撰稿：白默
陈万做付费投研简报快两年了。一周前，他从程序里删掉了两份维护已久的文件。
一份叫 cleaner.py，负责清洗网页；另一份叫 sites.yaml，列着 387 个网址，用来筛选信息来源。
过去，一个财经网站改版，前一份文件就要补一段规则；一位订阅者来邮件追问数字出处，后一份文件里往往又会多一个经过核实的网址。
陈万的研报聚焦新能源产业链，覆盖储能和锂电上下游，已有约 4000 名国内外订阅者。他们多数在券商研究所、公私募基金、产业资本和电池厂工作。研报里的信息会进入晨会讨论，也可能成为投资决策的参考。对陈万来说，找到信息只是第一步，弄清来源和时间同样重要。
他一直希望把网页清洗和信源筛选交给更成熟的服务，自己专注于研究和写作。他试过百度搜索，也用过其他产品内置的联网搜索，但这些工具不完全适合他的工作流。他需要一个能独立接入 Agent 的搜索接口：模型由自己选择，搜索结果先返回自己的程序，再决定如何使用。
删除文件前，陈万做了一次对比：把同一组问题分别发给使用了一年的海外搜索 API、一个国内 API，以及腾讯云刚升级的元宝搜索 API。在这次测试中，前两类接口返回的多是一两行摘要；元宝搜索的结果则提供了两类文字，一类概括网页内容，另一类围绕问题提取相关段落。
他还查了一家电池企业前一晚发布的业绩预告。海外接口返回了几条英文转载，时间与他的查询需求并不匹配；元宝搜索给出了交易所原始公告、几篇相关报道，以及各条结果的发布时间。
陈万需要的不是搬回整篇文章，而是知道信息在哪里、说了什么，并能带着读者回到源头。
这两份文件，不只陈万有。它也代表着不少开发者在搭建 Agent 时要处理的基础工作：从网页中提取有效内容，为信息来源建立筛选规则，再把不同接口返回的数据接入同一套流程。
网站结构会变，来源需要核实，公告、行情和新闻也各有更新节奏。应用做起来以后，这些工作就成了持续的维护任务。对于陈万这样的小团队，一套稳定、独立的搜索服务，意味着可以把更多时间留给业务本身。
2025 年 8 月，微软关停 Bing Search API，原有用户需要调整检索方案。与此同时，Agent 对联网信息的需求仍在增长。在 10 月 20 日元宝搜索的媒体沟通会上，元宝搜索提到，自今年 2 月“龙虾”走红以来，Agent 对热点新闻、财经财报等时效性外网信息的需求变得更为迫切。
模型能力之外，信息获取也在影响 Agent 的表现。腾讯云副总裁、腾讯元宝搜索研发负责人表示：“选择一个模型相对容易，但做好搜索却需要长期的建设与迭代。”
一、 两份文件的来历
要理解陈万为什么需要这两份文件，得先看看网页到了程序里是什么样子。
一份上市公司的业绩预告，在浏览器里可能只是标题、日期、几段正文和一张表。程序拿到的却是整页源码：嵌套的标签、样式表、统计脚本，以及广告、导航和版权声明。把这些内容直接交给模型，会让不少 Token 消耗在与任务无关的部分。
陈万的办法，是为不同网站编写抓取规则：正文在哪个标签里，日期藏在哪个位置。他估算，清洗之后，每次调用的 Token 成本降到了原来的不到 10％。
但代价是 cleaner.py 越写越长，网站一改版，相应规则就需要更新。
搜索服务商同样在处理效率问题。元宝搜索研发负责人介绍，内部团队训练了一个压缩模型，在建索引时把几千 Token 的网页压缩成约 100 个 Gist Token 并缓存，在线匹配时处理少量后续 Token，推理成本约可节省 30％。
成本之外，陈万更关心信息来源。
简报发出的头两个月，他陆续收到订阅者来信：某个电芯报价是哪天的？一个出货量数字，为什么和其他渠道看到的不一样？回查后他发现，有的转载内容标错了日期，也有信息来自自己并不熟悉的网站。
他开始为常用信源建立名单，记录哪些网站经过核实、适合优先检索。sites.yaml 里的 387 个网址，就是这样积累起来的。这份名单不能替代逐条核实，却给日常筛选提供了起点。
一位有多年财经报道经验、公众号订阅者超过 6 万的媒体人，也有类似需求。她说，关键事实信息通常要回到网页核实，行业观点则常常要到活跃的微信账号中寻找。她希望搜索能先呈现可靠且相关的信息，帮她判断哪些值得点开细读。
不同内容，需要不同的信源判断方式。据元宝搜索负责人介绍，政策、新闻和医疗法规等信息，优先参考政府、企业、交易所等官方来源；对于行业讨论，持续在相关领域发声的作者也有参考价值，团队会结合账号的历史发文进行交叉验证，再纳入整体排序。
此外，还要区分“多条结果”和“多个独立来源”。同一条新闻可能被几十个网站转载，标题不同，正文却相近。如果不识别这些关联，重复信息就可能被当作相互印证的依据。AI 批量生成的内容和针对 AI 搜索做的优化（GEO），还在让这类噪音变多。
搜索引擎的发展分为三个阶段：1.0 阶段，搜索把结果交给人，由人阅读、判断；2.0 阶段，模型读取结果，组织成答案；到了 3.0 阶段，Agent 会拆解任务、多次检索，并结合返回的信息继续推理。搜索因此不仅要找到内容，还要把来源、时效和相关性表达清楚，让后续判断有据可依。
相应地，质量评估也从单篇内容延伸到了发布者及其历史记录。团队既分析页面内容，也识别作者、账号和站点的持续表现，对频繁发布低质或 AI 生成的内容开展专项识别。这些评估信息会作为搜索信号反馈给 Agent，为模型判断内容提供参考。
腾讯云的联网搜索 API 于 2025 年上线，主力接口叫 SearchPro。今年产品更名为元宝搜索，并继续更新接入方式、检索参数和返回字段。
据腾讯在此次沟通会上介绍，升级后的元宝搜索已在内部 50 多个核心产品中得到验证，拥有千亿级融合索引，日增索引约 50 亿，日均调用量数亿次；底层检索平均耗时约 8 毫秒。这些指标对应服务的不同环节，其中底层检索耗时并不等于一次完整请求的响应时间。
陈万这次接入的，就是升级后的版本。
二、一次请求的路径
两份文件原来承担的工作，可以沿着一次搜索请求重新看一遍。
几乎每天早上，陈万都要整理产业链上市公司前一晚的公告、业绩预告，以及相关行情变动。他的 Agent 向 SearchPro 发出请求时，除了问题，还会根据任务设置检索范围和时间条件。
首先是去哪里找。
券商客户经理王宇对此有过切身体会。他的团队曾为客户搭建问答助手：回答一只股票的问题，要分别调用行情、新闻和通用搜索接口，再由代码整合结果。有一次，不同数据的时间没有对齐，团队花了两天才找到拼接逻辑中的问题。
SearchPro 通过 Mode 参数区分检索来源：0 以公开网页结果为主，1 返回优质权威垂直信源结果，2 混合返回两类结果；Industry 则可以进一步指定政府、新闻、学术或财经领域。开发者可以在一次请求中表达这些来源需求。
优质垂直信源结果也被称作 VR 卡。元宝搜索将其机制解释为一套类似“插件”的系统：接入各垂直领域的优质信源，再将返回内容封装成卡片，供模型使用。金融领域的例子，是与腾讯金融科技合作接入实时股票查询信息。
如果还要限定具体网站，可以使用 Site 参数：每次指定一个域名检索，或排除最多五个域名。它只作用于公开网页结果，不影响 VR 卡。这类边界也意味着，原有白名单和黑名单不能直接等同于一个接口参数，仍需结合业务重新安排筛选规则。
其次是信息有多新。
陈万的 cleaner.py 原来有一段专门提取日期的代码。有的网页不标日期，有的只标转载时间，还有的把日期放在图片里。取不到准确时间，程序就很难判断这条信息是否适合当天的简报。
SearchPro 提供 Freshness 参数，支持按天、月、年及指定日期区间检索。例如，d2 表示最近两天，m3 表示最近三个月；返回结果中的 date 字段则提供内容发布时间。陈万的晨报通常使用 d1 或 d2，周末回顾则放宽到 d7。
这让时间条件成为检索的一部分，减少了他自行解析网页日期的工作。但发布时间仍是判断线索，涉及重要数据时，还需要区分公告发布日、统计期和转载时间。
“对于财经、股票等时效性较高的网页，团队会考虑采用更高频的刷新策略，例如每小时刷新一次，以便搜索时获取更新的信息。”元宝搜索负责人表示。
接下来，是结果如何筛选。
每条结果包含两个不同的信号：score 表示与问题的相关程度，取值为 0 到 1；authority_level 表示信源权威度，取值为 0 到 5。一条结果可以很对题，但来源仍需进一步核实；也可能来自权威机构，却与当前问题关系不大。
元宝搜索会把“谁发布”和“内容如何”分开评估。权威性关注发布者在相关领域的身份与积累；内容质量则综合考察真实性、信息完整度和专业程度，包括能否追溯验证、是否注明引用来源等。评估覆盖站点、账号和页面三个层级，并结合大模型对页面结构的分析。
这些信号为筛选提供参考，但不替代具体任务中的判断。陈万据此设置了一套分层规则：权威度 4 分及以上的结果，作为简报正文的候选材料；3 分的放入延伸阅读；更低的暂不采用。分数帮助他缩小核实范围，关键数字和结论仍应回到原始来源确认。
过去，387 个网址承载着他逐个积累的信源经验；现在，接口评分提供了覆盖更广、便于调整的初筛依据。
筛选之后，是内容如何返回。
SearchPro 提供两类摘要：passage 是标准摘要，content 是动态摘要。据产品团队介绍，后者会围绕本次问题提取网页中的相关内容，而不是返回整页正文。同一篇文章，查询营收增速和查询订单进展，得到的相关片段可能不同。
这与陈万原来的流程形成了区别：cleaner.py 先清洗整页，再让模型寻找相关部分；动态摘要则先提供与问题有关的内容，减少后续的处理。他说：“我写了四百多行代码去逼近的东西，现在已经成为一个字段。”
内容获取也设置了边界。元宝搜索毕老师介绍，对设置反爬机制或需要借助额外工具才能访问的内容，“我们就不收录，以尊重对方的知识产权”，只提供公开、直接点开就能看到的内容；携程这类直接打不开的内容，则通过三方合作提供。
最后，搜索结果返回陈万的程序，由他的流程决定如何继续使用。马晋表示，智能体也会把一个问题拆成多个子问题，“搜索每慢一秒，都可能影响用户体验”；据他介绍，升级后，最快的版本可能在 300 毫秒以内返回，此前约为 400—500 毫秒。
三、接入与控制
搜索能力之外，接入方式决定了它能否顺畅进入现有工作流。
最早的接法走腾讯云标准 API：请求发往 wsa.tencentcloudapi.com，Action 写 SearchPro，Version 写 2025-05-08，再用 SecretId 和 SecretKey 按腾讯云的签名算法给每个请求签名。
某数据服务公司的后端工程师郑楠介绍说，他去年接入的时候花了半天，大半时间在调签名。后来腾讯云增加了单一 API Key 认证，请求发往 api.wsa.cloud.tencent.com 的/SearchPro 路径，签名这一步可以直接省掉。
另一种接入方式更快。在支持该协议的客户端中配置服务地址和 API Key，就能把 SearchPro 加入工具列表，不需要额外安装本地服务。元宝搜索目前提供的 MCP 工具只有 SearchPro，采用 Streamable HTTP 传输。
MCP 是 Anthropic 于 2024 年 11 月发布的开放协议，用于连接 AI 应用与外部数据、工具。此后一年多，主流 AI 开发工具陆续支持了它。元宝搜索的 MCP 页面写明，它使用的协议版本是 2025-06-18。腾讯做的，是把自己的搜索按这个协议重新包装了一层。
再往上，是指导 Agent 使用工具的 Skill。2026 年 3 月 10 日，腾讯上线 SkillHub，最初是 OpenClaw 技能生态 ClawHub 的国内镜像，收录一万三千多个技能；到年中，收录量超过八万个，支持 WorkBuddy、QClaw、ima 等工具。
SkillHub 上“元宝搜索 Skill”的安装方法只有一句话，大意是让 Agent 按 SkillHub 的安装说明装上 tencent-yuanbao-standard-search。Skill 的说明中列出了什么时候该用它：用户说“帮我搜索”“查找”，或者说“最新的……”“近期的新闻……”。
API、MCP 和 Skill 是解决不同问题的三类能力，并不是互相替代的三代产品。API 便于程序直接控制检索次数、时机和重试；MCP 解决兼容客户端如何调用工具；Skill 则补充使用说明，帮助助手判断何时检索、如何使用。
接入后的第二天，陈万在 Claude 终端里修改一份给订阅者的说明文档，说：“把昨晚最新的业绩预告也补进去。”随后，终端出现了 SearchPro 工具调用，耗时 681 毫秒，返回七条结果，其中五条是权威信息。
“它是听到‘最新’两个字就自己去的，”陈万说，“它已经成了任务流程中的一部分。”
一家做内容产品的团队有类似体会。团队的一位开发者在沟通会上说，他们原来自建的 AI 搜索“只能说能用”，Agent 常常先找出十几个网站，再反复调整。通过 Skill 接入元宝搜索后，“速度快，也不用反复检索”，以前问同一个问题要找 20 个网页，现在 5 个就能解决。
虽然接入变简单了，但没有简单到什么都不用管。保管密钥、配置账号权限，并根据调用量安排请求频率。接口承接了一部分基础工作，但应用如何运行，仍由开发者掌握。
接入的另一端，是成本。
根据腾讯云目前公开的计费文档，元宝搜索提供极速版、标准版和旗舰版，按量计费价格分别为每千次 14 元、20 元和 26 元，其中旗舰版仅限企业用户开通。
此外，填写邀请码后可获得每月 6000 次免费调用额度，按日均计算约 200 次，大致够一个小团队做日常测试或轻量使用。
陈万这次采用的权威度字段，属于旗舰版能力。
据腾讯产品团队介绍，元宝搜索定位于被调用方，不做独立的搜索入口，面向用户的入口由各个应用承担。它的调用方既包括腾讯内部产品，也包括外部 AI 助手和大模型公司。
过去一年，陈万的程序换过多次底层模型，从 DeepSeek 到通义千问，再到智谱，最后又换回 DeepSeek。搜索接口是第一次换。“模型换了，信息质量基本看不出来，”他说，“但搜索换了，效果完全不一样。”
四、交出去的，与留下的
删掉文件之后，陈万的搜索照常推送。
他做了一次最终检查。当天的公告和行业新闻覆盖率基本符合需要，来源里还多了几个此前较少涉及的站点；最近一周，也没有订阅者来信追问数据出处。
他维护了两年的文件，最终被一个搜索接口替代，“没有什么不舍，只是觉得轻松。”他说，并不是他的代码能力不好，而是这类代码本就不应该让每个人自己写。
现在，陈万可以把时间更多地用在核实关键数据、比较产业变化和组织研究判断上。
元宝搜索表示，这种调用需求正快速增长，目前最看好的是两个方向。一是智能办公和企业级 Agent：一款内部产品全面推广后，其调用量从低谷时的几十万次，增长到单日接近千万次。二是金融和投研：“研报和行情信息可能需要按小时甚至按分钟更新”，一个金融大模型接入后，近期的调用量增长了几十倍。
陈万的付费研报，正好落在这类需求上。
对中文 Agent 开发者来说，这类独立接口相当于增加了一种分工方式：把过去通用的信息检索、内容提取和来源评估，交给公共服务直接完成。
这也是这件事超出个案的地方。至于未来规模能做多大，取决于接下来有多少应用建在它上面，也取决于价格能不能降到小团队愿意把它设成默认能力。
陈万删掉的是两份文件，但也留下了对信息的更高要求。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Contribution: silence
Chen Wan has been doing paid research brief for almost two years. A week ago, he removed two long-maintained documents from the procedure.
One, cleaner.py, is responsible for cleaning web pages; the other, sites.yaml, lists 387 web sites used to screen information sources.
In the past, a financial web site has been re-edited with a rule added to the previous document; a subscriber has come to follow up on the digital source, and in the latter one often has an additional verified web site.
Chen Wan's research focused on the new energy industry chain, covering both storage and lithium power upstream and downstream, with about 400 domestic and foreign subscribers. Most of them work in voucher institutes, public and private fund-raising, industrial capital and battery plants. The information in the study is discussed at morning meetings and may serve as a reference for investment decisions. For Chen Wan, finding information is only the first step, and identifying sources and times is equally important.
He had always wished to refer web cleansing and filtering to more mature services, focusing on research and writing. He tried to search for Baidu, but these tools were not entirely suitable for his workflow. He needs a search interface with independent access to Agent: the model is chosen by himself, the search results go back to their own program and then decide how to use it.
Before deleting the document, Chen Wan made a comparison: sent the same set of questions to an overseas search for API, a domestic API, and a recently upgraded Tencent cloud for API. In this test, most of the first two types of interfaces returned one or two rows of summary; the results of the beagle search provided two types of text, one summarizing the web content and the other extracting the relevant paragraphs around the issue.
He also checked the performance forecast of publish the previous night for a battery company. The overseas interface returned several copies in English, the time of which did not match his queries; the betel search gave the original exchange bulletin, several related stories and publish time for the results.
What Chen needs is not to move back to the article as a whole, but to know where the information is, what it says and to be able to return to the source with the reader.
These two documents are not limited to anything else. It also represents the foundation that many developers have to deal with in building Agent: Take effective content from the web page, establish screening rules for information sources, and access data returned from different interfaces to the same process.
The structure of the website will change, sources will need to be verified, and announcements, newsletters and news will be updated. When applied, they become ongoing maintenance tasks. For a small team like Chen Wan, a stable, independent search service means that more time can be left to the business itself.
In August 2025, Microsoft shut down Bing Search API, and the original user needed to adjust the retrieval programme. Meanwhile, Agent's demand for networked information continues to grow. At the media conference on October 20, the Japanese yen search, the Yuanbao search mentioned that since February this year, when the lobster went red, the demand by Agent for time-bound offline information, such as news, financial and financial statements, became even more urgent.
In addition to model capabilities, access to information is also affecting Agent 's performance. Vice-President Tencent Cloud, Head of Tencent Search Research and Development, said: “The choice of a model is relatively easy, but a search requires long-term construction and iterative.
I. History of the two documents
To understand why Chen Wan needed the two documents, it was necessary to see what the web page looked like in the program.
A forecast of the performance of the listed company may be in the browser only by title, date, several body parts and a table. The program gets the entire page source code: embedded labels, style sheets, statistical scripts, and advertising, navigation and copyright statements. Turning these things over directly to the model would consume Token in parts unrelated to the mission.
Chen Wan's approach is to prepare capture rules for different websites: which label is in which text and where the date is hidden. He estimated that after cleaning, each call to Token cost had dropped to less than 10 per cent.
However, the cost is that the longer you write, the longer the website is, the more the rules need to be updated.
Search service providers also deal with efficiency issues. The head of the Renbao search research and development described an internal team training a compression model that would compress thousands of Token's web pages into about 100 Gist Token and cache, handling a small amount of follow-up Token when matching online, with approximately 30 per cent savings in reasoning costs.
In addition to costs, Chen Man is more concerned with sources of information.
During the first two months of the briefing, he received a letter from the subscriber: what day was the offer for a core? Why is a volume figure different from what other sources see? Upon return, he found that some of the reproductions were dated on the wrong dates and that there was information from websites he was not familiar with.
He began to create a list of commonly used sources of mail, recording which sites were verified and suitable for priority search. This is how 387 web sites are built up. This list is not a substitute for article-by-article verification but provides a starting point for daily screening.
A mediaman with years of financial and economic experience and more than 60,000 WeChat Official Account subscribers has similar needs. She said that key factual information was usually sent back to the web site for verification and that industry perspectives were often sought in active WeChat accounts. She hoped that the search would begin with reliable and relevant information that would help her judge what was worth reading.
Different content requires different source judgement. According to the head of the search, priority is given to official sources such as government, business, exchange, etc., for information on policies, news and medical regulations; for industry discussions, there is also a value for authors who continue to speak in the relevant fields, and teams cross-check the accounts in conjunction with historical texts and then integrate them into the overall order.
There is also a distinction between “multiple outcomes” and “multiple independent sources”. The same news could be reproduced on dozens of websites, with different titles and close to the body. If these linkages are not identified, duplicated information may be used as a basis for mutual corroboration. The content generated by the AI batch and the optimization of the search for the AI (GEO) are also increasing the noise.
The development of the search engine consists of three phases: phase 1.0, where the search results are handed over to the person and the person reads and judges; phase 2.0, where the model reads the results and organizes them into an answer; and phase 3.0, where Agent dismantles the task, retrieves it repeatedly and continues to reason with the information returned. The search must therefore be based not only on finding the content, but also on a clear statement of origin, timeliness and relevance so that subsequent judgements can be substantiated.
Accordingly, the quality assessment was extended from a single section to publishers and their historical records. The team analyzes both the content of the page and the continuous performance of the author, account number and site, and specifically identifies the content generated by frequent publish or AI. These assessments feed back to Agent as a search signal to inform model judgement.
The Tencent cloud network search API came online in 2025 and the main interface is called SearchPro. This year, the product is renamed the beacon search and continues to update access, search parameters and return to field.
According to Tencent, the upgraded Treasure Search has been validated in more than 50 core products, with an index of hundreds of billions of steps, with an increasing index of about 5 billion, with a daily average of hundreds of millions of calls; and a bottom search of average latency, about 8 milliseconds. These indicators correspond to the different segments of the service, where bottom search does not take time to respond to a full request.
Chen Wan's access is an upgraded version.
II. Path to a request
The two documents, which were originally undertaken, could be re-examined along a search request.
Almost every morning, Chen Man prepares the announcement of the night before the company was listed in the industrial chain, the performance forecast and the related changes. His Agent sends a request to SearchPro, setting the search scope and time conditions according to the task, in addition to the problem.
First, where to look.
The customer manager, Wang Woo, knew about this. His team had set up a question and answer assistant for the client: to answer the question about a stock, to call the business, news and general search interfaces separately, and then integrate the results by code. In one case, the timing of the different data was incomplete, and it took the team two days to find the problem in the collage logic.
SeerchPro distinguishes search sources by Mode parameter: 0, primarily with open web results,1 returns quality authoritative vertical source results,2 mixed returns of two types; Industry can further designate government, news, academic or financial areas. The developers can express these source needs in a single request.
The quality vertical source results are also known as VR cards. It interprets its mechanism as a system similar to a “plug-in” system: access to high-quality mail sources in all vertical fields, and then encapsulate the return content into cards for use in models. In the area of finance, an example is the cooperation with Tencent financial science and technology to access real-time equity information.
If a specific website is to be qualified, the Site parameter can be used: each time a domain name is specified, or a maximum of five domain names is excluded. It only serves to publicize web results without affecting VR cards. Such boundaries also mean that the original white and black lists cannot be directly equated with an interface parameter and that screening rules still need to be reorganized in conjunction with operations.
The second is how new the information is.
Chen Wan's cleaner.py used to have a code for extracting dates. Some pages do not indicate a date, others only indicate a time to reproduce and others appear in pictures. Without accurate time, it is difficult for the procedure to judge whether this information is suitable for the day ' s briefing.
The SearchPro provides Freshness parameters to support search by day, month, year and specified date range. For example, d2 indicates that the last two days, m3 indicates the last three months; the date field in the return results provides publish time. Chen Wan ' s morning newspapers usually use d1 or d2 and the weekend reviews relax to d7.
This made time conditions part of the search and reduced his own task of deciphering the page dates. However, publish time is still the point of judgement, and when it comes to important data, a distinction needs to be made between publish days, statistical periods and reproduction times.
“For more time-bound web pages, such as finance, stocks, the team would consider adopting a more high-frequency updating strategy, such as an hourly one, to obtain updated information when searching. The search manager for Yuanbao said:
Next is how the results are screened.
Each result contains two different signals: the score indicates the degree of relevance to the problem, with values ranging from 0 to 1; the power level indicates the source authority, with values ranging from 0 to 5. An outcome can be quite cross-cutting, but the sources need to be further verified; it may also come from authorities, but it has little to do with the current problem.
"Who is publish" and "What's the content?" are assessed separately by the Yuanbao search. Authoritatively concerned with the identity and accumulation of publishers in related fields; content quality examines the authenticity, integrity and professionalism of the information, including the possibility of retroactive validation, whether the source of the reference is indicated, etc. Assesses the three levels of coverage of sites, accounts and pages, and analyses the page structure in conjunction with LLM.
These signals provide a reference for screening, but are not a substitute for judgement in specific missions. Chen Man has thus set a hierarchy: 4 points of authority and above as a candidate for the main text of the briefing; 3 points for extended reading; and less for the time being. Scores had helped him to narrow the scope of the verification, and key figures and conclusions should still be confirmed from original sources.
In the past, 387 web sites carried his accumulated source-based experience; now, interface scoring provides a basis for a broader and easily adapted initial screening.
After filtering, the content returns.
SearchPro offers two types of summary:passage, standard summary, variable is dynamic summary. According to the product team, the latter would extract the relevant contents of the web page around the issue instead of returning the entire text. In the same article, the speed of collection and progress of searching for orders may vary.
This is different from Chen Van's original process: cleaner.py cleans the entire page before the model looks for the relevant part; dynamic summary provides problem-related content and reduces subsequent processing. He said, "I wrote over 400 lines of code to get close, and now it's a field."
The content acquisition also sets the boundary. Mr. Yuenbao's introduction to the subject, “We don't record it in order to respect each other's intellectual property rights”, provides only what can be seen in public and directly; such direct access is provided through tripartite cooperation.
Finally, the search results are returned to Chen Wan's program, which determines how to continue using it. Ma stated that AI agent would also break down a problem into sub-issues, “searching can affect user experience every second slow”; according to him, the fastest version of the upgrade could return within 300 milliseconds, before it was about 400-500 milliseconds.
Access and control
In addition to the search capability, the mode of access determines its smooth access to the existing workflow.
Tencent cloud standard API: Request sent to wsa.tencentcloudapi.com, Action write SearchPro, Version write 2025-05-08, then sign each request using SecretId and SecretKey as Tencent cloud signature algorithm.
Zheng Nan, a back-end engineer at a data service company, described the time he had spent in connection with his access last year, most of which was in the process of signing. The Tencent cloud subsequently added a single API Key authentication and requested to be sent to the api.wsa.cloud.tencent.com/SearchPro path, a step that can be directly saved.
Another way to get access is faster. Configure the service address and API Key among the client who supports the protocol will add SeachPro to the list of tools, without the need to install additional local services. The only MCP tool currently available is SaarchPro, which uses Streamable HTTP transmissions.
MCP is an open protocol for Anthropic to connect AI applications to external data and tools in November 2024. Over a year later, mainstream AI developed tools to support it. The MCP page of the Yuanbao search states that it uses protocol versions 2025-06-18. Tencent did it by repackaging its search under this protocol.
And up, it's Skill that guides Agent to use tools. On March 10, 2026, Tencent went online to SkyllHub, initially a domestic mirror of OpenClaw Skills Ecology ClawHub, with more than 13,000 skills; by mid-year, more than 80,000 entries were recorded, supporting tools such as WorldBuddy, QClaw and Ima.
SkillHub's method of installation for the "valuable search for Skill" is only one sentence to the effect that Agent installs tencent-yuanbao-standard-search with an installation description for SkillHub. The Skill note lists when it should be used: the user says “help me search” for “search” or “latest” “recent news...”.
API, MCP and Skill are three types of capacity to solve different problems, not three generations of alternative products. API allows the program to directly control the number, timing and retry; MCP solves how compatible clients access the tool; and Skill supplements the instructions to help the assistant judge when and how.
The day after the access, Chen Wan changed a note to the subscriber at the Claude terminal, saying, "Replace last night's latest performance forecast." The terminal then appeared at the SearchPro tool call, which took 681 ms and returned seven results, five of which were authoritative information.
“It went on its own hearing the last two words,” Chen Wan said, “It has become part of the task process.”
A team that does content products has a similar experience. One of the team's developers said at the conference that their original AI search, "It can only be said it works," that Agent often looked at a dozen websites first, and then adjusted them over and over again. After Skill's access search, "fast speed and no need to retrace" the same question was asked for 20 pages, and now 5 can be solved.
Although access has become simpler, it is not as simple as nothing. Keep the key, configure the account number permission and request the frequency according to the call schedule. The interface is part of the foundation work, but the way the application works is still in the hands of the developers.
The other end of access is cost.
According to the current billing documents of Tencent clouds, the Treasure Search provides an extremely fast, standard and flagship version at a rate of $14, $20 and $26 per thousand, respectively, of which the flagship version is open only to business users.
In addition, quota is called free of charge 6000 times a month upon completion of the invitation code, approximately 200 times a day, approximately enough for a small team to conduct daily testing or light use.
The authority field, which Chen Wan adopted this time, is a flagship capacity.
According to the Tencent product team, the beauties are located in the party to be called, without a separate search portal, and the user-oriented access is assumed by each application. It is called by both Tencent's internal product and external AI's assistant and LLM.
In the past year, Chen Man's program has changed several submersible models, from DeepSeek to Qwen, to Zhipu, and finally to DeepSeek. The search interface is the first change. "The model changed, and the quality of the information was almost invisible," he said, "but the search changed, and the results were completely different."
iv. Handed over and left behind.
After the documents were deleted, Chen's search was routinely carried out.
He did a final check. The announcements and industry news coverage on the same day were generally in line with needs, with several more sites previously less involved; no subscribers have written about data sources in the last week.
He maintained the document for two years and was eventually replaced by a search interface, “Nothing but ease”. He said that it was not his code capacity that was poor, but that such code books should not be written by everyone.
Chen Man can now devote more time to verifying key data, comparing industry changes and organizing research judgements.
The Yuanbao search indicates that this call demand is growing rapidly and that two directions are currently considered to be the best. One is Smart Office and Enterprise: After full promotion of an internal product, the number of calls rose from hundreds of thousands in the valley to nearly a million a day. Second, finance and research: “Reporting and reporting information may need to be up to date by hour or even by minutes”, and the recent number of calls by a financial LLM has increased several dozen times with access.
Chen Wan's report paid falls on such demand.
For Chinese Agent developers, such an independent interface would be tantamount to adding a division of labour: the retrieval of information, content extraction and source assessment, which used to be common, would be delivered directly to the public service.
This is where this goes beyond the case. As for the size of the future, it depends on how many applications are built on it, and on whether prices can be reduced to small teams willing to set it up as a default capability.
Chen Wan deleted two documents, but left behind a higher demand for information.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.