CINXE.COM
<!doctype html><html><head> <script> var e=function(e,t,a){if(Math.ceil(100*Math.random())<=100*t){var o="491065",d="24814",n={ev_type:"batch",list:[{ev_type:"custom",payload:{name:"sdk_glue_load",type:"event",metrics:{},categories:{sdk_glue_load_status:e,sdk_glue_load_err_src:a,payload_bdms_aid:o,payload_bdms_page_id:d}},common:{context:{ctx_bdms_aid:o,ctx_bdms_page_id:d},bid:"web_bdms_cn",pid:window.location.pathname,view_id:"/_1",user_id:"",session_id:"0-a-1-2-c",release:"",env:"production",url:window.location.href,timestamp:+new Date,sdk_version:"1.6.1",sdk_name:"SDK_SLARDAR_WEB"}}]},i=new XMLHttpRequest;i.open("POST","https://mon.zijieapi.com/monitor_browser/collect/batch/?biz_id=web_bdms_cn",!0),i.setRequestHeader("Content-type","application/json"),i.send(JSON.stringify(n))}};e("before_load",.1,""),window.addEventListener("error",(function(t){var a=t.target||t.srcElement;a instanceof HTMLElement&&"SCRIPT"==a.nodeName&&(-1!=(a.src||"").indexOf("sdk-glue")&&e("load_error",1,a.src))}),!0);</script><script src="https://lf-headquarters-speed.yhgfb-cn-static.com/obj/rc-client-security/web/glue/1.0.0.38/sdk-glue.js"></script><script> ;(function (){ var sdkInfo = { csrf: { init: function (options) {window.secsdk.csrf.setOptions(options)}, isLoaded: function () { return !!window.secsdk }, srcList: ["https://lf1-cdn-tos.bytegoofy.com/obj/goofy/secsdk/secsdk-lastest.umd.js","https://lf3-cdn-tos.bytegoofy.com/obj/goofy/secsdk/secsdk-lastest.umd.js","https://lf6-cdn-tos.bytegoofy.com/obj/goofy/secsdk/secsdk-lastest.umd.js"] }, bdms: { init: function (options) {window.bdms.init(options)}, isLoaded: function () { return !!window.bdms }, srcList: ["https://lf-c-flwb.bytetos.com/obj/rc-client-security/web/stable/1.0.0.43/bdms.js","https://lf-headquarters-speed.yhgfb-cn-static.com/obj/rc-client-security/web/stable/1.0.0.43/bdms.js"], }, verifyCenter: { init: function (options) {window.verifySDK.init(options)}, isLoaded: function () { return !!window.verifySDK }, srcList: ["https://lf-rc1.yhgfb-cn-static.com/obj/rc-client-security/secsdk-captcha/2.28.11/captcha.js","https://lf-rc2.yhgfb-cn-static.com/obj/rc-client-security/secsdk-captcha/2.28.11/captcha.js"] }, }; var options = { bdms: {aid:491065,pageId:24814,paths:["/api/fe"],ddrt:3},self: {aid:491065,pageId:24814,} }; window._SdkGlueInit(options, sdkInfo); })()</script> <script>window.gfdatav1={"env":"prod","idc":"lf","ver":"1.0.1.922","canary":0,"envName":"prod","region":"cn","runtime":"workerV2","extra":{"canaryType":null}}</script><script>window._SERVER_DATA={"router":{"baseUrl":"/","params":{}}}</script><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1,shrink-to-fit=no,viewport-fit=cover,minimum-scale=1,maximum-scale=1,user-scalable=no"><meta http-equiv="x-ua-compatible" content="ie=edge"><meta name="renderer" content="webkit"><meta name="layoutmode" content="standard"><meta name="imagemode" content="force"><meta name="wap-font-scale" content="no"><meta name="format-detection" content="telephone=no"><title data-react-helmet="true">Agent系列:AppAgent v2-屏幕智能Agent(详解版) - 文章 - 开发者社区 - 火山引擎</title><link rel="icon" href="//lf1-cdn-tos.bytegoofy.com/goofy/tech-fe/fav.png"/><link rel="apple-touch-icon" href="//lf1-cdn-tos.bytegoofy.com/goofy/tech-fe/logo193.png"/><meta name="referrer" content="always"><script>// Tea SDK 引导代码 (function (win, export_obj) { win['LogAnalyticsObject'] = export_obj; if (!win[export_obj]) { function _collect() { _collect.q.push(arguments); } _collect.q = _collect.q || []; win[export_obj] = _collect; } win[export_obj].l = +new Date(); })(window, 'collectEvent');</script><script>window.__assetPrefix__ = '//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325';</script><script>!function(){"use strict";var e,t,r,a,i,n,o,c={},f={};function s(e){var t=f[e];if(void 0!==t)return t.exports;var r=f[e]={id:e,loaded:!1,exports:{}};return c[e].call(r.exports,r,r.exports,s),r.loaded=!0,r.exports}s.m=c,e=[],s.O=function(t,r,a,i){if(!r){var n=1/0;for(d=0;d<e.length;d++){r=e[d][0],a=e[d][1],i=e[d][2];for(var o=!0,c=0;c<r.length;c++)(!1&i||n>=i)&&Object.keys(s.O).every((function(e){return s.O[e](r[c])}))?r.splice(c--,1):(o=!1,i<n&&(n=i));if(o){e.splice(d--,1);var f=a();void 0!==f&&(t=f)}}return t}i=i||0;for(var d=e.length;d>0&&e[d-1][2]>i;d--)e[d]=e[d-1];e[d]=[r,a,i]},s.n=function(e){var t=e&&e.__esModule?function(){return e.default}:function(){return e};return s.d(t,{a:t}),t},r=Object.getPrototypeOf?function(e){return Object.getPrototypeOf(e)}:function(e){return e.__proto__},s.t=function(e,a){if(1&a&&(e=this(e)),8&a)return e;if("object"==typeof e&&e){if(4&a&&e.__esModule)return e;if(16&a&&"function"==typeof e.then)return e}var i=Object.create(null);s.r(i);var n={};t=t||[null,r({}),r([]),r(r)];for(var o=2&a&&e;"object"==typeof o&&!~t.indexOf(o);o=r(o))Object.getOwnPropertyNames(o).forEach((function(t){n[t]=function(){return e[t]}}));return n.default=function(){return e},s.d(i,n),i},s.d=function(e,t){for(var r in t)s.o(t,r)&&!s.o(e,r)&&Object.defineProperty(e,r,{enumerable:!0,get:t[r]})},s.f={},s.e=function(e){return Promise.all(Object.keys(s.f).reduce((function(t,r){return s.f[r](e,t),t}),[]))},s.u=function(e){return"static/js/async/"+({572:"user/[id]/blog_moving/page",712:"user/[id]/draft/page",1225:"resource/[id]/page",1261:"user/edit/page",1687:"user/[id]/star/page",1826:"article/series/list/page",3180:"teams/[id]/articles/page",3487:"teams/page",4117:"mirror/page",4778:"articles/page",4786:"activity/[id]/page",4815:"user/[id]/qa/page",5400:"teams/[id]/page",5655:"teams/[id]/questions/page",5912:"user/[id]/page",6226:"questions/[id]/page",6242:"$",6480:"user/[id]/$",6772:"user/[id]/series/page",6827:"user/[id]/activity/page",6874:"articles/[id]/page",7281:"user/[id]/follow_list/page",7425:"resource/page",8079:"teams/[id]/layout",8138:"questions/page",8146:"user/$",8357:"user/page",8426:"article/series/detail/[id]/page",8528:"activities/[id]/page",8655:"team/[id]/page",8829:"activities/page",9130:"activity/page",9494:"team/page",9624:"user/[id]/articles/page",9768:"page",9844:"user/[id]/layout"}[e]||e)+"."+{407:"aefe6500",572:"88a5568c",585:"728b06d1",637:"32464f61",712:"dcdbbc36",756:"56e71d33",1008:"6f7bfd31",1094:"e7bbea48",1124:"830b6040",1225:"a6900ce3",1261:"b2e4a7dc",1289:"bf78b8e3",1687:"297ac6c9",1826:"4456d7f6",2057:"66169876",2128:"b6a13c1c",2417:"ae8d0598",2426:"48b8c954",2650:"28c0780d",3172:"b888ceda",3180:"034dce8a",3487:"e99070d3",3505:"fc596d93",3553:"27d193e6",4015:"0f26ce33",4117:"d449d71b",4211:"1803eb98",4433:"a0f3e45e",4778:"77241347",4786:"2be3f444",4815:"3258fa27",5086:"9c45d833",5092:"fc8ebe4d",5400:"2e7c6646",5435:"eccb96c2",5453:"b646a57c",5458:"9b8df362",5493:"c3044ac2",5505:"a3aa549d",5655:"49f52348",5777:"109859a7",5912:"cf73559a",5914:"1927ef78",5941:"59bcc29a",5998:"1d847129",6003:"48f9a706",6011:"22717107",6097:"8d0cacad",6226:"fceead16",6242:"8664c2ba",6336:"0ce21484",6480:"07b5287b",6637:"e866599a",6705:"a19b8e49",6772:"f9538e2d",6827:"c082f0a3",6874:"68d574c4",7056:"d6e157d4",7281:"8e9e57d5",7327:"9b4b3761",7425:"bf8f9096",7469:"24783622",7474:"4dbf3b1c",7518:"648426dd",7542:"a4ca3948",7644:"a5ed476c",7832:"cc5e6442",7921:"8b6f440c",8079:"1a309ced",8138:"55f1b21f",8146:"b9315b80",8357:"5acc5f96",8426:"98a6ffe6",8528:"befb1cc5",8626:"72051b87",8655:"b76ed95c",8829:"7bd373a2",8913:"3652e416",9107:"8db20dc7",9129:"69cb86a2",9130:"eb82bf48",9246:"633f6657",9404:"fa2c6b2f",9494:"ee37a04b",9624:"23bff0aa",9768:"fbf929cc",9791:"22e5ab4e",9844:"aafc33ee",9902:"ab79fd01"}[e]+".js"},s.miniCssF=function(e){return"static/css/async/"+({572:"user/[id]/blog_moving/page",712:"user/[id]/draft/page",1225:"resource/[id]/page",1261:"user/edit/page",1687:"user/[id]/star/page",1826:"article/series/list/page",3180:"teams/[id]/articles/page",3487:"teams/page",4117:"mirror/page",4778:"articles/page",4815:"user/[id]/qa/page",5400:"teams/[id]/page",5655:"teams/[id]/questions/page",5912:"user/[id]/page",6226:"questions/[id]/page",6242:"$",6772:"user/[id]/series/page",6827:"user/[id]/activity/page",6874:"articles/[id]/page",7281:"user/[id]/follow_list/page",7425:"resource/page",8079:"teams/[id]/layout",8138:"questions/page",8426:"article/series/detail/[id]/page",8528:"activities/[id]/page",8829:"activities/page",9624:"user/[id]/articles/page",9768:"page",9844:"user/[id]/layout"}[e]||e)+"."+{572:"5c7b055b",712:"72db1217",1225:"c771a8fa",1261:"556cf1b3",1687:"3de9a0fc",1826:"112c3666",2426:"01a02ce7",3180:"4fb2abd1",3487:"60e48adf",4117:"00b315f5",4778:"11c4f69f",4815:"06c5a7e1",5400:"43e4ce1a",5655:"4e016773",5912:"96b3cbfa",6226:"5c2bbaa5",6242:"6e3970d3",6772:"e6f14c09",6827:"f9a60515",6874:"68a000e0",7056:"365859d0",7281:"9b1d94e8",7425:"95aa4a33",7542:"f7fe289f",8079:"bc06c4c6",8138:"aed8f20f",8426:"00a692ef",8528:"7279aa4b",8829:"5f0e0daa",9107:"c4cafd96",9624:"96b3cbfa",9768:"394efaa3",9844:"9344d547"}[e]+".css"},s.g=function(){if("object"==typeof globalThis)return globalThis;try{return this||new Function("return this")()}catch(e){if("object"==typeof window)return window}}(),s.hmd=function(e){return(e=Object.create(e)).children||(e.children=[]),Object.defineProperty(e,"exports",{enumerable:!0,set:function(){throw new Error("ES Modules may not assign module.exports or exports.*, Use ESM export syntax, instead: "+e.id)}}),e},s.o=function(e,t){return Object.prototype.hasOwnProperty.call(e,t)},a={},i="volc-developer-fe:",s.l=function(e,t,r,n){if(a[e])a[e].push(t);else{var o,c;if(void 0!==r)for(var f=document.getElementsByTagName("script"),d=0;d<f.length;d++){var u=f[d];if(u.getAttribute("src")==e||u.getAttribute("data-webpack")==i+r){o=u;break}}o||(c=!0,(o=document.createElement("script")).charset="utf-8",o.timeout=120,s.nc&&o.setAttribute("nonce",s.nc),o.setAttribute("data-webpack",i+r),o.src=e,0!==o.src.indexOf(window.location.origin+"/")&&(o.crossOrigin="anonymous")),a[e]=[t];var l=function(t,r){o.onerror=o.onload=null,clearTimeout(p);var i=a[e];if(delete a[e],o.parentNode&&o.parentNode.removeChild(o),i&&i.forEach((function(e){return e(r)})),t)return t(r)},p=setTimeout(l.bind(null,void 0,{type:"timeout",target:o}),12e4);o.onerror=l.bind(null,o.onerror),o.onload=l.bind(null,o.onload),c&&document.head.appendChild(o)}},s.r=function(e){"undefined"!=typeof Symbol&&Symbol.toStringTag&&Object.defineProperty(e,Symbol.toStringTag,{value:"Module"}),Object.defineProperty(e,"__esModule",{value:!0})},s.nmd=function(e){return e.paths=[],e.children||(e.children=[]),e},s.p="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/",n=function(e){return new Promise((function(t,r){var a=s.miniCssF(e),i=s.p+a;if(function(e,t){for(var r=document.getElementsByTagName("link"),a=0;a<r.length;a++){var i=(o=r[a]).getAttribute("data-href")||o.getAttribute("href");if("stylesheet"===o.rel&&(i===e||i===t))return o}var n=document.getElementsByTagName("style");for(a=0;a<n.length;a++){var o;if((i=(o=n[a]).getAttribute("data-href"))===e||i===t)return o}}(a,i))return t();!function(e,t,r,a){var i=document.createElement("link");i.rel="stylesheet",i.type="text/css",i.onerror=i.onload=function(n){if(i.onerror=i.onload=null,"load"===n.type)r();else{var o=n&&("load"===n.type?"missing":n.type),c=n&&n.target&&n.target.href||t,f=new Error("Loading CSS chunk "+e+" failed.\n("+c+")");f.code="CSS_CHUNK_LOAD_FAILED",f.type=o,f.request=c,i.parentNode.removeChild(i),a(f)}},i.href=t,0!==i.href.indexOf(window.location.origin+"/")&&(i.crossOrigin="anonymous"),document.head.appendChild(i)}(e,i,t,r)}))},o={6272:0},s.f.miniCss=function(e,t){o[e]?t.push(o[e]):0!==o[e]&&{572:1,712:1,1225:1,1261:1,1687:1,1826:1,2426:1,3180:1,3487:1,4117:1,4778:1,4815:1,5400:1,5655:1,5912:1,6226:1,6242:1,6772:1,6827:1,6874:1,7056:1,7281:1,7425:1,7542:1,8079:1,8138:1,8426:1,8528:1,8829:1,9107:1,9624:1,9768:1,9844:1}[e]&&t.push(o[e]=n(e).then((function(){o[e]=0}),(function(t){throw delete o[e],t})))},function(){var e={6272:0};s.f.j=function(t,r){var a=s.o(e,t)?e[t]:void 0;if(0!==a)if(a)r.push(a[2]);else if(/^(2426|6272|9107)$/.test(t))e[t]=0;else{var i=new Promise((function(r,i){a=e[t]=[r,i]}));r.push(a[2]=i);var n=s.p+s.u(t),o=new Error;s.l(n,(function(r){if(s.o(e,t)&&(0!==(a=e[t])&&(e[t]=void 0),a)){var i=r&&("load"===r.type?"missing":r.type),n=r&&r.target&&r.target.src;o.message="Loading chunk "+t+" failed.\n("+i+": "+n+")",o.name="ChunkLoadError",o.type=i,o.request=n,a[1](o)}}),"chunk-"+t,t)}},s.O.j=function(t){return 0===e[t]};var t=function(t,r){var a,i,n=r[0],o=r[1],c=r[2],f=0;if(n.some((function(t){return 0!==e[t]}))){for(a in o)s.o(o,a)&&(s.m[a]=o[a]);if(c)var d=c(s)}for(t&&t(r);f<n.length;f++)i=n[f],s.o(e,i)&&e[i]&&e[i][0](),e[i]=0;return s.O(d)},r=self.__LOADABLE_LOADED_CHUNKS__=self.__LOADABLE_LOADED_CHUNKS__||[];r.forEach(t.bind(null,0)),r.push=t.bind(null,r.push.bind(r))}()}();</script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-arco.a74ce714.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-polyfill.e32e6eae.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-lodash.66bf9186.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-react.fc946b50.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-router.91ad9bff.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/4316.6db4945c.js" crossorigin="anonymous"></script><script defer="defer" src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/main.d3e44856.js" crossorigin="anonymous"></script><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/lib-arco.2bbbdcb8.css" rel="stylesheet" crossorigin="anonymous"><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/4316.27e0ba1f.css" rel="stylesheet" crossorigin="anonymous"><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/main.ff1f8027.css" rel="stylesheet" crossorigin="anonymous"><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/lib-arco.2bbbdcb8.css" rel="stylesheet" /><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/async/2426.01a02ce7.css" rel="stylesheet" /><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/async/9107.c4cafd96.css" rel="stylesheet" /><link href="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/css/async/articles/[id]/page.68a000e0.css" rel="stylesheet" /> <meta data-react-helmet="true" name="keywords" content="移动开发、大模型、数据库、算法"/><meta data-react-helmet="true" name="description" content="引言简介方法Agent 框架Agent 交互探索阶段部署阶段文档生成高级功能实验结果总结局限性未来工作引言 =======大语言模型(LLM)如 ChatGPT 和 GPT-4 显著提升了自然语言处理能力,并且推动了智能体在自主决策中的应用。最初,这些智能体专为基于文本的交互方式设计,展现了卓越的表现,包括记忆自适应性和多任务处理能力。然而,现实世界的应用程序不仅仅局限于文本输入,还涉及视觉和其他"/> <script data-react-helmet="true" type="application/ld+json">{"@context":"https://schema.org","@type":"Article","headline":"Agent系列:AppAgent v2-屏幕智能Agent(详解版)","image":["https://p3-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/8d019bcbb7244532a63a8702460e7135~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=YKf7UVsRNW22T642hIz1rCbo7kI%3D"],"datePublished":"2024-11-12T10:43:56.000Z","dateModified":"2024-11-12T11:00:52.000Z","author":[{"@type":"Person","name":"小窗幽记机器学习","url":"https://developer.volcengine.com/user/1977369294279187"}]}</script> </head><body><noscript>We're sorry but react app doesn't work properly without JavaScript enabled. Please enable it to continue.</noscript><div id="root"><div class="WFpVe NVl6r"><div class="yYbKk"><div class="gcUkQ"><a href="https://www.volcengine.com" target="_blank"><img src="https://lf1-cdn-tos.bytegoofy.com/goofy/tech-fe/assets/vocl_logo_dark.c678a292.svg" draggable="false"/></a></div><div class="rpPns"><a href="https://developer.volcengine.com" target="_blank"><img src="https://lf1-cdn-tos.bytegoofy.com/goofy/tech-fe/assets/community-logo-text-icon.4b9f2115.svg" draggable="false"/></a></div><div class="Q44pY"><div class="yYwDU"><div class="arco-input-group-wrapper arco-input-group-wrapper-default PqPNy"><span class="arco-input-group"><span class="arco-input-inner-wrapper arco-input-inner-wrapper-has-prefix arco-input-inner-wrapper-default"><span class="arco-input-group-prefix"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" class="arco-icon arco-icon-search"><path d="M33.072 33.071c6.248-6.248 6.248-16.379 0-22.627-6.249-6.249-16.38-6.249-22.628 0-6.248 6.248-6.248 16.379 0 22.627 6.248 6.248 16.38 6.248 22.628 0Zm0 0 8.485 8.485"></path></svg></span><input placeholder="在社区搜文章/找答案" class="arco-input arco-input-size-default" value=""/></span></span></div><a href="https://www.volcengine.com/docs" target="blank">文档</a><a href="https://www.volcengine.com/beian" target="blank">备案</a><a href="https://console.volcengine.com/home" target="blank">控制台</a><a href="https://console.volcengine.com/auth/login?redirectURI=https%3A%2F%2Fdeveloper.volcengine.com%2Farticles%2F7436341741323550771">登录</a><a href="https://console.volcengine.com/auth/signup?redirectURI=https%3A%2F%2Fdeveloper.volcengine.com%2Farticles%2F7436341741323550771" class="arco-btn arco-btn-primary arco-btn-size-default arco-btn-shape-square arco-btn-link iqsRJ"><span>立即注册</span></a></div></div></div><div class="stbmn"><div class="ASi46"><div class="qZzL_"><a href="/" target="_blank" class="">首页</a><a href="/articles" target="_blank" class="">文章</a><a href="/questions" target="_blank" class="hide">问答</a><a href="/videos" target="_blank" class="hide">视频</a><a href="/activities" target="_blank" class="">活动</a><a href="/resource" target="_blank" class="hide">下载资源</a><a href="/teams" target="_blank" class="hide">团队号</a><a href="/mirror" target="_blank" class="">镜像站</a></div></div><div class="NugHr"><div class="CQeSf">发布</div></div></div></div><div class="gSJtT"><div class="NabeC" id="main"><h1 id="article_title" class="BeWCA">Agent系列:AppAgent v2-屏幕智能Agent(详解版)</h1><div class="NSxaQ"><a href="/user/1977369294279187" target="_blank"><div style="width:20px;height:20px;font-size:10px" class="arco-avatar arco-avatar-circle"><span class="arco-avatar-image"><img src="https://p26-passport.byteacctimg.com/img/user-avatar/1816ccc0ebc35f4a89457e0b4ffe692b~300x300.image" alt="小窗幽记机器学习"/></span></div><div class="I2c0Q"><span>小窗幽记机器学习</span></div></a><div class="vCvrn"></div><div class="category"><a target="_blank" href="/articles?category=5">移动开发</a></div></div><div class="PN9l6"><span>大模型</span><span>数据库</span><span>算法</span></div><div class="uKIrB"><img src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/image/rec-product.424eb1f0.png"/></div><article><div class="markdown-body"><ul> <li>引言</li> <li>简介</li> <li>方法</li> </ul> <ul> <li>Agent 框架</li> <li>Agent 交互</li> <li>探索阶段</li> <li>部署阶段</li> <li>文档生成</li> <li>高级功能</li> </ul> <ul> <li>实验结果</li> <li>总结</li> <li>局限性</li> <li>未来工作</li> </ul> <ol> <li>引言 =======</li> </ol> <p>大语言模型(LLM)如 ChatGPT 和 GPT-4 显著提升了自然语言处理能力,并且推动了智能体在自主决策中的应用。最初,这些智能体专为基于文本的交互方式设计,展现了卓越的表现,包括记忆自适应性和多任务处理能力。然而,现实世界的应用程序不仅仅局限于文本输入,还涉及视觉和其他模态数据的处理,这对传统智能体提出了挑战。</p> <p>为应对这一需求,近年来涌现了多模态智能体,如 AssistGPT、Vipergpt 和 Visual ChatGPT 等,它们能够在复杂环境中执行多步骤推理、信息提取和集成,并根据用户输入作出适应性响应。此前的工作如 AppAgent v1 和 MobileAgent 通过模拟人类的交互方式,已经在智能手机应用程序中展示了较大的潜力。</p> <p>尽管如此,准确识别图形用户界面仍然是多模态智能体面临的关键挑战。依赖视觉特征的传统方法往往因识别模型的局限性而不够准确。而移动环境的动态特性和频繁的应用更新也进一步加剧了这一挑战。即便是如 GPT-4 这样的先进模型,在处理不熟悉的视觉元素时仍显不足,尤其是在应对小众应用时。</p> <p>为了解决这些问题,AppAgent v1 通过自动化探索和观看演示的方式增强了智能体的决策能力,但其过于依赖现成的解析器,限制了智能体在处理不标准界面元素时的灵活性。为此,AppAgent v2 提出了一种新的多模态智能体框架,扩展了动作空间,使其能够更广泛地适应不同类型的应用程序和界面元素。</p> <ol> <li>简介 =======</li> </ol> <p>随着多模态大语言模型(MLLM)的发展,视觉智能体(Agent)正逐渐在软件界面中,尤其是图形用户界面(GUI)中,发挥重要作用。本文设计了一个专为移动设备打造的智能体框架 <strong>AppAgent v2</strong> ,该框架基于大语言模型,能够在移动设备上导航并模拟用户交互行为。AppAgent v2 通过构建灵活的动作空间,增强了其在不同应用程序中的适应性,支持解析器、文本和视觉描述的结合。智能体的工作分为两个阶段:探索(Exploration)和部署(Deployment)。</p> <p>在探索阶段,智能体通过自动或手动的方式记录用户界面各元素的功能,并将其存储在结构化的知识库中。部署阶段,智能体通过 RAG(检索增强生成)技术从知识库中检索信息并实时更新,使其能够准确高效地完成任务,包括跨多个应用程序执行复杂的多步骤操作。实验结果表明,AppAgent v2 在各种基准测试中的表现优异,验证了其在实际场景中的有效性。AppAgent v2官方表示代码将很快开源。更多关于端侧大模型、Agent及其实践,欢迎关注微信公众号《小窗幽记机器学习》:</p> <ol start="2"> <li>方法 =======</li> </ol> <p>AppAgent v2 的多模态智能体框架分为两个阶段:<strong>探索</strong> 和<strong>部署</strong> 。在探索阶段,智能体通过自主或手动方式分析用户界面,记录 UI 元素及其功能,形成知识库。在部署阶段,智能体利用知识库中的信息,根据任务要求进行操作。</p> <p><img src="https://p6-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/9e671c8173a941fcb7fc34ea1485332f~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=mxS%2FTrjqAn1PS4Yg2B5oouQBFt0%3D" alt="picture.image"></p> <p>图1:appAgent 流程。探索模块采用Agent驱动或手动探索,将元素信息收集到文档中。部署阶段使用RAG实时检索和更新文档,从而快速准备执行任务</p> <h2>2.1 Agent 框架</h2> <p>AppAgent v2 在 Android 15 环境下使用 Android Studio 模拟器实现。智能体通过调用 AndroidController 发送指令与手机交互,结合 OCR 和检测模型,从屏幕截图中提取 UI 元素的信息(如 Android ID、文本、坐标等),并将其用于任务推断和决策。</p> <h2>2.2 Agent 交互</h2> <p>智能体通过一套标准化的指令与 Android 系统交互,指令包括点击、输入文本、滑动、返回、长按等操作。这些指令被转换为 Android 系统可执行的命令,确保与界面的交互精确无误。</p> <h2>2.3 探索阶段</h2> <p>在探索阶段,智能体通过两种方式分析用户界面:智能体驱动的探索和手动探索。</p> <ul> <li><strong>智能体驱动的探索</strong> :智能体自动识别需要交互的 UI 元素并执行操作,记录交互前后的变化,并根据结果反思操作的有效性。</li> <li><strong>手动探索</strong> :当智能体对某些界面的理解有限时,引入手动探索,GPT-4 通过观察人类的操作</li> </ul> <p>记录 UI 元素的变化和任务流程,帮助智能体在未来的任务中更好地理解和应对复杂的界面和操作。</p> <h2>2.4 部署阶段</h2> <p>在部署阶段,智能体利用自查检索器(Self-query Retriever)从知识库中检索相关文档,结合当前的 GUI 信息和任务要求,执行操作。智能体可以根据任务的需要动态更新知识库,确保在不同应用场景中快速适应。通过不断更新历史信息和操作结果,智能体能够优化其决策过程,直到任务完成。</p> <h2>2.5 文档生成</h2> <p>探索阶段收集的 UI 元素信息被存储在结构化的知识库中。该知识库包括元素的 Android ID、可见标签、文本内容、视觉特征(如颜色、形状)及其屏幕坐标等。知识库支持实时更新和动态检索,确保智能体可以随着任务的变化不断调整其操作策略。</p> <h2>2.6 高级功能</h2> <p>AppAgent v2 还具有以下增强功能:</p> <ul> <li><strong>视觉特征决策</strong> :当无法通过标准解析器识别 UI 元素时,智能体会转向基于视觉特征的决策,利用 OCR 技术识别界面中的文本和图标,确保在面对全新界面时仍能有效操作。</li> <li><strong>安全检查</strong> :智能体在执行涉及敏感信息的操作(如账户密码或支付)时,会自动切换到手动模式,确保用户隐私不受侵犯。</li> <li><strong>跨应用任务管理</strong> :智能体能够在多个应用程序之间切换,执行复杂的跨应用任务,如从一个应用中提取信息并在另一个应用中执行操作。</li> </ul> <ol start="3"> <li>实验 =======</li> </ol> <p>为了评估 AppAgent v2 的性能,本文在 Android 平台上进行了一系列实验,涵盖多个基准测试和用户研究。</p> <h4></h4> <p>3.1 定量结果</p> <p>AppAgent v2 在三个基准测试中进行了评估:<strong>DroidTask</strong> 、<strong>AppAgent v1 基准</strong> 和 <strong>Mobile-Eval</strong> 。</p> <ul> <li><strong>DroidTask</strong> :在 13 个流行应用程序衍生出的 158 个任务中,AppAgent v2 在完成率上表现优异,显著超过了 GPT-4 和带有记忆组件的 AutoDroid,证明了其强大的任务执行能力。</li> </ul> <p><img src="https://p6-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/9cffaa4e217347f4a47893e8d33a06be~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=7%2F7O2pEHuc%2FKxihLL%2FsoAxsKXks%3D" alt="picture.image"></p> <ul> <li><strong>AppAgent v1 基准</strong> :在 Twitter、Telegram 和其他常用应用的测试中,AppAgent v2 通过探索和部署阶段的结合,展示了其在多种任务和界面中的适应能力,成功率大幅提升。</li> <li><strong>Mobile-Eval</strong> :在 10 个常用移动应用程序中,AppAgent v2 在成功率、过程分数、相对效率和完成率四个指标上均表现出色,验证了其高效、准确地执行复杂任务的能力。</li> </ul> <p><img src="https://p6-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/d1ede5cf43df4260bcfbfb21f1668407~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=TkWFTgTWVlnPIraQhdk6zVYa5ME%3D" alt="picture.image"></p> <p><img src="https://p6-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/fa6d96ccb71f4ab88c2bcfc11831f145~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=lZQB2tpRCUy11LB%2BpmeYghpEqA8%3D" alt="picture.image"></p> <p>表2: MobileAgent和appAgent在Mobile-Eval上的指标对比结果</p> <h2>3.2 用户研究</h2> <p>为了进一步展示 AppAgent v2 的实际应用效果,本文进行了用户研究,模拟了复杂的跨应用任务。结果表明,AppAgent v2 能够准确执行长时间、多步骤的任务,展现了其在实际场景中处理复杂任务的能力和灵活性。</p> <p><img src="https://p6-volc-community-sign.byteimg.com/tos-cn-i-tlddhu82om/faa3d2b154f44c8b9e276759bf03e9e8~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=%2BI%2BtShTHv5rxFD4KOC%2BfHmhcozg%3D" alt="picture.image"></p> <p>图3:跨app应用上的定性结果</p> <h2>3.3 UI 界面解析分析</h2> <p>AppAgent v2 结合了结构化数据和视觉特征两种方法解析 UI 界面。对于常规应用,智能体依赖从 Android 系统解析的 XML 数据,而对于自定义开发的应用,智能体则通过 OCR 和视觉特征识别图标和文本。该双重策略显著提高了智能体的适应性和任务执行的可靠性。</p> <ol start="4"> <li>总结 =======</li> </ol> <p>本文提出了一个多模态智能体框架 AppAgent v2,显著提升了智能体在移动设备上的交互能力。通过整合解析器和视觉特征,AppAgent v2 构建了一个灵活的动作空间,能够适应不同类型的应用程序界面,并通过探索和部署两个阶段有效管理移动设备的动态环境。实验证明了该框架在图形用户界面识别和复杂任务执行方面的优越性,尤其是在跨应用任务和新环境中的适应能力。</p> <p>AppAgent v2 的主要贡献包括:</p> <ul> <li><strong>多模态智能体框架的引入</strong> :结合解析器和视觉特征,构建了灵活的动作空间,增强了智能体与图形用户界面的交互能力,提高了其在新环境中的适应性。</li> <li><strong>结构化存储格式的开发</strong> :结合 RAG 技术,实现了对知识库的实时更新和访问,增强了智能体的决策精度和任务执行效率。</li> <li><strong>广泛的实验证明</strong> :通过多个基准测试和用户研究,验证了智能体在各种移动应用程序中的有效性,展示了其适应性、用户友好性和操作效率。</li> </ul> <p>AppAgent v2 在现有技术基础上进一步提升了智能体在复杂移动环境中的表现,为未来的多模态智能体研究提供了重要的参考和改进方向。</p> <ol start="5"> <li>局限性 ========</li> </ol> <p>尽管 AppAgent v2 在多个应用场景中表现优异,但仍存在一些局限性:</p> <ul> <li><strong>数字标签识别的局限性</strong> :AppAgent v2 依赖智能体识别 UI 上的数字标签来确定特定元素。当 UI 元素本身包含数字时,可能会导致混淆。这种问题可以通过预先的手动探索和记录来缓解。</li> <li><strong>隐藏 UI 元素的交互难题</strong> :智能体在处理一些隐藏 UI 元素(如加速视频播放的按钮)时,由于缺乏先验知识,可能无法检测到相关元素,进而限制了其执行某些操作的能力。未来的工作将侧重于增强 UI 识别能力,并结合先验知识来解决这些问题。</li> </ul> <ol start="6"> <li>未来工作 =========</li> </ol> <p>未来的研究工作将集中在以下几个方面:</p> <ul> <li><strong>增强跨应用任务处理能力</strong> :进一步优化智能体在多个应用程序之间切换时的决策流程,提升其处理复杂跨应用任务的效率。</li> <li><strong>优化视觉特征识别</strong> :改进视觉特征的识别与操作,特别是在面对复杂或定制化的应用界面时,确保智能体始终能够做出正确的交互决策。</li> <li><strong>提高智能体的用户体验</strong> :通过优化智能体的反馈机制和任务执行过程,进一步提升用户体验,使其更加直观和高效。</li> </ul> <p>总之,AppAgent v2 为多模态智能体的发展提供了坚实的基础,并为未来的移动设备智能体研究指明了方向。</p></div></article><div class="pNjfy"><div class="lugd4"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" style="font-size:16px" class="arco-icon arco-icon-eye"><path d="M24 37c6.627 0 12.627-4.333 18-13-5.373-8.667-11.373-13-18-13-6.627 0-12.627 4.333-18 13 5.373 8.667 11.373 13 18 13Z" clip-rule="evenodd"></path><path d="M29 24a5 5 0 1 1-10 0 5 5 0 0 1 10 0Z"></path></svg><span class="MRrY3">0</span></div><div class="lugd4"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" style="font-size:16px" class="arco-icon arco-icon-thumb-up"><path d="M7 17v26m35.17-21.394-5.948 18.697a1 1 0 0 1-.953.697H14V19h3l9.403-12.223a1 1 0 0 1 1.386-.196l2.535 1.87a6 6 0 0 1 2.044 6.974L31 19h9.265a2 2 0 0 1 1.906 2.606Z"></path></svg><span class="MRrY3">0</span></div><div class="lugd4"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" style="font-size:16px" class="arco-icon arco-icon-message"><path d="M15 20h18m-18 9h9M7 41h17.63C33.67 41 41 33.67 41 24.63V24c0-9.389-7.611-17-17-17S7 14.611 7 24v17Z"></path></svg><span class="MRrY3">0</span></div><div class="lugd4"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" style="font-size:16px" class="arco-icon arco-icon-star"><path d="M22.552 6.908a.5.5 0 0 1 .896 0l5.02 10.17a.5.5 0 0 0 .376.274l11.224 1.631a.5.5 0 0 1 .277.853l-8.122 7.916a.5.5 0 0 0-.143.443l1.917 11.178a.5.5 0 0 1-.726.527l-10.038-5.278a.5.5 0 0 0-.466 0L12.73 39.9a.5.5 0 0 1-.726-.527l1.918-11.178a.5.5 0 0 0-.144-.443l-8.122-7.916a.5.5 0 0 1 .278-.853l11.223-1.63a.5.5 0 0 0 .376-.274l5.02-10.17Z"></path></svg><span class="MRrY3">0</span></div></div><div class="M6be4 article-action-warp"><div class="NEq5K"><button class="arco-btn arco-btn-secondary arco-btn-size-default arco-btn-shape-square" type="button"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" class="arco-icon arco-icon-thumb-up"><path d="M7 17v26m35.17-21.394-5.948 18.697a1 1 0 0 1-.953.697H14V19h3l9.403-12.223a1 1 0 0 1 1.386-.196l2.535 1.87a6 6 0 0 1 2.044 6.974L31 19h9.265a2 2 0 0 1 1.906 2.606Z"></path></svg><span>点赞</span></button><button class="arco-btn arco-btn-secondary arco-btn-size-default arco-btn-shape-square" type="button"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" class="arco-icon arco-icon-message"><path d="M15 20h18m-18 9h9M7 41h17.63C33.67 41 41 33.67 41 24.63V24c0-9.389-7.611-17-17-17S7 14.611 7 24v17Z"></path></svg><span>评论</span></button><button class="arco-btn arco-btn-secondary arco-btn-size-default arco-btn-shape-square" type="button"><svg fill="none" stroke="currentColor" stroke-width="4" viewBox="0 0 48 48" aria-hidden="true" focusable="false" class="arco-icon arco-icon-star"><path d="M22.552 6.908a.5.5 0 0 1 .896 0l5.02 10.17a.5.5 0 0 0 .376.274l11.224 1.631a.5.5 0 0 1 .277.853l-8.122 7.916a.5.5 0 0 0-.143.443l1.917 11.178a.5.5 0 0 1-.726.527l-10.038-5.278a.5.5 0 0 0-.466 0L12.73 39.9a.5.5 0 0 1-.726-.527l1.918-11.178a.5.5 0 0 0-.144-.443l-8.122-7.916a.5.5 0 0 1 .278-.853l11.223-1.63a.5.5 0 0 0 .376-.274l5.02-10.17Z"></path></svg><span>收藏</span></button></div></div></div><div class="ZZhEV"><div class="sgjEZ"><div class="jeBJ7">关于作者</div><div class="NSxaQ"><a href="/user/1977369294279187" target="_blank"><div style="width:48px;height:48px;font-size:24px" class="arco-avatar arco-avatar-circle"><span class="arco-avatar-image"><img src="https://p26-passport.byteacctimg.com/img/user-avatar/1816ccc0ebc35f4a89457e0b4ffe692b~300x300.image" alt="小窗幽记机器学习"/></span></div></a><div class="I2c0Q"><a href="/user/1977369294279187" target="_blank">小窗幽记机器学习</a></div></div><div></div></div><div class="eDNOE"><div class="ZotsP">相关产品</div></div><div class="Ex5bA"><div class="ZotsP">推荐阅读</div><a class="gZFRc" href="/articles/7419482941090168870" target="_blank"><div class="yEtsM">Bun 1.0,新的 JavaScript 运行时,内置了打包器、转译器、任务运行器和 np</div><div class="TNweF"></div></a><a class="gZFRc" href="/articles/7418740225456734227" target="_blank"><div class="yEtsM">使用 KMP 构建 Compose Multiplatform 共享 UI</div><div class="TNweF"></div></a><a class="gZFRc" href="/articles/7437331886491648026" target="_blank"><div class="yEtsM">无需网络,轻松在手机上离线布署本地大模型</div><div class="TNweF"></div></a><a class="gZFRc" href="/articles/7418398064655728681" target="_blank"><div class="yEtsM">Kotlin新技巧: combine 函数组合多个 Flow</div><div class="TNweF"></div></a><a class="gZFRc" href="/articles/7419128332123373622" target="_blank"><div class="yEtsM">在 Jetpack Compose 中实现 Android 15 的预测性返回</div><div class="TNweF"></div></a></div></div><div id="comment-panel" class="cKKVV"><div class="oTXbe"><div class="oBFjQ">评论</div><div class="arco-comment Ly94D"><div class="arco-comment-avatar"><div style="width:32px;height:32px;font-size:16px" class="arco-avatar arco-avatar-circle"><span class="arco-avatar-text">未登录</span></div></div><div class="arco-comment-inner"><div class="arco-comment-inner-content"><div class="arco-comment-content"><div class="jAxwI"><textarea class="arco-textarea" placeholder="输入评论"></textarea><div class="lFZfU">看完啦,<span>登录</span>分享一下感受吧~</div></div></div><div class="arco-comment-actions arco-comment-actions-align-right"></div></div></div></div><div class="iJPOx"></div><div class="qks6h">暂无评论</div></div></div></div></div><script id="__LOADABLE_REQUIRED_CHUNKS__" type="application/json">[2649,2986,7991,8626,3505,7469,9246,9902,6874]</script><script id="__LOADABLE_REQUIRED_CHUNKS___ext" type="application/json">{"namedChunks":["articles/[id]/page"]}</script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-arco.a74ce714.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-polyfill.e32e6eae.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/lib-lodash.66bf9186.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/8626.72051b87.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/3505.fc596d93.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/7469.24783622.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/9246.633f6657.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/9902.ab79fd01.js" ></script><script src="//portal.volccdn.com/obj/volcfe-scm/deploy/volc_developer_/42325/static/js/async/articles/[id]/page.68d574c4.js" ></script><script>window._SSR_DATA = {"data":{},"context":{"request":{"params":{},"query":{},"pathname":"\u002Farticles\u002F7436341741323550771","host":"developer.volcengine.com","url":"https:\u002F\u002Fdeveloper.volcengine.com\u002Farticles\u002F7436341741323550771"}},"renderLevel":2}</script> <script>window._ROUTER_DATA = {"loaderData":{"layout":{"err_no":401,"err_msg":"NotAuthorized"},"articles\u002Flayout":null,"articles\u002F[id]\u002Flayout":null,"articles\u002F[id]\u002Fpage":{"article":{"data":{"content":{"item_id":"7436341741323550771","item_type":2,"app_id":3569,"user_id":"1977369294279187","version":1,"status":2,"create_time":1731408236,"update_time":1731409252,"publish_time":1731408236,"name":"Agent系列:AppAgent v2-屏幕智能Agent(详解版)","abstract":"引言简介方法Agent 框架Agent 交互探索阶段部署阶段文档生成高级功能实验结果总结局限性未来工作引言\n=======大语言模型(LLM)如 ChatGPT 和 GPT-4 显著提升了自然语言处理能力,并且推动了智能体在自主决策中的应用。最初,这些智能体专为基于文本的交互方式设计,展现了卓越的表现,包括记忆自适应性和多任务处理能力。然而,现实世界的应用程序不仅仅局限于文本输入,还涉及视觉和其他","cover_image":{"key":"","url":"https:\u002F\u002Fp3-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002F8d019bcbb7244532a63a8702460e7135~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=YKf7UVsRNW22T642hIz1rCbo7kI%3D","size":0,"mime_type":"","rid":""},"mime_type":"","content":"\u003Cul\u003E\n\u003Cli\u003E引言\u003C\u002Fli\u003E\n\u003Cli\u003E简介\u003C\u002Fli\u003E\n\u003Cli\u003E方法\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cul\u003E\n\u003Cli\u003EAgent 框架\u003C\u002Fli\u003E\n\u003Cli\u003EAgent 交互\u003C\u002Fli\u003E\n\u003Cli\u003E探索阶段\u003C\u002Fli\u003E\n\u003Cli\u003E部署阶段\u003C\u002Fli\u003E\n\u003Cli\u003E文档生成\u003C\u002Fli\u003E\n\u003Cli\u003E高级功能\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cul\u003E\n\u003Cli\u003E实验结果\u003C\u002Fli\u003E\n\u003Cli\u003E总结\u003C\u002Fli\u003E\n\u003Cli\u003E局限性\u003C\u002Fli\u003E\n\u003Cli\u003E未来工作\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Col\u003E\n\u003Cli\u003E引言\n=======\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E大语言模型(LLM)如 ChatGPT 和 GPT-4 显著提升了自然语言处理能力,并且推动了智能体在自主决策中的应用。最初,这些智能体专为基于文本的交互方式设计,展现了卓越的表现,包括记忆自适应性和多任务处理能力。然而,现实世界的应用程序不仅仅局限于文本输入,还涉及视觉和其他模态数据的处理,这对传统智能体提出了挑战。\u003C\u002Fp\u003E\n\u003Cp\u003E为应对这一需求,近年来涌现了多模态智能体,如 AssistGPT、Vipergpt 和 Visual ChatGPT 等,它们能够在复杂环境中执行多步骤推理、信息提取和集成,并根据用户输入作出适应性响应。此前的工作如 AppAgent v1 和 MobileAgent 通过模拟人类的交互方式,已经在智能手机应用程序中展示了较大的潜力。\u003C\u002Fp\u003E\n\u003Cp\u003E尽管如此,准确识别图形用户界面仍然是多模态智能体面临的关键挑战。依赖视觉特征的传统方法往往因识别模型的局限性而不够准确。而移动环境的动态特性和频繁的应用更新也进一步加剧了这一挑战。即便是如 GPT-4 这样的先进模型,在处理不熟悉的视觉元素时仍显不足,尤其是在应对小众应用时。\u003C\u002Fp\u003E\n\u003Cp\u003E为了解决这些问题,AppAgent v1 通过自动化探索和观看演示的方式增强了智能体的决策能力,但其过于依赖现成的解析器,限制了智能体在处理不标准界面元素时的灵活性。为此,AppAgent v2 提出了一种新的多模态智能体框架,扩展了动作空间,使其能够更广泛地适应不同类型的应用程序和界面元素。\u003C\u002Fp\u003E\n\u003Col\u003E\n\u003Cli\u003E简介\n=======\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E随着多模态大语言模型(MLLM)的发展,视觉智能体(Agent)正逐渐在软件界面中,尤其是图形用户界面(GUI)中,发挥重要作用。本文设计了一个专为移动设备打造的智能体框架 \u003Cstrong\u003EAppAgent v2\u003C\u002Fstrong\u003E ,该框架基于大语言模型,能够在移动设备上导航并模拟用户交互行为。AppAgent v2 通过构建灵活的动作空间,增强了其在不同应用程序中的适应性,支持解析器、文本和视觉描述的结合。智能体的工作分为两个阶段:探索(Exploration)和部署(Deployment)。\u003C\u002Fp\u003E\n\u003Cp\u003E在探索阶段,智能体通过自动或手动的方式记录用户界面各元素的功能,并将其存储在结构化的知识库中。部署阶段,智能体通过 RAG(检索增强生成)技术从知识库中检索信息并实时更新,使其能够准确高效地完成任务,包括跨多个应用程序执行复杂的多步骤操作。实验结果表明,AppAgent v2 在各种基准测试中的表现优异,验证了其在实际场景中的有效性。AppAgent v2官方表示代码将很快开源。更多关于端侧大模型、Agent及其实践,欢迎关注微信公众号《小窗幽记机器学习》:\u003C\u002Fp\u003E\n\u003Col start=\"2\"\u003E\n\u003Cli\u003E方法\n=======\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003EAppAgent v2 的多模态智能体框架分为两个阶段:\u003Cstrong\u003E探索\u003C\u002Fstrong\u003E 和\u003Cstrong\u003E部署\u003C\u002Fstrong\u003E 。在探索阶段,智能体通过自主或手动方式分析用户界面,记录 UI 元素及其功能,形成知识库。在部署阶段,智能体利用知识库中的信息,根据任务要求进行操作。\u003C\u002Fp\u003E\n\u003Cp\u003E\u003Cimg src=\"https:\u002F\u002Fp6-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002F9e671c8173a941fcb7fc34ea1485332f~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=mxS%2FTrjqAn1PS4Yg2B5oouQBFt0%3D\" alt=\"picture.image\"\u003E\u003C\u002Fp\u003E\n\u003Cp\u003E图1:appAgent 流程。探索模块采用Agent驱动或手动探索,将元素信息收集到文档中。部署阶段使用RAG实时检索和更新文档,从而快速准备执行任务\u003C\u002Fp\u003E\n\u003Ch2\u003E2.1 Agent 框架\u003C\u002Fh2\u003E\n\u003Cp\u003EAppAgent v2 在 Android 15 环境下使用 Android Studio 模拟器实现。智能体通过调用 AndroidController 发送指令与手机交互,结合 OCR 和检测模型,从屏幕截图中提取 UI 元素的信息(如 Android ID、文本、坐标等),并将其用于任务推断和决策。\u003C\u002Fp\u003E\n\u003Ch2\u003E2.2 Agent 交互\u003C\u002Fh2\u003E\n\u003Cp\u003E智能体通过一套标准化的指令与 Android 系统交互,指令包括点击、输入文本、滑动、返回、长按等操作。这些指令被转换为 Android 系统可执行的命令,确保与界面的交互精确无误。\u003C\u002Fp\u003E\n\u003Ch2\u003E2.3 探索阶段\u003C\u002Fh2\u003E\n\u003Cp\u003E在探索阶段,智能体通过两种方式分析用户界面:智能体驱动的探索和手动探索。\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003E智能体驱动的探索\u003C\u002Fstrong\u003E :智能体自动识别需要交互的 UI 元素并执行操作,记录交互前后的变化,并根据结果反思操作的有效性。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E手动探索\u003C\u002Fstrong\u003E :当智能体对某些界面的理解有限时,引入手动探索,GPT-4 通过观察人类的操作\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cp\u003E记录 UI 元素的变化和任务流程,帮助智能体在未来的任务中更好地理解和应对复杂的界面和操作。\u003C\u002Fp\u003E\n\u003Ch2\u003E2.4 部署阶段\u003C\u002Fh2\u003E\n\u003Cp\u003E在部署阶段,智能体利用自查检索器(Self-query Retriever)从知识库中检索相关文档,结合当前的 GUI 信息和任务要求,执行操作。智能体可以根据任务的需要动态更新知识库,确保在不同应用场景中快速适应。通过不断更新历史信息和操作结果,智能体能够优化其决策过程,直到任务完成。\u003C\u002Fp\u003E\n\u003Ch2\u003E2.5 文档生成\u003C\u002Fh2\u003E\n\u003Cp\u003E探索阶段收集的 UI 元素信息被存储在结构化的知识库中。该知识库包括元素的 Android ID、可见标签、文本内容、视觉特征(如颜色、形状)及其屏幕坐标等。知识库支持实时更新和动态检索,确保智能体可以随着任务的变化不断调整其操作策略。\u003C\u002Fp\u003E\n\u003Ch2\u003E2.6 高级功能\u003C\u002Fh2\u003E\n\u003Cp\u003EAppAgent v2 还具有以下增强功能:\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003E视觉特征决策\u003C\u002Fstrong\u003E :当无法通过标准解析器识别 UI 元素时,智能体会转向基于视觉特征的决策,利用 OCR 技术识别界面中的文本和图标,确保在面对全新界面时仍能有效操作。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E安全检查\u003C\u002Fstrong\u003E :智能体在执行涉及敏感信息的操作(如账户密码或支付)时,会自动切换到手动模式,确保用户隐私不受侵犯。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E跨应用任务管理\u003C\u002Fstrong\u003E :智能体能够在多个应用程序之间切换,执行复杂的跨应用任务,如从一个应用中提取信息并在另一个应用中执行操作。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Col start=\"3\"\u003E\n\u003Cli\u003E实验\n=======\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E为了评估 AppAgent v2 的性能,本文在 Android 平台上进行了一系列实验,涵盖多个基准测试和用户研究。\u003C\u002Fp\u003E\n\u003Ch4\u003E\u003C\u002Fh4\u003E\n\u003Cp\u003E3.1 定量结果\u003C\u002Fp\u003E\n\u003Cp\u003EAppAgent v2 在三个基准测试中进行了评估:\u003Cstrong\u003EDroidTask\u003C\u002Fstrong\u003E 、\u003Cstrong\u003EAppAgent v1 基准\u003C\u002Fstrong\u003E 和 \u003Cstrong\u003EMobile-Eval\u003C\u002Fstrong\u003E 。\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003EDroidTask\u003C\u002Fstrong\u003E :在 13 个流行应用程序衍生出的 158 个任务中,AppAgent v2 在完成率上表现优异,显著超过了 GPT-4 和带有记忆组件的 AutoDroid,证明了其强大的任务执行能力。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cp\u003E\u003Cimg src=\"https:\u002F\u002Fp6-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002F9cffaa4e217347f4a47893e8d33a06be~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=7%2F7O2pEHuc%2FKxihLL%2FsoAxsKXks%3D\" alt=\"picture.image\"\u003E\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003EAppAgent v1 基准\u003C\u002Fstrong\u003E :在 Twitter、Telegram 和其他常用应用的测试中,AppAgent v2 通过探索和部署阶段的结合,展示了其在多种任务和界面中的适应能力,成功率大幅提升。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003EMobile-Eval\u003C\u002Fstrong\u003E :在 10 个常用移动应用程序中,AppAgent v2 在成功率、过程分数、相对效率和完成率四个指标上均表现出色,验证了其高效、准确地执行复杂任务的能力。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cp\u003E\u003Cimg src=\"https:\u002F\u002Fp6-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002Fd1ede5cf43df4260bcfbfb21f1668407~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=TkWFTgTWVlnPIraQhdk6zVYa5ME%3D\" alt=\"picture.image\"\u003E\u003C\u002Fp\u003E\n\u003Cp\u003E\u003Cimg src=\"https:\u002F\u002Fp6-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002Ffa6d96ccb71f4ab88c2bcfc11831f145~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=lZQB2tpRCUy11LB%2BpmeYghpEqA8%3D\" alt=\"picture.image\"\u003E\u003C\u002Fp\u003E\n\u003Cp\u003E表2: MobileAgent和appAgent在Mobile-Eval上的指标对比结果\u003C\u002Fp\u003E\n\u003Ch2\u003E3.2 用户研究\u003C\u002Fh2\u003E\n\u003Cp\u003E为了进一步展示 AppAgent v2 的实际应用效果,本文进行了用户研究,模拟了复杂的跨应用任务。结果表明,AppAgent v2 能够准确执行长时间、多步骤的任务,展现了其在实际场景中处理复杂任务的能力和灵活性。\u003C\u002Fp\u003E\n\u003Cp\u003E\u003Cimg src=\"https:\u002F\u002Fp6-volc-community-sign.byteimg.com\u002Ftos-cn-i-tlddhu82om\u002Ffaa3d2b154f44c8b9e276759bf03e9e8~tplv-tlddhu82om-image.image?=&rk3s=8031ce6d&x-expires=1732831354&x-signature=%2BI%2BtShTHv5rxFD4KOC%2BfHmhcozg%3D\" alt=\"picture.image\"\u003E\u003C\u002Fp\u003E\n\u003Cp\u003E图3:跨app应用上的定性结果\u003C\u002Fp\u003E\n\u003Ch2\u003E3.3 UI 界面解析分析\u003C\u002Fh2\u003E\n\u003Cp\u003EAppAgent v2 结合了结构化数据和视觉特征两种方法解析 UI 界面。对于常规应用,智能体依赖从 Android 系统解析的 XML 数据,而对于自定义开发的应用,智能体则通过 OCR 和视觉特征识别图标和文本。该双重策略显著提高了智能体的适应性和任务执行的可靠性。\u003C\u002Fp\u003E\n\u003Col start=\"4\"\u003E\n\u003Cli\u003E总结\n=======\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E本文提出了一个多模态智能体框架 AppAgent v2,显著提升了智能体在移动设备上的交互能力。通过整合解析器和视觉特征,AppAgent v2 构建了一个灵活的动作空间,能够适应不同类型的应用程序界面,并通过探索和部署两个阶段有效管理移动设备的动态环境。实验证明了该框架在图形用户界面识别和复杂任务执行方面的优越性,尤其是在跨应用任务和新环境中的适应能力。\u003C\u002Fp\u003E\n\u003Cp\u003EAppAgent v2 的主要贡献包括:\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003E多模态智能体框架的引入\u003C\u002Fstrong\u003E :结合解析器和视觉特征,构建了灵活的动作空间,增强了智能体与图形用户界面的交互能力,提高了其在新环境中的适应性。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E结构化存储格式的开发\u003C\u002Fstrong\u003E :结合 RAG 技术,实现了对知识库的实时更新和访问,增强了智能体的决策精度和任务执行效率。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E广泛的实验证明\u003C\u002Fstrong\u003E :通过多个基准测试和用户研究,验证了智能体在各种移动应用程序中的有效性,展示了其适应性、用户友好性和操作效率。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cp\u003EAppAgent v2 在现有技术基础上进一步提升了智能体在复杂移动环境中的表现,为未来的多模态智能体研究提供了重要的参考和改进方向。\u003C\u002Fp\u003E\n\u003Col start=\"5\"\u003E\n\u003Cli\u003E局限性\n========\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E尽管 AppAgent v2 在多个应用场景中表现优异,但仍存在一些局限性:\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003E数字标签识别的局限性\u003C\u002Fstrong\u003E :AppAgent v2 依赖智能体识别 UI 上的数字标签来确定特定元素。当 UI 元素本身包含数字时,可能会导致混淆。这种问题可以通过预先的手动探索和记录来缓解。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E隐藏 UI 元素的交互难题\u003C\u002Fstrong\u003E :智能体在处理一些隐藏 UI 元素(如加速视频播放的按钮)时,由于缺乏先验知识,可能无法检测到相关元素,进而限制了其执行某些操作的能力。未来的工作将侧重于增强 UI 识别能力,并结合先验知识来解决这些问题。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Col start=\"6\"\u003E\n\u003Cli\u003E未来工作\n=========\u003C\u002Fli\u003E\n\u003C\u002Fol\u003E\n\u003Cp\u003E未来的研究工作将集中在以下几个方面:\u003C\u002Fp\u003E\n\u003Cul\u003E\n\u003Cli\u003E\u003Cstrong\u003E增强跨应用任务处理能力\u003C\u002Fstrong\u003E :进一步优化智能体在多个应用程序之间切换时的决策流程,提升其处理复杂跨应用任务的效率。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E优化视觉特征识别\u003C\u002Fstrong\u003E :改进视觉特征的识别与操作,特别是在面对复杂或定制化的应用界面时,确保智能体始终能够做出正确的交互决策。\u003C\u002Fli\u003E\n\u003Cli\u003E\u003Cstrong\u003E提高智能体的用户体验\u003C\u002Fstrong\u003E :通过优化智能体的反馈机制和任务执行过程,进一步提升用户体验,使其更加直观和高效。\u003C\u002Fli\u003E\n\u003C\u002Ful\u003E\n\u003Cp\u003E总之,AppAgent v2 为多模态智能体的发展提供了坚实的基础,并为未来的移动设备智能体研究指明了方向。\u003C\u002Fp\u003E","resource":{},"extra":{"source":{"app_id":3569,"name":"","author":""},"html_content":""},"parent_id":"0","parent_type":0,"last_version":1,"last_status":2,"last_create_time":1731408236,"last_update_time":1731408236,"status_tags":[],"create_user_id":"0","item_source":"2","arcosite_id":"","text_url":""},"user":{"user_id":"1977369294279187","user_type":14,"name":"小窗幽记机器学习","avatar":{"key":"","url":"https:\u002F\u002Fp26-passport.byteacctimg.com\u002Fimg\u002Fuser-avatar\u002F1816ccc0ebc35f4a89457e0b4ffe692b~300x300.image","size":0,"mime_type":"","rid":""},"status":2},"categories":[{"category_id":"5","name":"移动开发","parent_id":"0","create_time":1637495002,"update_time":1637495249,"op_user_id":"0","abstract":"","cover_image":{"key":"","url":"","size":0,"mime_type":"","rid":""},"rank":0,"category_url":""}],"tags":[{"tag_id":"7307103161498632233","name":"大模型","create_time":1701317532,"update_time":1701917647,"op_user_id":"3945470841857399","create_user_id":"620546920289582","abstract":"","cover_image":{"key":"","url":"","size":0,"mime_type":"","rid":""},"status":2},{"tag_id":"7146114517833875464","name":"数据库","create_time":1663834443,"update_time":1663834443,"op_user_id":"743694848765533","create_user_id":"743694848765533","abstract":"","cover_image":{"key":"","url":"","size":0,"mime_type":"","rid":""},"status":2},{"tag_id":"7146205183884656670","name":"算法","create_time":1663855553,"update_time":1663855553,"op_user_id":"4015857634130296","create_user_id":"4015857634130296","abstract":"","cover_image":{"key":"","url":"","size":0,"mime_type":"","rid":""},"status":2}],"interact_status":{"is_like":false,"is_collect":false,"is_join":false,"like_count":0,"collect_count":0,"comment_count":0,"join_count":0,"evaluate_count":0,"evaluate_avg":0,"evaluate_self":0},"pool_status":{"is_top":false,"is_good":false,"is_recommend":false},"content_count":{"view_count":284,"child_count":0},"op_info":{"op_user":null,"op_time":0,"op_resource":""}},"err_msg":"","err_no":0},"recomConfig":{"err_no":0,"err_msg":"success","data":{"recom_result":[],"service_info_map":[{"name":"万有商城","data":[[{"title":"万有商城","item":[]}]]},{"name":"企业服务与云通信","data":[[{"title":"域名与网站","item":[{"title":"SSL证书","link":"www.volcengine.com\u002Fproduct\u002Fcertificate-center","des":"可信赖的数字证书全生命周期管理平台,提供SSL证书和私有CA服务","desShort":"可信赖的数字证书全生命周期管理平台","enTitle":"certificate_service"},{"title":"边缘云商标服务","link":"www.volcengine.com\u002Fproduct\u002Ftrademark","des":"智能注册,快速提交,便捷管理,即时保护您的品牌;资深顾问,专业方案,全程跟进,助您注册省心无忧","desShort":"专业、高效的商标注册管理服务平台","enTitle":"trademark"},{"title":"域名注册","link":"www.volcengine.com\u002Fproduct\u002Fdomain-service","des":"提供域名注册、转入、委托购买、管理等服务,可选择的域名后缀丰富","desShort":"便捷的域名注册、转入、委托购买、管理等服务","enTitle":"domain_service"},{"title":"云解析DNS","link":"www.volcengine.com\u002Fproduct\u002Ftrafficroute","des":"安全、可靠、快速的权威DNS、递归DNS、私网DNS及全局流量管理GTM服务","desShort":"全系列云解析DNS和全局流量管理GTM","enTitle":"dns"},{"title":"账户管理","link":"","des":"16字以上描述16字以上描述16字以上描述","desShort":"账户管理","enTitle":""},{"title":"全局流量管理","link":"","des":"火山引擎云调度(GTM)是基于 DNS 的流量路由服务。 假设您的服务部署在多个位于不同地理位置的 IDC 机房。每个 IDC 机房针对不同网络运营商有不同的接入 IP 地址。用户通过域名访问您的服务。 当分布在不同地理位置、使用不同网络运营商的用户访问您的域名时,您希望您的服务可以满足以下要求: 用户可以获取访问延时最低的可用 IP 地址。 用户请求可以根据每个 IDC 机房的容量合理分配,实现用户请求数量上的负载均衡。 某个 IDC 机房因发生故障而不可用时,用户请求不会被路由到该 IDC 机房。 云调度 GTM 可以通过以下功能满足这些要求: 降低用户访问延时:云调度 GTM 会根据用户的地理位置,把用户请求路由到延时最低的 IP 地址。 负载均衡:云调度 GTM 会根据您配置的权重、容量或负载信息,合理分配各个 IP 地址收到的用户请求数。 故障转移:云调度 GTM 会主动进行基于多种类型的定时健康检查,实时检测各个 IP 地址的可用性。用户请求不会被路由到不可用的 IP 地址。","desShort":"火山引擎云调度(GTM)是基于 DNS 的流量路由服务。","enTitle":"gtm"},{"title":"移动解析 HTTPDNS","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Ftrafficroute","des":"火山引擎移动解析(HTTPDNS)为互联网域名提供递归 DNS 服务。您的客户端应用在集成移动解析 HTTPDNS 提供的客户端 SDK 或 HTTP API 后,就可以绕过 Local DNS 服务器,通过 HTTP 或 HTTPS 协议向移动解析 HTTPDNS 服务端发送 DNS 查询请求。 相比基于 Local DNS 的传统 DNS 解析流程,移动解析 HTTPDNS 可以避免以下问题: 客户端应用访问的域名的 DNS 被劫持 客户端应用的 IP 地址和 DNS 查询返回的的 IP 地址分别属于不同的网络运营商或地理位置相距较远,导致应用的网络延时较高 Local DNS 服务器的 DNS 缓存更新不及时,导致客户端应用无法及时得到最新的 IP 地址 另外,您还可以基于网络运营商、地理位置或自定义标识对 HTTPDNS 服务端返回的 DNS 查询结果进行自定义配置。","desShort":"火山引擎移动解析(HTTPDNS)为互联网域名提供递归 DNS 服务。","enTitle":"httpdns"}]},{"title":"云通信","item":[{"title":"语音服务","link":"www.volcengine.com\u002Fproduct\u002Fcloud-vms","des":"为企业提供专业的语音服务,支持语音SIP、语音通知、语音验证码、隐私号等产品","desShort":"高质量、高并发的语音通信能力","enTitle":"vms"},{"title":"云通信 短信服务","link":"www.volcengine.com\u002Fproduct\u002Fcloud-sms","des":"提供安全可靠、便捷高效的全球短信服务,服务范围覆盖国内外230多个国家和地区,适用于推广短信、通知短信、验证码短信等多种场景,稳定触达客户","desShort":"安全可靠、便捷高效的全球短信服务","enTitle":"volcSMS"},{"title":"语音服务海外版","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Fcloud-vms","des":"为企业提供专业的语音服务,支持语音SIP、语音通知、语音验证码、隐私号等产品","desShort":"高质量、高并发的语音通信能力","enTitle":""}]},{"title":"教育科研","item":[{"title":"生信操作系统","link":"www.volcengine.com\u002Fproduct\u002Fbioos","des":"火山引擎生信操作系统 Bio-OS 是火山引擎提供的低门槛,低成本,极速的云上基因分析平台;基于Serverless 架构实现了标准生信工作流引擎和多种三级分析引擎支持,帮助生信科研人员、生物科技公司,农业育种公司快速构建基础能力","desShort":"低门槛、低成本的极速云上生信分析服务","enTitle":"bio"}]}]]},{"name":"人工智能与机器学习","data":[[{"title":"机器学习","item":[{"title":"机器学习平台","link":"www.volcengine.com\u002Fproduct\u002Fml-platform","des":"机器学习平台是面向机器学习应用开发者,提供 WebIDE 和自定义训练等丰富建模工具、多框架高性能模型推理服务的企业级云原生机器学习平台。","desShort":"企业级云原生机器学习平台","enTitle":"ml_platform"},{"title":"火山方舟","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Fark","des":"字节跳动推出的自研大模型。通过字节内部50+业务场景实践验证,每日千亿级tokens大使用量持续打磨,提供多模态能力,以优质模型效果为企业打造丰富的业务体验","desShort":"一站式大模型服务平台","enTitle":""},{"title":"量化科学计算平台","link":"","des":"企业级量化科学计算平台是面向量子化学研究员的高通量计算化学任务调度服务。","desShort":"企业级量化科学计算平台","enTitle":""}]},{"title":"智能体验与创作","item":[{"title":"Para云平台","link":"www.volcengine.com\u002Fproduct\u002FAR","des":"提供全场景沉浸式的AR互动体验,促进用户体验升级和商业价值转化提升,打造身临其境的交互体验","desShort":"全链路AR营销服务及平台","enTitle":"para"},{"title":"视觉智能","link":"www.volcengine.com\u002Fproduct\u002Fintelligent-interactive-effects","des":"提供图像视频美化和计算机视觉算法能力,支持移动端、PC端等跨平台使用","desShort":"多项智能图像渲染和视觉算法能力","enTitle":"cv"},{"title":"智能创作云","link":"www.volcengine.com\u002Fproduct\u002FICC","des":"为企业和媒体推出的智能内容生产平台","desShort":"企业级智能内容生产SaaS","enTitle":"ic_iam"},{"title":"test","link":"","des":"testttttttttttttt","desShort":"test","enTitle":""},{"title":"智能美化特效","link":"","des":"智能美化特效,目前仅包含基于大模型的ai音乐创作","desShort":"智能美化特效","enTitle":""}]},{"title":"智能应用","item":[{"title":"智能创意","link":"www.volcengine.com\u002Fproduct\u002FCMP","des":"企业抖音账号矩阵管理运营后台,通过丰富运营工具,数据分析,视频工具,提升运营质量","desShort":"企业抖音账号矩阵管理运营管家","enTitle":"CreativeManagementPlatform"}]},{"title":"智能语音交互","item":[{"title":"AVATAR","link":"www.volcengine.com\u002Fproduct\u002Favatar","des":"基于多模态交互体系打造全场景数字人产品方案,为企业提供拟人化服务","desShort":"全场景数字人产品方案","enTitle":"avatar"},{"title":"有声内容创作平台","link":"www.volcengine.com\u002Fproduct\u002Fvoice-tech","des":"基于语音识别、语音合成、自然语言理解等技术,适用于电话客服、阅读、会议纪要、视频字幕等场景,让您的产品能“听”会“说”","desShort":"语音合成、语音识别等AI语音服务","enTitle":"accp"},{"title":"智能外呼","link":"www.volcengine.com\u002Fproduct\u002Fsmart-outbound","des":"基于语音对话能力,结合外呼场景构建的交互真实,应答智能的外呼系统","desShort":"真实互动语音外呼机器人","enTitle":"smart-outbound"},{"title":"live_insight","link":"","des":"完整直播回放,支持查看逐字稿,直播内容标签化,可结构化查看内容分布。 - 支持自由视频切片及灵活内容检索,可以根据关键词、标签在场次内\u002F跨场次搜索符合条件的话术内容。","desShort":"直播洞察","enTitle":""},{"title":"直播洞察","link":"","des":"完整直播回放,支持查看逐字稿,直播内容标签化,可结构化查看内容分布。支持自由视频切片及灵活内容检索,可以根据关键词、标签在场次内\u002F跨场次搜索符合条件的话术内容。","desShort":"直播洞察","enTitle":""}]},{"title":"智能内容","item":[{"title":"内容定制API","link":"www.volcengine.com\u002Fproduct\u002Fcontent-customization","des":"帮助客户快速构建内容信息流的SaaS产品","desShort":"快速构建内容信息流的SaaS产品","enTitle":"content"},{"title":"内容管理平台","link":"www.volcengine.com\u002Fproduct\u002Fcms","des":"提供内容引入、质检、理解、分发、数据的全链路一站式管理工具,为企业内容精细化运营赋能","desShort":"提供一站式运营管理能力","enTitle":"mp"},{"title":"内容洞察平台","link":"","des":"内容管理平台依托字节跳动长期沉淀的先进的内容管理能力、质检能力、理解能力、运营能力、推荐能力、数据能力,为合作伙伴提供从内容引入、内容加工、内容分发、数据分析全链路一站式运营管理平台,为企业的内容精细化运营赋能","desShort":"提供一站式运营管理能力","enTitle":"insight"},{"title":"内容管理平台 Platform","link":"","des":"内容管理平台依托字节跳动长期沉淀的领先的内容管理能力、质检能力、理解能力、运营能力、推荐能力、数据能力,为合作伙伴提供从内容引入、内容加工、内容分发、数据分析全链路一站式运营管理平台,为企业的内容精细化运营赋能。","desShort":"内容管理平台以内容为驱动,从底层系统到上层客户端一站式赋能,为企业带来全链路业务增长","enTitle":""}]},{"title":"推荐算法","item":[{"title":"智能推荐","link":"www.volcengine.com\u002Fproduct\u002Frec","des":"为客户提供端到端的一站式推荐服务能力搭建平台","desShort":"一站式企业级推荐平台","enTitle":"air"}]},{"title":"内容安全与风控","item":[{"title":"业务风险识别","link":"www.volcengine.com\u002Fproduct\u002Fbusiness-security","des":"精准识别各类场景黑灰产风险,持续守护业务的健康增长","desShort":"精准识别各类场景黑灰产风险","enTitle":"BusinessSecurity"}]},{"title":"AI开放平台","item":[{"title":"机器翻译","link":"www.volcengine.com\u002Fproduct\u002Fmachine-translation","des":"本地化一站式解决方案,简化本地化流程,提高协作效率,助力产品出海转型","desShort":"通用机器翻译、视频翻译和智能同传","enTitle":"translate"},{"title":"智能视频分析平台","link":"www.volcengine.com\u002Fproduct\u002Fiva","des":"基于AI视频结构化技术,可对视频或图片中的人、车、物、行为事件等进行检测并识别的通用性视频分析平台","desShort":"涵盖人、车、事件的全自研AI视频算法","enTitle":"iva"}]}]]},{"name":"大数据","data":[[{"title":"数据中台","item":[{"title":"ByteHouse","link":"www.volcengine.com\u002Fproduct\u002Fbytehouse","des":"极致性能弹性伸缩的分析型数据库","desShort":"极致性能弹性伸缩的分析型数据库","enTitle":"bytehouse"},{"title":"全域数据集成","link":"www.volcengine.com\u002Fproduct\u002Fdatasail","des":"高速稳定的批量和流式数据同步引擎","desShort":"高速稳定的批量和流式数据同步引擎","enTitle":"datasail"},{"title":"E-MapReduce","link":"www.volcengine.com\u002Fproduct\u002Femr","des":"云原生开源大数据平台","desShort":"云原生开源大数据平台","enTitle":"emr"},{"title":"大数据研发治理套件","link":"www.volcengine.com\u002Fproduct\u002Fdataleap","des":"一站式数据中台套件","desShort":"一站式数据中台套件","enTitle":"dataleap"},{"title":"湖仓一体分析服务 LAS","link":"www.volcengine.com\u002Fproduct\u002Flas","des":"全托管一站式智能大数据分析服务","desShort":"全托管一站式智能大数据分析服务","enTitle":"las"},{"title":"流式计算 Flink版","link":"www.volcengine.com\u002Fproduct\u002Fflink","des":"云原生全托管实时数据处理分析服务","desShort":"云原生全托管实时数据处理分析服务","enTitle":"flink"},{"title":"智能数据洞察","link":"www.volcengine.com\u002Fproduct\u002Fdatawind","des":"一站式大量级数据展示与分析平台","desShort":"一站式大量级数据展示与分析平台","enTitle":"datawind"}]},{"title":"增长营销","item":[{"title":"数据安全中心-私有化平台","link":"www.volcengine.com\u002Fproduct\u002FveCDP","des":"面向业务增长的客户全域数据中台","desShort":"面向业务增长的客户全域数据中台","enTitle":"cdp"},{"title":"增长分析staging","link":"www.volcengine.com\u002Fproduct\u002Fdatafinder","des":"一站式用户分析与运营平台","desShort":"一站式用户分析与运营平台","enTitle":"clouduba_staging"},{"title":"智能广告营销平台","link":"www.volcengine.com\u002Fproduct\u002Fgmp","des":"全域营销触达与多策略管理平台","desShort":"全域营销触达与多策略管理平台","enTitle":"adopenapi"},{"title":"A\u002FB测试 DataTester","link":"www.volcengine.com\u002Fproduct\u002Fdatatester","des":"一站式A\u002FB测试与智能优化平台","desShort":"一站式A\u002FB测试与智能优化平台","enTitle":"datatester"},{"title":"增长分析","link":"","des":"一站式用户分析与运营平台 为企业提供数字化消费者行为分析洞见 发现业务的关键增长点,提升企业效益","desShort":"DataFinder 热门 一站式用户分析与运营平台","enTitle":""}]}]]},{"name":"CDN与边缘","data":[[{"title":"CDN与加速","item":[{"title":"多云CDN管理平台","link":"www.volcengine.com\u002Fproduct\u002Fmcdn","des":"基于数百Tbps业务规模实践打造的多云CDN运维管理服务,提供流量调度、运维监控、智能容灾、成本管理等能力","desShort":"多云架构下的CDN运维管理服务","enTitle":"mcdn"},{"title":"内容分发网络","link":"www.volcengine.com\u002Fproduct\u002FCDN","des":"基于覆盖全球骨干网以及各地域的海量加速节点,为图文、音视频等多种类型内容提供高性能,稳定的加速服务,提升访问体验","desShort":"稳定、弹性、高性能的全球内容分发服务","enTitle":"CDN"},{"title":"全球加速","link":"www.volcengine.com\u002Fproduct\u002Fga","des":"基于火山引擎边缘云覆盖全球的节点和骨干网基础设施,为游戏、协同办公和互联网应用提供安全、合规和可靠的全球网络加速服务","desShort":"安全、合规、可靠的全球网络加速服务","enTitle":"ga"},{"title":"全站加速","link":"www.volcengine.com\u002Fproduct\u002Fdcdn","des":"基于火山引擎边缘云覆盖全球的节点和骨干网基础设施,为您的API和互联网应用提供安全、全链路和高质量的网络加速解决方案","desShort":"安全、全链路、高质量的网络加速解决方案","enTitle":"dcdn"}]},{"title":"边缘计算","item":[{"title":"边缘计算节点","link":"www.volcengine.com\u002Fproduct\u002Fveen","des":"基于全域覆盖的边缘节点,提供弹性、可靠、分布式的异构算力资源和低时延的网络资源","desShort":"全域覆盖、弹性、分布式的异构算力资源","enTitle":"veen_edge"},{"title":"边缘渲染农场","link":"www.volcengine.com\u002Fproduct\u002FveRender","des":"高效、安全、稳定的渲染平台,依托底层海量边缘算力资源,多节点并行渲染,显著提升渲染效率","desShort":"高效、安全、稳定的分布式渲染平台","enTitle":"verender"}]}]]},{"name":"视频云","data":[[{"title":"视频服务","item":[{"title":"veImageX","link":"www.volcengine.com\u002Fproducts\u002Fimagex","des":"灵活、高效的图像、文档等各类素材上传、托管、智能处理和分发一站式解决方案","desShort":"素材托管、智能处理与分发全链路优化","enTitle":"ImageX"},{"title":"企业直播","link":"www.volcengine.com\u002Fproduct\u002Flive-saas","des":"0开发拥有定制直播间,千万并发流畅体验","desShort":"0开发、高可靠,打造企业级直播平台","enTitle":"livesaas"},{"title":"实时音视频","link":"www.volcengine.com\u002Fproduct\u002FveRTC","des":"亿级 DAU 产品验证打磨,提供全球范围内高质量、低延时的实时音视频通信能力","desShort":"首创支持单房间超千人开麦实时互动","enTitle":"rtc"},{"title":"视频点播","link":"www.volcengine.com\u002Fproduct\u002Fvod","des":"视频点播包含媒资上传、管理、处理、音视频分发播放的全链路能力","desShort":"端云一体、开箱即用的一站式点播方案","enTitle":"vod"},{"title":"视频直播","link":"www.volcengine.com\u002Fproduct\u002Flive","des":"为企业和个人提供安全、稳定、专业的云端直播服务","desShort":"低延迟、高画质、更流畅的直播服务","enTitle":"live"},{"title":"智能处理","link":"www.volcengine.com\u002Fproduct\u002Fimp","des":"弹性稳定智能的多媒体处理服务,助力企业提升音视频处理效率和质量","desShort":"全流程多媒体智能处理打造超高清画质","enTitle":"imp"},{"title":"WebRTC传输网络","link":"","des":"依托抖音、飞书等产品多年的经验技术积累,提供全球范围内高品质、超低延时的实时通信能力及配套场景化解决方案。","desShort":"WebRTC传输网络","enTitle":"wtn"},{"title":"视联网","link":"","des":"一站式物联网视频服务,便捷接入各类摄像机硬件和平台,行业视频互联网化","desShort":"一站式物联网视频服务","enTitle":"aiotvideo"}]},{"title":"云渲染与流化平台","item":[{"title":"云手机","link":"www.volcengine.com\u002Fproduct\u002FACEP","des":"高性能多场景云手机服务及解决方案","desShort":"自研ARM服务器,最大化模拟真机性能","enTitle":"ACEP"},{"title":"云游戏","link":"www.volcengine.com\u002Fproduct\u002FveGame","des":"结合云计算和超低延迟音视频传输技术的多场景云游戏解决方案","desShort":"一站式云游戏方案,小时级快速接入","enTitle":"vegame"}]}]]},{"name":"安全","data":[[{"title":"云安全","item":[{"title":"DDoS防护","link":"www.volcengine.com\u002Fproduct\u002FAntiDDoS-basic","des":"面向易遭受 DDoS 攻击的用户提供精准防护服务","desShort":"轻松抵御TB级大流量DDoS攻击","enTitle":"DDoS"},{"title":"Web应用防火墙","link":"www.volcengine.com\u002Fproduct\u002Fwaf","des":"有效防御恶意入侵和攻击,解决数据泄露以及合规、隐私保护等问题,从而保障数据安全性和应用程序可用性","desShort":"防御常见Web攻击,保障可用性","enTitle":"waf"},{"title":"高级网络威胁检测系统","link":"www.volcengine.com\u002Fproduct\u002Fnta","des":"高级网络威胁检测系统使用行为分析、特征检测、关联分析等技术检测网络威胁","desShort":"精准定位已知和未知威胁","enTitle":"nta"},{"title":"攻击面管理 SaaS版","link":"www.volcengine.com\u002Fproduct\u002Fasm","des":"从攻击者视角出发,全面、精准识别暴露给攻击者的客户资产及其安全风险,帮助客户从传统的安全事件响应,转变为“先发现、先预防、先处理”的主动持续监测模式","desShort":"从攻击者视角发现风险","enTitle":"asm"},{"title":"云安全中心","link":"www.volcengine.com\u002Fproduct\u002Fsecurity-center","des":"云安全中心是一个集资产管理、威胁检测、智能分析、协同响应于一体的云安全运营平台","desShort":"实时识别、分析、预警安全威胁","enTitle":"seccenter"},{"title":"云堡垒机","link":"www.volcengine.com\u002Fproduct\u002Fvbh","des":"云堡垒机是一款 4A 统一的运维安全审计平台,保障企业的安全合规","desShort":"4A 统一的运维安全审计平台","enTitle":"vbh"},{"title":"云防火墙","link":"www.volcengine.com\u002Fproduct\u002Fcloudfirewall","des":"基于公有云环境下的 SaaS 化防火墙,为您的云上资产提供访问控制、流量可视化和日志审计等功能","desShort":"一键开启云上资产安全访问控制","enTitle":"fw_center"},{"title":"多云安全平台","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Fcspm","des":"火山引擎旗下多云安全管理平台,统一管理资产、服务、事件、身份四个安全要素,为企业提供一站式安全运营、安全运维、安全数据共享能力","desShort":"连接多云和混合环境以获得全面的安全可见性。","enTitle":"msc"},{"title":"多云安全管理平台","link":"","des":"多云安全平台为企业客户提供SaaS化多云安全管理服务。通过统一管理多云、云上云下、中心+边缘等环境的资产、服务、告警、身份四要素,提供一站式安全运营、安全运维、安全数据共享能力。","desShort":"连接多云和混合环境以获得全面的安全可见性","enTitle":""}]},{"title":"办公安全","item":[{"title":"飞连","link":"www.volcengine.com\u002Fproduct\u002Ffeilian","des":"企业级数字化办公平台,帮助员工随时随地,安全连接内部网络与应用","desShort":"一站式 IT 管理与办公安全平台","enTitle":"corplink"}]},{"title":"终端安全","item":[{"title":"云工作负载保护平台","link":"www.volcengine.com\u002Fproduct\u002Fcwpp","des":"云工作负载保护平台提供主机、容器的反入侵、RASP、资产、漏洞与基线管理等能力","desShort":"为主机与容器提供统一可视的安全防护","enTitle":"cwpp"}]},{"title":"数据安全","item":[{"title":"密钥管理系统","link":"www.volcengine.com\u002Fproduct\u002Fkms","des":"一站式、全托管的密钥管理和数据加密服务平台","desShort":"提供一站式、全托管的密钥管理能力","enTitle":"kms"},{"title":"云加密机","link":"www.volcengine.com\u002Fproduct\u002Fcloudhsm","des":"满足金融、互联网等行业的加密需求,保护云上业务数据隐私安全","desShort":"满足加密合规需求,保护云上数据安全","enTitle":"hsm"},{"title":"公域数据报告","link":"","des":"基于多方安全计算、联邦学习、可信执行环境、同态加密和差分隐私等技术,结合火山引擎的各项业务实践,为合作伙伴提供综合全面的隐私计算解决方案,降低数据要素流通安全风险。","desShort":"基于多方安全计算、联邦学习、可信执行环境、同态加密和差分隐私等技术,结合火山引擎的各项业务实践,为合作伙伴提供综合全面的隐私计算解决方案,降低数据要素流通安全风险","enTitle":""}]},{"title":"安全专家服务"}]]},{"name":"容器与中间件","data":[[{"title":"容器服务","item":[{"title":"vci","link":"www.volcengine.com\u002Fproduct\u002Fvke","des":"通过深度融合新一代云原生技术,提供以容器为核心的高性能 Kubernetes 容器集群管理服务","desShort":"高性能Kubernetes容器集群管理","enTitle":"vci_pre"},{"title":"分布式云原生平台","link":"www.volcengine.com\u002Fproduct\u002Fdcp","des":"分布式云原生平台是面向多云多集群场景的企业级云原生统一管理平台","desShort":"面向多云多集群场景的统一管理平台","enTitle":"dcp"},{"title":"镜像仓库","link":"www.volcengine.com\u002Fproduct\u002Fcr","des":"提供安全高可用的容器镜像托管服务,方便用户对容器镜像进行全生命周期管理","desShort":"安全高可用的容器镜像托管服务","enTitle":"cr"},{"title":"容器服务","link":"","des":"容器服务(Volcengine Kubernetes Engine,VKE)提供高性能、高可靠、极致弹性的企业级容器管理能力,无缝集成弹性容器(VCI)的灵活弹性能力,支持云下弹云上、跨云弹性等分布式云原生能力。","desShort":"高性能Kubernetes容器集群管理","enTitle":"vke"},{"title":"vci","link":"","des":"高性能、稳定的弹性容器VCI服务","desShort":"弹性容器VCI","enTitle":""}]},{"title":"Serverless","item":[{"title":"函数服务","link":"www.volcengine.com\u002Fproduct\u002Fvefaas","des":"事件驱动的无服务器函数托管计算平台,支持快速创建和部署函数,连接云上中间件和数据库产品,帮助企业低成本构建复杂应用","desShort":"无服务器函数托管计算平台","enTitle":"vefaas"}]},{"title":"微服务治理","item":[{"title":"API网关","link":"www.volcengine.com\u002Fproduct\u002Fapig","des":"高扩展、高可用的云原生网关","desShort":"高扩展、高可用的云原生网关","enTitle":"apig"},{"title":"微服务引擎","link":"www.volcengine.com\u002Fproduct\u002Fmse","des":"一键快速构建云原生微服务体系","desShort":"一键快速构建云原生微服务体系","enTitle":"mse"},{"title":"应用韧性增强服务","link":"","des":"提供丰富的故障原子、跨集群的故障注入和管理、灵活的实验编排、多维度的指标统计,帮助客户提升系统韧性","desShort":"丰富的混沌演练能力以提升系统韧性","enTitle":""}]},{"title":"中间件","item":[{"title":"消息队列 Kafka版","link":"www.volcengine.com\u002Fproduct\u002FMessage-Queue-for-Kafka","des":"基于 Apache Kafka 构建,提供高可用、高吞吐量的分布式消息队列服务","desShort":"企业级高可用、全托管的Kafka服务","enTitle":"Kafka"},{"title":"消息队列 RabbitMQ版","link":"www.volcengine.com\u002Fproduct\u002FMessage-Queue-for-RabbitMQ","des":"消息队列 RabbitMQ版兼容开源 RabbitMQ,提供高可用、低延迟、高并发的消息服务","desShort":"企业级高可用、全托管的RabbitMQ","enTitle":"RabbitMQ"},{"title":"消息队列 RocketMQ版","link":"www.volcengine.com\u002Fproduct\u002FMessage-Queue-for-RocketMQ","des":"基于 Apache RocketMQ 构建的低延迟、高并发、高可用的分布式消息中间件","desShort":"企业级高可用、全托管的RocketMQ","enTitle":"RocketMQ"},{"title":"云搜索服务","link":"www.volcengine.com\u002Fproduct\u002Felasticsearch-service","des":"云搜索服务是基于Elasticsearch的在线分布式搜索服务,为您提供多条件的检索、统计等","desShort":"完全托管的在线分布式搜索服务","enTitle":"ESCloud"},{"title":"云原生消息引擎","link":"www.volcengine.com\u002Fproduct\u002Fbmq","des":"基于云原生的全托管、高吞吐、动态扩缩、高稳定性的分布式存算分离消息服务","desShort":"深度自研且兼容Kafka的消息中间件","enTitle":"bmq"}]},{"title":"云原生可观测","item":[{"title":"托管Prometheus","link":"www.volcengine.com\u002Fproduct\u002Fprometheus","des":"完全继承和对接 Prometheus 开源生态,为用户提供免运维、高可用的新一代云原生监控引擎系统","desShort":"免运维、高可用的云原生监控引擎","enTitle":"vmp"},{"title":"应用性能监控全链路版","link":"www.volcengine.com\u002Fproduct\u002Fapmplus","des":"提供针对应用服务的品质、性能以及自定义埋点的APM服务","desShort":"快速定位异常,优化应用品质","enTitle":"apmplus_api"},{"title":"云监控","link":"www.volcengine.com\u002Fproduct\u002Fcloudmonitor","des":"监控云上资源健康状态,并及时对异常资源告警,确保云上业务及资源顺畅运行","desShort":"7×24小时监控云资源健康状态","enTitle":"Volc_Observe"},{"title":"云拨测","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Fcloud-detect","des":"云拨测(CloudDetect)通过分布在全球各地的监测节点来模拟用户访问,实现端到端网络可用性、稳定性以及用户体验的可观测。帮助客户快速发现、定位和诊断网络服务问题,提升用户体验","desShort":"主动拨测,先于用户感知故障","enTitle":"cloud_detect"},{"title":"服务网格","link":"","des":"服务网格(Application Mesh Service,AMS)通过无侵入方式 为微服务应用提供可观测、服务治理和安全管控能力。","desShort":"服务网格(Application Mesh Service,AMS)通过无侵入方式 为微服务应用提供可观测、服务治理和安全管控能力。","enTitle":"ams"},{"title":"应用性能监控全链路版-inner","link":"","des":"APMPlus Client inner API","desShort":"APMPlus Client inner API","enTitle":""}]},{"title":"开发与运维","item":[{"title":"持续交付","link":"www.volcengine.com\u002Fproduct\u002Fcp","des":"企业级CI\u002FCD平台,助力构建集成自动化、应用交付自动化","desShort":"企业级CI\u002FCD,敏捷高效交付业务","enTitle":"cp"},{"title":"制品仓库","link":"www.volcengine.com\u002Fproduct\u002Fartifacts","des":"制品仓库(Artifacts)为用户提供开箱即用、高安全的云端制品托管服务,方便用户对源代码编译后的制品进行全生命周期管理","desShort":"开箱即用、高安全的制品托管服务","enTitle":"artifacts"}]}]]},{"name":"数据库","data":[[{"title":"关系型数据库","item":[{"title":"云数据库 PostgreSQL 版","link":"www.volcengine.com\u002Fproduct\u002Frds-pg","des":"云数据库 PostgreSQL 版是一种基于云平台的即开即用、稳定可靠、灵活弹性、易于使用的关系型数据库服务,RDS for PostgreSQL完全兼容原生PostgreSQL","desShort":"稳定可靠的弹性PostgreSQL服务","enTitle":"rds_postgresql"},{"title":"云数据库 SQL Server 版","link":"www.volcengine.com\u002Fproduct\u002Fsql-server","des":"云数据库 SQL Server 版拥有正版授权,能够高度兼容微软生态下的各类应用,并提供实例管理、备份恢复、日志管理、监控告警、数据迁移等全套解决方案,帮助企业更多聚焦于自己的核心业务。","desShort":"微软官方授权,广泛兼容各类应用","enTitle":"rds_mssql"},{"title":"云数据库 RDS MySQL 版","link":"www.volcengine.com\u002Fproduct\u002Frds-mysql","des":"基于云平台的即开即用、稳定可靠、灵活弹性、易于使用的关系型数据库服务","desShort":"即开即用稳定可靠的弹性MySQL服务","enTitle":"rds_mysql"},{"title":"云数据库 veDB MySQL 版","link":"www.volcengine.com\u002Fproduct\u002Fvedb-mysql","des":"新一代分布式关系型数据库,完全兼容 MySQL,提供安全可靠、简单易用和极致性价比的企业级数据库服务","desShort":"自研新一代云原生高性能MySQL服务","enTitle":"vedbm"},{"title":"图数据库 veGraph","link":"","des":"veStack 全栈版是火山引擎自主研发的云平台,包含计算、存储、网络、安全、数据库、容器、大数据、AI等产品能力,助力客户建设简单、智能、高效、全栈式的云平台,实现客户的数字化创新和智能升级","desShort":"自研新一代云原生高性能MySQL服务","enTitle":"graph"}]},{"title":"NoSQL数据库","item":[{"title":"表格数据库 HBase 版","link":"www.volcengine.com\u002Fproduct\u002Fhbase","des":"火山引擎表格数据库 HBase 版是基于 Apache HBase 提供的全托管数据库服务","desShort":"低成本高吞吐,兼容标准HBase","enTitle":"hbase"},{"title":"缓存数据库 Redis 版","link":"www.volcengine.com\u002Fproduct\u002Fredis","des":"与Redis兼容的全托管缓存和存储服务,以其超高读写性能为企业应用赋能","desShort":"兼具缓存的高性能与存储的持久化","enTitle":"Redis"},{"title":"文档数据库 MongoDB 版","link":"www.volcengine.com\u002Fproduct\u002Fmongodb","des":"开箱即用、稳定可靠、灵活弹性的云数据库服务,完全兼容原生 MongoDB","desShort":"开箱即用的文档数据库MongoDB","enTitle":"mongodb"}]},{"title":"数据库管理服务","item":[{"title":"数据库传输服务","link":"www.volcengine.com\u002Fproduct\u002Fdts","des":"提供一体化数据库数据传输管理服务,降低数据库之间数据流通复杂性,可帮助用户在业务不停服的前提下轻松完成数据库迁移上云","desShort":"数据库不停服,业务轻松迁移云端","enTitle":"dts"},{"title":"数据库工作台","link":"www.volcengine.com\u002Fproduct\u002Fdbw","des":"一款面向多类型数据库生命周期管理的统一云管平台。集数据库图形用户界面(GUI)、故障排查、审计于一体的数据库SaaS产品","desShort":"面向多类型数据库生命周期管理的统一云管平台","enTitle":"dbw"}]}]]},{"name":"存储","data":[[{"title":"基础存储服务","item":[{"title":"大数据统一存储 CloudFS","link":"www.volcengine.com\u002Fproduct\u002Fcfs","des":"大数据文件存储,支持HDFS语义,提供数据加速,多模式支持等能力","desShort":"高性能的存储加速服务","enTitle":"cfs"},{"title":"弹性块存储","link":"www.volcengine.com\u002Fproduct\u002Febs","des":"弹性块存储EBS是火山引擎提供的高可用、高可靠、高性能的块存储设备","desShort":"高可靠、高性能、低延时的块存储","enTitle":"storage_ebs"},{"title":"对象存储","link":"www.volcengine.com\u002Fproduct\u002FTOS","des":"海量、安全、低成本、易用、高可靠、高可用的分布式云存储服务","desShort":"高可用分布式存储,部署规模超10EB级","enTitle":"tos"},{"title":"日志服务","link":"www.volcengine.com\u002Fproduct\u002Ftls","des":"日志服务是火山引擎的实时日志平台,提供日志采集、检索分析等功能","desShort":"针对日志类数据的一站式服务","enTitle":"TLS"},{"title":"文件存储 NAS","link":"www.volcengine.com\u002Fproduct\u002Fnas","des":"NAS 是一种共享访问、持续在线、弹性扩展、跨地域访问的高性价比云存储服务","desShort":"云端共享文件存储,支持PB级数据量","enTitle":"FileNAS"},{"title":"文件存储 vePFS","link":"www.volcengine.com\u002Fproduct\u002Fvepfs","des":"vePFS 是一种高吞吐、低延时、可扩展的并行文件存储服务","desShort":"高吞吐、低延时的并行文件存储服务","enTitle":"vepfs"},{"title":"存储thallo管控服务","link":"","des":"存储通用管控服务,代理三方服务请求","desShort":"存储通用管控服务","enTitle":""}]},{"title":"数据迁移与工具","item":[{"title":"存储迁移服务","link":"www.volcengine.com\u002Fproduct\u002Fdms","des":"存储迁移服务用于将其他云服务商对象存储数据迁移到火山引擎对象存储","desShort":"快速平滑、高效安全的数据迁移服务","enTitle":"dms"}]}]]},{"name":"网络","data":[[{"title":"云上网络","item":[{"title":"NAT网关","link":"www.volcengine.com\u002Fproduct\u002Fnat","des":"为私有网络内的云服务器实例提供网络地址转换能力,使子网内的多个云服务器可以共享公网IP访问公网","desShort":"灵活部署,内外网连接统一地址转换","enTitle":"natgateway"},{"title":"负载均衡","link":"www.volcengine.com\u002Fproduct\u002Fclb","des":"一种将访问流量按策略分发给多台后端服务器的服务,可以提高系统的整体可用性","desShort":"高性能流量分发,提升系统可用性","enTitle":"clb"},{"title":"公网IP","link":"www.volcengine.com\u002Fproduct\u002Feip","des":"一种可以独立购买和持有的公网IP地址资源。通过和云资源绑定,实现云资源与公网的连接","desShort":"弹性、灵活、安全可靠的公网服务","enTitle":"eip"},{"title":"私有网络","link":"www.volcengine.com\u002Fproduct\u002Fvpc","des":"为云上资源构建隔离的虚拟网络环境,便于您自主规划、配置和管理云上网络","desShort":"安全隔离的云上网络环境","enTitle":"vpc"},{"title":"应用型负载均衡","link":"www.volcengine.com\u002Fproduct\u002Falb","des":"面向 7 层互联网应用及云原生应用,基于内容精细化调度,提供稳定、弹性、安全、高性能的应用层负载均衡服务","desShort":"稳定、弹性、安全、高性能的应用层负载均衡服务","enTitle":"alb"},{"title":"私网连接","link":"","des":"通过终端节点和终端节点服务建立私密的VPC间网络连接,避免了公网访问潜在的安全风险,提供更加安全的组网方式","desShort":"安全可靠的单向网络访问","enTitle":"privatelink"},{"title":"中转路由器","link":"","des":"中转路由器是地域级路由器产品,通过丰富的路由功能支持客户灵活构建任意拓扑云上网络","desShort":"可灵活自定义云上任意网络拓扑","enTitle":"transitrouter"}]},{"title":"混合云网络","item":[{"title":"VPN连接","link":"www.volcengine.com\u002Fproduct\u002Fvpn","des":"通过特定的加密方式,在Internet网络中建立的临时、安全、可靠的通信隧道","desShort":"公网加密隧道,让业务安全访问云端","enTitle":"vpn"},{"title":"专线连接","link":"www.volcengine.com\u002Fproduct\u002Fdirectconnect","des":"专线连接是一种连接用户本地数据中心与火山引擎云上网络的服务,能够搭建高速、低时延、稳定安全的专属通道","desShort":"高速、稳定、安全的专属连接通道","enTitle":"directconnect"},{"title":"bandwidthquota","link":"","des":"该项目旨在提供一种简单且用户友好的方式来管理物理专线资源","desShort":"物理专线资源水位管理系统","enTitle":""}]},{"title":"跨地域网络","item":[{"title":"云企业网","link":"www.volcengine.com\u002Fproduct\u002Fcen","des":"通过自动路由分发学习,提供云上网络实例之间高速稳定的网络互通能力","desShort":"高速稳定的企业级云网互通","enTitle":"cen"}]}]]},{"name":"计算","data":[[{"title":"弹性计算","item":[{"title":"弹性伸缩","link":"www.volcengine.com\u002Fproduct\u002Fas","des":"根据用户的业务需求自动灵活经济的调整计算资源,保证业务平稳健康运行","desShort":"自动化、灵活经济的云资源管理","enTitle":"auto_scaling"},{"title":"服务器迁移中心","link":"https:\u002F\u002Fwww.volcengine.com\u002Fproduct\u002Fsmc","des":"可以通过SMC将本地服务器、虚拟机、其他云平台的云服务器迁移至火山引擎","desShort":"火山引擎提供的服务器迁移平台","enTitle":"smc"},{"title":"云服务器","link":"www.volcengine.com\u002Fproduct\u002Fecs","des":"提供稳定的弹性计算服务。通过实时增减计算资源,适应业务变动,降低维护成本","desShort":"灵活稳定的弹性计算服务","enTitle":"ecs"},{"title":"云原生计算引擎","link":"","des":"为泛计算产品(VCI\u002FECS\u002FTCE)提供池化的资源池","desShort":"泛计算(VCI\u002FECS\u002FTCE)底座","enTitle":"ubiquitous compute engine"},{"title":"高性能计算服务EHPC","link":"","des":"高性能计算服务EHPC是一款火山引擎自研的高性能计算集群管理工具。它集成了多款作业调度软件、支持云上弹性伸缩,让您只需几步操作就能快速创建出配置完整的HPC集群,旨在简化火山引擎云环境中运行并行工作负载的过程,为您提供弹性灵活、自助化的计算服务。","desShort":"高性能计算服务EHPC是一套帮助客户在云上快速搭建HPC环境并提供任务调度与集群管理的工具。","enTitle":""}]},{"title":"混合云"},{"title":"操作系统","item":[{"title":"veLinux","link":"www.volcengine.com\u002Fproduct\u002Fvelinux","des":"针对火山引擎公有云环境进行深度定制与优化的自研操作系统,提供了更安全高效的系统服务和环境","desShort":"云端定制,支持海量业务和亿级并发","enTitle":"velinux"}]}]]},{"name":"服务支持","data":[[]]},{"name":"解决方案","data":[[]]},{"name":"销售体系","data":[[]]},{"name":"合作与生态","data":[[]]}]}},"recomArticles":{"err_no":0,"err_msg":"","data":[{"item_id":"7419482941090168870","name":"Bun 1.0,新的 JavaScript 运行时,内置了打包器、转译器、任务运行器和 np","user_id":3402360430542128,"publish_time":"2024-09-29T00:59:12.000Z","category_ids":"[5]","tag_ids":"[7397779777916633142]"},{"item_id":"7436341741323550771","name":"Agent系列:AppAgent v2-屏幕智能Agent(详解版)","user_id":1977369294279187,"publish_time":"2024-11-12T10:43:56.000Z","category_ids":"[5]","tag_ids":"[7307103161498632233,7146114517833875464,7146205183884656670]"},{"item_id":"7418740225456734227","name":"使用 KMP 构建 Compose Multiplatform 共享 UI","user_id":1484818005369147,"publish_time":"2024-09-26T03:00:39.000Z","category_ids":"[5]","tag_ids":"[7397779777916633142,7347591544491638803,7146205340617408543]"},{"item_id":"7437331886491648026","name":"无需网络,轻松在手机上离线布署本地大模型","user_id":2417169195086699,"publish_time":"2024-11-15T07:29:11.000Z","category_ids":"[5]","tag_ids":"[7307103161498632233,7146205471601328164,7146205340643065870]"},{"item_id":"7418398064655728681","name":"Kotlin新技巧: combine 函数组合多个 Flow","user_id":1484818005369147,"publish_time":"2024-09-25T13:21:54.000Z","category_ids":"[5]","tag_ids":"[7397779777916633142,7347591544491638803,7146205340617408543]"},{"item_id":"7419128332123373622","name":"在 Jetpack Compose 中实现 Android 15 的预测性返回","user_id":1484818005369147,"publish_time":"2024-09-27T02:27:09.000Z","category_ids":"[5]","tag_ids":"[7397779777916633142,7347591544491638803,7146205340617408543]"}]}}},"errors":null}</script> <script> const __ENV = 'prod'; function slardarInit(w, d, u, b, n, pc, ga, ae, po, s, p, e, t, pp) { pc = 'precollect'; ga = 'getAttribute'; ae = 'addEventListener'; po = 'PerformanceObserver'; s = function (m) { p = [].slice.call(arguments); p.push(Date.now(), location.href); (m == pc ? s.p.a : s.q).push(p); }; s.q = []; s.p = { a: [] }; w[n] = s; e = document.createElement('script'); e.src = u + '?bid=' + b + '&globalName=' + n; e.crossOrigin = u.indexOf('sdk-web') > 0 ? 'anonymous' : 'use-credentials'; d.getElementsByTagName('head')[0].appendChild(e); if (ae in w) { s.pcErr = function (e) { e = e || w.event; t = e.target || e.srcElement; if (t instanceof Element || t instanceof HTMLElement) { if (t[ga]('integrity')) { w[n](pc, 'sri', t[ga]('href') || t[ga]('src')); } else { w[n](pc, 'st', { tagName: t.tagName, url: t[ga]('href') || t[ga]('src'), }); } } else { w[n](pc, 'err', e.error || e.message); } }; s.pcRej = function (e) { e = e || w.event; w[n](pc, 'err', e.reason || (e.detail && e.detail.reason)); }; w[ae]('error', s.pcErr, true); w[ae]('unhandledrejection', s.pcRej, true); } if ('PerformanceLongTaskTiming' in w) { pp = s.pp = { entries: [] }; pp.observer = new PerformanceObserver(function (l) { pp.entries = pp.entries.concat(l.getEntries()); }); pp.observer.observe({ entryTypes: ['longtask', 'largest-contentful-paint', 'layout-shift'], }); } } slardarInit( window, document, 'https://lf3-short.ibytedapm.com/slardar/fe/sdk-web/browser.cn.js', 'eps_platform_fe_tech', 'Slardar', ); const env = __ENV === 'prod' ? 'production' : 'development'; window.Slardar('init', { bid: 'eps_platform_fe_tech', env, release: '1.0.1.922', // 区分上报版本 }); window.Slardar('start'); </script> <!-- <script src="/sdk/mermaid/mermaid@9.3.0.min.js"></script> <script> window.mermaid.init({ noteMargin: 10 }, '.language-mermaid'); </script> --> <script type="module"> import mermaid from 'https://cdn.jsdelivr.net/npm/mermaid@9/dist/mermaid.esm.min.mjs'; mermaid.init({ noteMargin: 10 }, '.language-mermaid'); </script> <script type="text/javascript" src="https://res2.wx.qq.com/open/js/jweixin-1.6.0.js" ></script> </body></html>