，建立类型、业务、权限与幂等检查。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>监督学习算法：线性、树、集成与距离</title><link>https://blog.foov.cc/notes/supervised-models/</link><guid isPermaLink="true">https://blog.foov.cc/notes/supervised-models/</guid><description>对比回归和分类中的常用算法，理解假设、预处理、复杂度与基线选择。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>时序预测：窗口、基线与滚动验证</title><link>https://blog.foov.cc/notes/time-series-forecasting/</link><guid isPermaLink="true">https://blog.foov.cc/notes/time-series-forecasting/</guid><description>围绕预测时点组织 lag、季节性和窗口特征，避免未来信息泄漏和随机切分的乐观误差。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>迁移学习与数据增强：冻结、微调和标签一致性</title><link>https://blog.foov.cc/notes/transfer-learning-augmentation/</link><guid isPermaLink="true">https://blog.foov.cc/notes/transfer-learning-augmentation/</guid><description>从预训练表示到目标任务，检查冻结参数、BatchNorm 状态、增强标签和训练验证隔离。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Transformer：注意力、位置与生成</title><link>https://blog.foov.cc/notes/transformer-attention/</link><guid isPermaLink="true">https://blog.foov.cc/notes/transformer-attention/</guid><description>看懂 Q、K、V 的作用、因果掩码与三类架构，并把训练并行和生成顺序分开理解。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>无监督学习：聚类、降维与异常检测</title><link>https://blog.foov.cc/notes/unsupervised-learning/</link><guid isPermaLink="true">https://blog.foov.cc/notes/unsupervised-learning/</guid><description>理解 k-means、密度聚类、混合模型、PCA 和异常检测，区分发现结构与证明业务事实。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>视觉与视觉语言模型：像素、任务与坐标</title><link>https://blog.foov.cc/notes/vision-foundations/</link><guid isPermaLink="true">https://blog.foov.cc/notes/vision-foundations/</guid><description>从图像张量、分类、检测和分割进入视觉问答，明确预处理、空间信息与文档理解边界。</description><pubDate>Wed, 07 Oct 2026 00:00:00 GMT</pubDate></item><item><title>评估先行：固定问题，再比较改动</title><link>https://blog.foov.cc/notes/evaluation-baseline/</link><guid isPermaLink="true">https://blog.foov.cc/notes/evaluation-baseline/</guid><description>先明确成功条件，保存固定输入与失败记录。小样本适合发现回归，但不能代表真实用户总体表现。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate></item><item><title>动手实验：关键词检索会漏掉什么</title><link>https://blog.foov.cc/notes/keyword-retrieval-lab/</link><guid isPermaLink="true">https://blog.foov.cc/notes/keyword-retrieval-lab/</guid><description>在 3 条自写语料、5 个固定查询上比较精确与规范化匹配。修复大小写和全角字符后，同义表达依然未命中。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate></item><item><title>检索增强 RAG：让答案回到证据</title><link>https://blog.foov.cc/notes/rag-evidence/</link><guid isPermaLink="true">https://blog.foov.cc/notes/rag-evidence/</guid><description>把检索、生成和引用检查分开。找到一段相关文本只是起点，还需要判断它能否支持答案以及何时应该不作答。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate></item><item><title>Token 与上下文：先分清计量单位</title><link>https://blog.foov.cc/notes/token-context/</link><guid isPermaLink="true">https://blog.foov.cc/notes/token-context/</guid><description>Token 的切分由分词器决定。设计输入预算时，先确认模型与分词器，再把输入、历史和输出需求一起列出来。</description><pubDate>Sat, 03 Oct 2026 00:00:00 GMT</pubDate></item></channel></rss>