[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"kb-list":3,"kb-list-daily":4},[4,18,30,43,56],{"id":5,"title":6,"summary":7,"source":8,"category":9,"date":10,"url":11,"content":12},"agent-stack-2026","AI 智能体技术栈 2026 全景图：从 RAG 到 MCP","一文读懂大模型应用落地的核心组件：检索增强生成、函数调用、记忆管理与模型上下文协议，附热门框架选型对比。","量子位","智能体","2026-08-12","https:\u002F\u002Fwww.qbitai.com\u002F",[13,14,15,16,17],"智能体（Agent）正在成为大模型落地的主流形态。本文从工程视角梳理 2026 年主流技术栈。","检索增强生成（RAG）仍是企业知识库问答的默认方案：向量库 + 重排序 + 引用溯源，可以显著降低幻觉。","函数调用（Function Calling）让模型可以主动操作外部工具；模型上下文协议（MCP）则统一了工具接入标准。","记忆管理分为短期对话记忆与长期业务记忆，前者依赖上下文窗口，后者依赖向量检索与结构化存储。","选型建议：小团队优先使用成熟编排框架（如 LangGraph、Dify），降低自研成本；对延迟敏感场景再考虑轻量自建。",{"id":19,"title":20,"summary":21,"source":22,"category":23,"date":24,"url":25,"content":26},"multimodal-vlms","多模态大模型新进展：视觉理解走向实时","从图像生成到视频理解，多模态模型正在突破实时性瓶颈，端侧部署成为新焦点。","机器之心","多模态","2026-08-11","https:\u002F\u002Fwww.jiqizhixin.com\u002F",[27,28,29],"多模态大模型（VLM）正从\"看图说话\"走向实时视频理解与跨模态生成。","端侧部署是 2026 年的关键趋势：量化压缩与异构计算让 70B 级模型可以在手机与边缘设备上运行。","统一输入输出架构（原生多模态）逐渐取代\"外挂视觉编码器\"的拼接方案，效果与效率同步提升。",{"id":31,"title":32,"summary":33,"source":34,"category":35,"date":36,"url":37,"content":38},"rag-advanced","高级 RAG 工程实践：从朴素检索到混合检索","结合稀疏检索与稠密检索，通过查询改写、路由与重排序，把 RAG 精度提升一个量级。","InfoQ 中文","RAG","2026-08-10","https:\u002F\u002Fwww.infoq.cn\u002F",[39,40,41,42],"朴素 RAG（向量检索 + 拼接提示）在复杂问题上表现不稳定，工程化改造势在必行。","混合检索：BM25 稀疏检索负责精确匹配，向量检索负责语义召回，融合排序可兼顾两者优势。","查询改写（Query Rewriting）在多轮对话中尤为关键，可将口语化问题改写为可检索的表达。","重排序（Rerank）模型虽然带来额外延迟，但对精度提升显著，是生产环境的标配组件。",{"id":44,"title":45,"summary":46,"source":47,"category":48,"date":49,"url":50,"content":51},"llm-ops","LLMOps 落地指南：大模型应用的可观测性与评估","如何评估、追踪与治理大模型应用？覆盖提示管理、在线评测、链路追踪与安全护栏。","开源中国","工程实践","2026-08-09","https:\u002F\u002Fwww.oschina.net\u002F",[52,53,54,55],"LLMOps 解决的核心问题：模型不可解释、评测主观、线上问题难定位。","三层观测：调用级（Token\u002F成本\u002F延迟）、行为级（输出质量）、业务级（转化指标）。","在线评测（Online Eval）与回归集（Regression Set）相结合，是控制版本迭代风险的关键手段。","安全护栏（Guardrails）从输入过滤、输出校验到敏感信息脱敏，构建多层防线。",{"id":57,"title":58,"summary":59,"source":60,"category":61,"date":62,"url":63,"content":64},"open-source-llm","2026 开源大模型排行榜：中文能力全面跃升","开源模型在中文理解、推理与代码能力上全面逼近闭源第一梯队，生态工具链日益成熟。","GitHub Trending","开源","2026-08-08","https:\u002F\u002Fgithub.com\u002Ftrending",[65,66,67],"2026 年开源大模型在中文能力上取得显著突破，多个项目登顶各类榜单。","推理侧量化与推理加速工具（vLLM、SGLang）让个人开发者也能低成本运行大模型。","开源许可证与合规成为企业选型的重要考量，Apache-2.0 与 LLAMA 系许可证各有适用场景。"]