返回列表
🪞 Uota学 · 💬 讨论题

网页搜索真正的“提效点”不是搜得更快,而是先把垃圾过滤掉

把“搜索→把一堆网页塞进上下文→再推理”改成“搜索→写代码先过滤→只把相关内容塞进上下文”,这是对 Agent 成本/准确率最有效的结构优化。

2026-02-19
阅读简报
双语对照
完整翻译
原文
讨论归档

核心观点

  • 动态过滤本质是把推理前置成“可执行的筛选器” Claude 在 web search / web fetch 过程中会原生写并执行代码(如 Python)对搜索结果做后处理:提取相关段落、丢弃无关内容,然后再进入上下文窗口。关键不是更聪明的推理,而是更干净的输入。
  • 它解决的是“web 搜索的 token 污染” 传统做法把大量无关 HTML/片段灌进上下文,会直接拉低回答质量;动态过滤在“进入上下文之前”把噪音切掉,从结构上避免污染。
  • 数据:平均少 24% 输入 token,性能平均 +11% 他们在 BrowseComp 与 DeepsearchQA 上对 Sonnet 4.6 / Opus 4.6 做评测:动态过滤平均减少 24% 输入 token,同时把整体表现拉高 11%。BrowseComp 准确率:Sonnet 33.3%→46.6%,Opus 45.3%→61.6%;DeepsearchQA 的 F1:Sonnet 52.6%→59.4%,Opus 69.8%→77.3%。
  • 成本不是无脑下降:Opus 可能更贵 过滤需要写代码、执行代码;他们提到按价格加权的 token 在 Sonnet 上下降,但在 Opus 上可能上升。结论:这是一种“用结构换质量”的手段,不是单纯省钱工具。
  • 一个暗示:未来的 Agent 优化是“pipeline 工程”,不是 prompt 小修小补 搜索/抓取/过滤/抽取/验证 这种编排层的改进,会比改一句 prompt 更决定最终质量。

跟我们的关联

  • 读书管线/行研管线可以直接抄这个思路:先过滤再喂贵模型 现在很多文章/网页内容其实噪音极高(导航、站点菜单、重复段落)。可以加一个“轻量抽取/过滤”步骤,输出干净的正文+关键表格/引用,再交给 opusft 写简报。
  • OpenClaw 的工具层可以内置“结构化抓取器” 与其让模型在上下文里读 HTML,不如工具直接产出:`{title, key_paragraphs, quotes, urls, tables}`。这会显著减少上下文污染。
  • 评测方式值得学:用代表你生产查询的测试集做 AB 他们强调不要只看公开 benchmark,应该用“你自己业务里会问的搜索问题”跑对比,统计准确率/成本/失败率。

讨论引子

  • 动态过滤这件事,在我们的阅读管线里应该落在哪一层?(抓取脚本、tool 层、还是简报前的预处理?)
  • 如果 Opus 变贵但质量显著上升,我们接受的“成本上限”是多少?该怎么用测试集把这条线画清楚?
  • 我们要不要把“引用核验/交叉验证”也做成过滤阶段的一部分(先用代码找证据,再让模型写结论)?

动态过滤让网页搜索更准确高效 | Claude

认识 Claude

产品

功能

模型

平台

解决方案

使用场景

行业

定价

资源

洞见

学习

工具

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

认识 Claude

产品

功能

模型

平台

解决方案

使用场景

行业

定价

资源

洞见

学习

工具

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

联系销售

联系销售联系销售

试用 Claude

试用 Claude试用 Claude

博客

博客

/

通过动态过滤提升网页搜索的准确性与效率

在此探索

通过动态过滤提升网页搜索的准确性与效率

动态过滤让 Claude 在复杂的网页搜索任务上更准确、更高效。下面介绍它的工作原理,以及如何在 API 上启用。

分类

产品公告

产品

Claude 开发者平台

日期

2026 年 2 月 17 日

阅读时间

5

分钟

分享

复制链接

https://claude.com/blog/improved-web-search-with-dynamic-filtering

在发布 Claude Opus 4.6Sonnet 4.6 的同时,我们也推出了新版的 web searchweb fetch 工具。Claude 现在可以在网页搜索过程中原生编写并执行代码,在结果进入上下文窗口之前先进行过滤,从而提升准确性与 token 使用效率。

带动态过滤的网页搜索

网页搜索是一项高度消耗 token 的任务。使用基础网页搜索工具的代理通常需要先发起查询,把搜索结果拉入上下文,再从多个网站抓取完整的 HTML 文件,并在此基础上进行推理后作答。但从搜索中拉入上下文的内容往往与问题无关,这会降低回答质量。

为提升 Claude 的网页搜索表现,我们的 web search 与 web fetch 工具现在会自动编写并执行代码,对查询结果进行后处理。Claude 不再需要对完整 HTML 文件逐一推理,而是可以在加载进上下文之前动态过滤搜索结果,只保留相关内容并丢弃其余部分。

我们此前已发现这种技术在其他代理式工作流中同样有效,并在 API 上加入了诸如代码执行编程式工具调用等工具以提供原生支持。如今,我们也把这些同样的技术带到 web search 与 web fetch 上。

评估 Claude 的网页搜索能力‍

我们在不启用其他工具的前提下,分别在启用与不启用动态过滤两种条件下,对 Sonnet 4.6 与 Opus 4.6 的网页搜索能力进行了评估。在两个基准测试(BrowseCompDeepsearchQA)上,动态过滤在平均减少 24% 输入 token 的同时,将性能平均提升了 11%。

BrowseComp:在网页上搜索以找到一个答案‍

BrowseComp 用来测试代理是否能够在大量网站中导航,找到一条刻意被设计得很难在网上找到的特定信息。动态过滤显著提升了 Claude 的准确率:Sonnet 4.6 从 33.3% 提升到 46.6%,Opus 4.6 从 45.3% 提升到 61.6%。 ‍

DeepsearchQA:在网页上搜索以找到多个答案‍

DeepsearchQA 会向代理提出具有多个正确答案的研究型查询,所有答案都必须通过网页搜索找到。它测试代理能否系统地规划并执行多步骤搜索,同时不遗漏任何答案。其衡量指标是 “F1 分数”,用于平衡精确率与召回率——既反映返回答案的准确性,也体现搜索的完整性。

动态过滤将 Sonnet 4.6 的 F1 分数从 52.6% 提升到 59.4%,将 Opus 4.6 的 F1 分数从 69.8% 提升到 77.3%。

token 成本会随模型为过滤上下文所需编写代码的多少而变化。按价格加权后的 token 数在两个基准上对 Sonnet 4.6 均有所下降,但对 Opus 4.6 则有所上升。为了更好地理解你自身的成本,我们建议用一组能代表你的代理在生产环境中可能遇到的网页搜索查询来评估此工具。

客户聚焦:Quora

QuoraPoe 是最大的多模型 AI 平台之一,让数百万用户通过单一界面访问 200 多个模型。Quora 的内部团队发现,启用动态过滤的 Opus 4.6 “在我们的内部评测中,与其他前沿模型对比测试时达到了最高准确率”,产品与研究负责人 Gareth Jones 表示。“这个模型的行为就像真正的研究员:它会写 Python 来解析、过滤并交叉引用结果,而不是在上下文里对原始 HTML 直接推理。”

web search 与 fetch 工具中的动态过滤

在 Claude API 上使用新版 web search 与 web fetch 工具并选择 Sonnet 4.6 或 Opus 4.6 时,动态过滤将默认开启。对于复杂的网页搜索查询(例如筛查技术文档或核验引用来源),你可以期待获得与上文所示类似的性能提升。

下面是在 API 中的使用方式:

{
  model: claude-opus-4-6,
  max_tokens: 4096,
  tools: [
    {
      type: web_search_20260209,
      name: web_search
    },
    {
      type: web_fetch_20260209,
      name: web_fetch
    }
  ],
  messages: [
    {
      role: user,
      content: Search for the current prices of AAPL and GOOGL, then calculate which has a better P/E ratio.
    }
  ]
}

代码执行、记忆等更多工具现已全面可用

我们还将多项工具提升为全面可用(GA),以帮助代理在高 token 消耗任务中表现更佳:

  • 代码执行:为代理提供一个沙箱,使其能在对话过程中运行代码以过滤上下文、分析数据或进行计算。

  • 记忆:通过持久化文件目录跨对话存取信息,让代理无需把所有内容都留在上下文窗口中也能保留背景。

  • 编程式工具调用:以代码执行复杂的多工具工作流,把中间结果留在上下文窗口之外。

  • 工具搜索:从大型工具库中动态发现工具,而无需将所有定义都加载到上下文窗口中。

  • 工具使用示例:在工具定义中直接提供示例调用,以展示用法模式并减少参数错误。

快速开始

改进后的 web search 与 web fetch——以及代码执行、记忆、编程式工具调用、工具搜索与工具使用示例——现已在 Claude 开发者平台上线。阅读我们的 API 文档 即可开始使用。

未找到任何内容。

上一页上一页

0/5

下一页下一页

电子书

常见问题(FAQ)

未找到任何内容。

相关文章

探索更多产品动态与团队使用 Claude 的最佳实践。

2026 年 1 月 12 日

Cowork:让 Claude Code 覆盖你其余的工作

产品公告

Cowork:让 Claude Code 覆盖你其余的工作Cowork:让 Claude Code 覆盖你其余的工作

Cowork:让 Claude Code 覆盖你其余的工作Cowork:让 Claude Code 覆盖你其余的工作

2026 年 1 月 26 日

你最常用的工作工具现在已成为 Claude 内可交互的连接器

产品公告

你最常用的工作工具现在已成为 Claude 内可交互的连接器你最常用的工作工具现在已成为 Claude 内可交互的连接器

你最常用的工作工具现在已成为 Claude 内可交互的连接器你最常用的工作工具现在已成为 Claude 内可交互的连接器

2025 年 12 月 8 日

Claude Code 与 Slack

产品公告

Claude Code 与 SlackClaude Code 与 Slack

Claude Code 与 SlackClaude Code 与 Slack

2025 年 8 月 25 日

在 Chrome 中试点 Claude

产品公告

在 Chrome 中试点 Claude在 Chrome 中试点 Claude

在 Chrome 中试点 Claude在 Chrome 中试点 Claude

用 Claude 改变你的组织运作方式

查看定价

查看定价查看定价

联系销售

联系销售联系销售

获取开发者通讯

产品更新、操作指南、社区聚焦,等等。每月发送至你的收件箱。

订阅订阅

如果你希望接收我们的每月开发者通讯,请提供你的电子邮箱地址。你可随时取消订阅。

谢谢!你已订阅。

抱歉,提交时出现问题,请稍后再试。

主页主页

下一页下一页

谢谢!已收到你的提交!

哎呀!提交表单时出了点问题。

写作

按钮文字按钮文字

学习

按钮文字按钮文字

代码

按钮文字按钮文字

写作

帮我为受众打造独特的写作声音

嗨,Claude!你能帮我为某个受众打造独特的写作声音吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

改进我的写作风格

嗨,Claude!你能改进我的写作风格吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

头脑风暴创意点子

嗨,Claude!你能帮我头脑风暴一些创意点子吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

学习

用简单方式解释复杂主题

嗨,Claude!你能用简单的方式解释一个复杂主题吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

帮我理清这些想法

嗨,Claude!你能帮我理清这些想法吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

备考或面试准备

嗨,Claude!你能帮我备考或准备面试吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

代码

解释一个编程概念

嗨,Claude!你能解释一个编程概念吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

帮我看看代码并给建议

嗨,Claude!你能帮我看看我的代码并给一些建议吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

和我一起 Vibe Coding

嗨,Claude!你能和我一起 vibe code 吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

更多

撰写案例研究

这是另一个测试

撰写拨款/资助申请

嗨,Claude!你能帮我撰写拨款/资助申请吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

撰写视频脚本

这是个测试

AnthropicAnthropic

© [year] Anthropic PBC

产品

Claude

ClaudeClaude

Claude Code

Claude CodeClaude Code

Cowork

CoworkCowork

Max 方案

Max planMax 方案

团队方案

Team plan团队方案

企业方案

Enterprise plan企业方案

下载应用

Download app下载应用

定价

Pricing定价

登录

Log in登录

功能

Chrome 中的 Claude

Claude in ChromeChrome 中的 Claude

Slack 中的 Claude

Claude in SlackSlack 中的 Claude

Excel 中的 Claude

Claude in ExcelExcel 中的 Claude

PowerPoint 中的 Claude

Claude in PowerPointPowerPoint 中的 Claude

Skills

SkillsSkills

模型

Opus

OpusOpus

Sonnet

SonnetSonnet

Haiku

HaikuHaiku

解决方案

AI 代理

AI agentsAI 代理

代码现代化

Code modernization代码现代化

编程

Coding编程

客户支持

Customer support客户支持

教育

Education教育

金融服务

Financial services金融服务

政府

Government政府

医疗健康

Healthcare医疗健康

生命科学

Life sciences生命科学

非营利组织

Nonprofits非营利组织

Claude 开发者平台

概览

Overview概览

开发者文档

Developer docs开发者文档

定价

Pricing定价

区域合规

Regional compliance区域合规

Amazon Bedrock

Amazon BedrockAmazon Bedrock

Google Cloud 的 Vertex AI

Google Cloud’s Vertex AIGoogle Cloud 的 Vertex AI

控制台登录

Console login控制台登录

学习

博客

Blog博客

Claude 合作伙伴网络

Claude partner networkClaude 合作伙伴网络

课程

Courses课程

连接器

Connectors连接器

客户案例

Customer stories客户案例

Anthropic 工程团队

Engineering at AnthropicAnthropic 工程团队

活动

Events活动

插件

Plugins插件

Powered by Claude

Powered by ClaudePowered by Claude

服务合作伙伴

Service partners服务合作伙伴

初创企业项目

Startups program初创企业项目

教程

Tutorials教程

使用场景

Use cases使用场景

公司

Anthropic

AnthropicAnthropic

招聘

Careers招聘

经济未来

Economic Futures经济未来

研究

Research研究

新闻

News新闻

负责任规模化政策

Responsible Scaling Policy负责任规模化政策

安全与合规

Security and compliance安全与合规

透明度

Transparency透明度

帮助与安全

可用性

Availability可用性

状态

Status状态

支持中心

Support center支持中心

条款与政策

隐私选择

Cookie 设置

我们使用 Cookie 来交付并改进服务、分析站点使用情况,并在你同意的情况下,为你定制或个性化体验并向你推广服务。你可以在此处阅读我们的 Cookie 政策。

自定义 Cookie 设置 拒绝所有 Cookie 接受所有 Cookie

必要

启用安全与基础功能。

必需

分析

启用站点性能追踪。

关闭

营销

启用广告个性化与追踪。

关闭

保存偏好设置

隐私政策

隐私政策隐私政策

负责任披露政策

负责任披露政策负责任披露政策

服务条款:商业

服务条款:商业服务条款:商业

服务条款:消费者

服务条款:消费者服务条款:消费者

使用政策

使用政策使用政策

x.comx.com

LinkedInLinkedIn

YouTubeYouTube

InstagramInstagram

英语(美国)

英语(美国)

日语(日本)

德语(德国)

法语(法国)

韩语(韩国)

Claude 开发者平台

代理

相关笔记

Improved Web Search with Dynamic Filtering | Claude

-

-

认识 Claude

Meet Claude

产品

Products

功能

模型

Models

-

平台

Platform

-

解决方案

Solutions

使用场景

Use cases

行业

Industries

-

定价

Pricing

-

资源

Resources

洞见

Insights

Learn

工具

Tools

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

认识 Claude

Meet Claude

产品

Products

功能

模型

Models

-

平台

Platform

-

解决方案

Solutions

使用场景

Use cases

行业

Industries

-

定价

Pricing

-

资源

Resources

洞见

Insights

Learn

工具

Tools

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

-

-

试用 Claude

Try Claude

试用 Claude试用 Claude

Try ClaudeTry Claude

-

-

博客

Blog

/

-

-

通过动态过滤提升网页搜索的准确性与效率

Increase web search accuracy and efficiency with dynamic filtering

在此探索

Explore here

通过动态过滤提升网页搜索的准确性与效率

Increase web search accuracy and efficiency with dynamic filtering

动态过滤让 Claude 在复杂的网页搜索任务上更准确、更高效。下面介绍它的工作原理,以及如何在 API 上启用。

Dynamic filtering makes Claude more accurate and efficient on complex web search tasks. Here’s how it works, and how to enable it on the API.

-

-

分类

Category

-

产品

Product

Claude 开发者平台

Claude Developer Platform

-

-

日期

Date

2026 年 2 月 17 日

February 17, 2026

-

-

阅读时间

Reading time

5

5

分钟

min

-

-

分享

https://claude.com/blog/improved-web-search-with-dynamic-filtering

https://claude.com/blog/improved-web-search-with-dynamic-filtering

在发布 Claude Opus 4.6Sonnet 4.6 的同时,我们也推出了新版的 web searchweb fetch 工具。Claude 现在可以在网页搜索过程中原生编写并执行代码,在结果进入上下文窗口之前先进行过滤,从而提升准确性与 token 使用效率。

Alongside Claude Opus 4.6 and Sonnet 4.6, we’re releasing new versions of our web search and web fetch tools. Claude can now natively write and execute code during web searches to filter results before they reach the context window, improving its accuracy and token efficiency.

带动态过滤的网页搜索

Web search with dynamic filtering

网页搜索是一项高度消耗 token 的任务。使用基础网页搜索工具的代理通常需要先发起查询,把搜索结果拉入上下文,再从多个网站抓取完整的 HTML 文件,并在此基础上进行推理后作答。但从搜索中拉入上下文的内容往往与问题无关,这会降低回答质量。

Web search is a highly token-intensive task. Agents using basic web search tools need to make a query, pull search results into context, fetch full HTML files from multiple websites, and reason over it all before responding. But the context being pulled in from search is often irrelevant, which degrades the quality of the response.

为提升 Claude 的网页搜索表现,我们的 web search 与 web fetch 工具现在会自动编写并执行代码,对查询结果进行后处理。Claude 不再需要对完整 HTML 文件逐一推理,而是可以在加载进上下文之前动态过滤搜索结果,只保留相关内容并丢弃其余部分。

To improve Claude’s performance on web searches, our web search and web fetch tools now automatically write and execute code to post-process query results. Instead of reasoning over full HTML files, Claude can dynamically filter the search results before loading them into context, keeping only what’s relevant and discarding the rest.

我们此前已发现这种技术在其他代理式工作流中同样有效,并在 API 上加入了诸如代码执行编程式工具调用等工具以提供原生支持。如今,我们也把这些同样的技术带到 web search 与 web fetch 上。

We’ve previously foundthis technique to be effective across other agentic workflows, and we’ve added tools such as code execution and programmatic tool calling for native support on our API. We’re now bringing these same techniques to web search and web fetch.

评估 Claude 的网页搜索能力‍

Evaluating Claude’s ability to search the web‍

我们在不启用其他工具的前提下,分别在启用与不启用动态过滤两种条件下,对 Sonnet 4.6 与 Opus 4.6 的网页搜索能力进行了评估。在两个基准测试(BrowseCompDeepsearchQA)上,动态过滤在平均减少 24% 输入 token 的同时,将性能平均提升了 11%。

We evaluated web search on Sonnet 4.6 and Opus 4.6 with and without dynamic filtering and no other tools enabled. Across two benchmarks, BrowseComp and DeepsearchQA, dynamic filtering improved performance by an average of 11% while using 24% fewer input tokens.

BrowseComp:在网页上搜索以找到一个答案‍

BrowseComp: Searching the web to find one answer‍

BrowseComp 用来测试代理是否能够在大量网站中导航,找到一条刻意被设计得很难在网上找到的特定信息。动态过滤显著提升了 Claude 的准确率:Sonnet 4.6 从 33.3% 提升到 46.6%,Opus 4.6 从 45.3% 提升到 61.6%。 ‍

BrowseComp tests whether an agent can navigate many websites to find a specific piece of information that is deliberately hard to find online. Dynamic filtering improved Claude’s accuracy significantly, bringing Sonnet 4.6 from 33.3% to 46.6% and Opus 4.6 from 45.3% to 61.6%. ‍

DeepsearchQA:在网页上搜索以找到多个答案‍

DeepsearchQA: Searching the web to find many answers‍

DeepsearchQA 会向代理提出具有多个正确答案的研究型查询,所有答案都必须通过网页搜索找到。它测试代理能否系统地规划并执行多步骤搜索,同时不遗漏任何答案。其衡量指标是 “F1 分数”,用于平衡精确率与召回率——既反映返回答案的准确性,也体现搜索的完整性。

DeepsearchQA presents agents with research queries that have many correct answers, all of which must be found via web search. It tests whether an agent can systematically plan and execute multi-step searches without missing any answers. It’s measured by an “F1 score,” which balances precision and recall—capturing both the accuracy of returned answers and the completeness of the search.

动态过滤将 Sonnet 4.6 的 F1 分数从 52.6% 提升到 59.4%,将 Opus 4.6 的 F1 分数从 69.8% 提升到 77.3%。

Dynamic filtering improved Claude’s F1 score from 52.6% to 59.4% for Sonnet 4.6 and from 69.8% to 77.3% for Opus 4.6.

token 成本会随模型为过滤上下文所需编写代码的多少而变化。按价格加权后的 token 数在两个基准上对 Sonnet 4.6 均有所下降,但对 Opus 4.6 则有所上升。为了更好地理解你自身的成本,我们建议用一组能代表你的代理在生产环境中可能遇到的网页搜索查询来评估此工具。

Token costs will vary depending on how much code the model needs to write to filter context. Price-weighted tokens decreased for Sonnet 4.6 on both benchmarks but increased for Opus 4.6. To better understand your own costs, we recommend evaluating this tool against a representative set of web search queries your agent is likely to encounter in production.

客户聚焦:Quora

Customer spotlight: Quora

QuoraPoe 是最大的多模型 AI 平台之一,让数百万用户通过单一界面访问 200 多个模型。Quora 的内部团队发现,启用动态过滤的 Opus 4.6 “在我们的内部评测中,与其他前沿模型对比测试时达到了最高准确率”,产品与研究负责人 Gareth Jones 表示。“这个模型的行为就像真正的研究员:它会写 Python 来解析、过滤并交叉引用结果,而不是在上下文里对原始 HTML 直接推理。”

Poe by Quora is one of the largest multi-model AI platforms, giving millions of users access to over 200 models through a single interface. Internal teams at Quora found that Opus 4.6 with dynamic filtering “achieved the highest accuracy on our internal evals when tested against other frontier models,” said Gareth Jones, Product and Research Lead. “The model behaves like an actual researcher, writing Python to parse, filter, and cross-reference results rather than reasoning over raw HTML in context.”

web search 与 fetch 工具中的动态过滤

Dynamic filtering in the web search and fetch tools

在 Claude API 上使用新版 web search 与 web fetch 工具并选择 Sonnet 4.6 或 Opus 4.6 时,动态过滤将默认开启。对于复杂的网页搜索查询(例如筛查技术文档或核验引用来源),你可以期待获得与上文所示类似的性能提升。

Dynamic filtering will be turned on by default when using our new web search and web fetch tools with Sonnet 4.6 and Opus 4.6 on the Claude API. For complex web search queries, such as sifting through technical documentation or verifying citations, you can expect similar performance improvements to those shown above.

下面是在 API 中的使用方式:

Here’s how to use it in the API:

{
  model: claude-opus-4-6,
  max_tokens: 4096,
  tools: [
    {
      type: web_search_20260209,
      name: web_search
    },
    {
      type: web_fetch_20260209,
      name: web_fetch
    }
  ],
  messages: [
    {
      role: user,
      content: Search for the current prices of AAPL and GOOGL, then calculate which has a better P/E ratio.
    }
  ]
}

{
  model: claude-opus-4-6,
  max_tokens: 4096,
  tools: [
    {
      type: web_search_20260209,
      name: web_search
    },
    {
      type: web_fetch_20260209,
      name: web_fetch
    }
  ],
  messages: [
    {
      role: user,
      content: Search for the current prices of AAPL and GOOGL, then calculate which has a better P/E ratio.
    }
  ]
}

代码执行、记忆等更多工具现已全面可用

Code execution, memory, and more tools are now generally available

我们还将多项工具提升为全面可用(GA),以帮助代理在高 token 消耗任务中表现更佳:

We’re also graduating several tools to general availability to help agents perform better across token-intensive tasks:

  • 代码执行:为代理提供一个沙箱,使其能在对话过程中运行代码以过滤上下文、分析数据或进行计算。
  • Code execution: Provides a sandbox for agents to run code during a conversation to filter context, analyze data, or perform calculations.
  • 记忆:通过持久化文件目录跨对话存取信息,让代理无需把所有内容都留在上下文窗口中也能保留背景。
  • Memory: Store and retrieve information across conversations through a persistent file directory, so agents can retain context without keeping everything in the context window.
  • 编程式工具调用:以代码执行复杂的多工具工作流,把中间结果留在上下文窗口之外。
  • Programmatic tool calling: Execute complex multi-tool workflows in code, keeping intermediate results out of the context window.
  • 工具搜索:从大型工具库中动态发现工具,而无需将所有定义都加载到上下文窗口中。
  • Tool search: Dynamically discover tools from large libraries without loading all definitions into the context window.
  • 工具使用示例:在工具定义中直接提供示例调用,以展示用法模式并减少参数错误。
  • Tool use examples: Provide sample tool calls directly in your tool definitions to demonstrate usage patterns and reduce parameter errors.

快速开始

Getting started

改进后的 web search 与 web fetch——以及代码执行、记忆、编程式工具调用、工具搜索与工具使用示例——现已在 Claude 开发者平台上线。阅读我们的 API 文档 即可开始使用。

Improved web search and web fetch—as well as code execution, memory, programmatic tool calling, tool search, and tool use examples—are available now on the Claude Developer Platform. Read our API documentation to get started.

未找到任何内容。

No items found.

上一页上一页

PrevPrev

0/5

0/5

下一页下一页

NextNext

电子书

eBook

常见问题(FAQ)

FAQ

未找到任何内容。

No items found.

相关文章

Related posts

探索更多产品动态与团队使用 Claude 的最佳实践。

Explore more product news and best practices for teams building with Claude.

2026 年 1 月 12 日

Jan 12, 2026

Cowork:让 Claude Code 覆盖你其余的工作

Cowork: Claude Code for the rest of your work

产品公告

Product announcements

Cowork:让 Claude Code 覆盖你其余的工作Cowork:让 Claude Code 覆盖你其余的工作

Cowork: Claude Code for the rest of your workCowork: Claude Code for the rest of your work

Cowork:让 Claude Code 覆盖你其余的工作Cowork:让 Claude Code 覆盖你其余的工作

Cowork: Claude Code for the rest of your workCowork: Claude Code for the rest of your work

2026 年 1 月 26 日

Jan 26, 2026

你最常用的工作工具现在已成为 Claude 内可交互的连接器

Your favorite work tools are now interactive connectors inside Claude

产品公告

Product announcements

你最常用的工作工具现在已成为 Claude 内可交互的连接器你最常用的工作工具现在已成为 Claude 内可交互的连接器

Your favorite work tools are now interactive connectors inside ClaudeYour favorite work tools are now interactive connectors inside Claude

你最常用的工作工具现在已成为 Claude 内可交互的连接器你最常用的工作工具现在已成为 Claude 内可交互的连接器

Your favorite work tools are now interactive connectors inside ClaudeYour favorite work tools are now interactive connectors inside Claude

2025 年 12 月 8 日

Dec 8, 2025

Claude Code 与 Slack

Claude Code and Slack

产品公告

Product announcements

Claude Code 与 SlackClaude Code 与 Slack

Claude Code and SlackClaude Code and Slack

Claude Code 与 SlackClaude Code 与 Slack

Claude Code and SlackClaude Code and Slack

2025 年 8 月 25 日

Aug 25, 2025

在 Chrome 中试点 Claude

Piloting Claude in Chrome

产品公告

Product announcements

在 Chrome 中试点 Claude在 Chrome 中试点 Claude

Piloting Claude in ChromePiloting Claude in Chrome

在 Chrome 中试点 Claude在 Chrome 中试点 Claude

Piloting Claude in ChromePiloting Claude in Chrome

用 Claude 改变你的组织运作方式

Transform how your organization operates with Claude

查看定价

See pricing

查看定价查看定价

See pricingSee pricing

联系销售

Contact sales

联系销售联系销售

Contact salesContact sales

获取开发者通讯

Get the developer newsletter

产品更新、操作指南、社区聚焦,等等。每月发送至你的收件箱。

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

订阅订阅

SubscribeSubscribe

如果你希望接收我们的每月开发者通讯,请提供你的电子邮箱地址。你可随时取消订阅。

Please provide your email address if youd like to receive our monthly developer newsletter. You can unsubscribe at any time.

谢谢!你已订阅。

Thank you! You’re subscribed.

抱歉,提交时出现问题,请稍后再试。

Sorry, there was a problem with your submission, please try again later.

主页主页

HomepageHomepage

下一页下一页

NextNext

谢谢!已收到你的提交!

Thank you! Your submission has been received!

哎呀!提交表单时出了点问题。

Oops! Something went wrong while submitting the form.

写作

Write

按钮文字按钮文字

Button TextButton Text

学习

Learn

按钮文字按钮文字

Button TextButton Text

代码

Code

按钮文字按钮文字

Button TextButton Text

写作

Write

-

-

帮我为受众打造独特的写作声音

Help me develop a unique voice for an audience

嗨,Claude!你能帮我为某个受众打造独特的写作声音吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you help me develop a unique voice for an audience? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

改进我的写作风格

Improve my writing style

嗨,Claude!你能改进我的写作风格吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you improve my writing style? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

头脑风暴创意点子

Brainstorm creative ideas

嗨,Claude!你能帮我头脑风暴一些创意点子吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you brainstorm creative ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

学习

Learn

-

-

用简单方式解释复杂主题

Explain a complex topic simply

嗨,Claude!你能用简单的方式解释一个复杂主题吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you explain a complex topic simply? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

帮我理清这些想法

Help me make sense of these ideas

嗨,Claude!你能帮我理清这些想法吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you help me make sense of these ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

备考或面试准备

Prepare for an exam or interview

嗨,Claude!你能帮我备考或准备面试吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you prepare for an exam or interview? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

代码

Code

-

-

解释一个编程概念

Explain a programming concept

嗨,Claude!你能解释一个编程概念吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you explain a programming concept? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

帮我看看代码并给建议

Look over my code and give me tips

嗨,Claude!你能帮我看看我的代码并给一些建议吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you look over my code and give me tips? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

和我一起 Vibe Coding

Vibe code with me

嗨,Claude!你能和我一起 vibe code 吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you vibe code with me? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

更多

More

-

-

撰写案例研究

Write case studies

这是另一个测试

This is another test

-

-

撰写拨款/资助申请

Write grant proposals

嗨,Claude!你能帮我撰写拨款/资助申请吗?如果你需要我提供更多信息,请立刻问我 1-2 个关键问题。如果你觉得我上传一些文档能帮助你做得更好,也请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——只要它们能帮助你更好地完成这个任务。不要使用分析工具。请让你的回复友好、简短、对话式。

Hi Claude! Could you write grant proposals? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to — like Google Drive, web search, etc. — if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

请尽快执行任务——如果合适的话,给我一个可交付成果(artifact)会很棒。如果使用 artifact,请考虑什么类型(交互式、可视化、清单等)对这个具体任务最有帮助。谢谢你的帮助!

Please execute the task as soon as you can - an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

-

-

撰写视频脚本

Write video scripts

这是个测试

this is a test

AnthropicAnthropic

AnthropicAnthropic

© [year] Anthropic PBC

© [year] Anthropic PBC

产品

Products

-

-

Claude

Claude

ClaudeClaude

ClaudeClaude

-

-

Claude Code

Claude Code

Claude CodeClaude Code

Claude CodeClaude Code

-

-

Cowork

Cowork

CoworkCowork

CoworkCowork

-

-

Max 方案

Max plan

Max planMax 方案

Max planMax plan

-

-

团队方案

Team plan

Team plan团队方案

Team planTeam plan

-

-

企业方案

Enterprise plan

Enterprise plan企业方案

Enterprise planEnterprise plan

-

-

下载应用

Download app

Download app下载应用

Download appDownload app

-

-

定价

Pricing

Pricing定价

PricingPricing

-

-

登录

Log in

Log in登录

Log inLog in

功能

Features

-

-

Chrome 中的 Claude

Claude in Chrome

Claude in ChromeChrome 中的 Claude

Claude in ChromeClaude in Chrome

-

-

Slack 中的 Claude

Claude in Slack

Claude in SlackSlack 中的 Claude

Claude in SlackClaude in Slack

-

-

Excel 中的 Claude

Claude in Excel

Claude in ExcelExcel 中的 Claude

Claude in ExcelClaude in Excel

-

-

PowerPoint 中的 Claude

Claude in PowerPoint

Claude in PowerPointPowerPoint 中的 Claude

Claude in PowerPointClaude in PowerPoint

-

-

Skills

Skills

SkillsSkills

SkillsSkills

模型

Models

-

-

Opus

Opus

OpusOpus

OpusOpus

-

-

Sonnet

Sonnet

SonnetSonnet

SonnetSonnet

-

-

Haiku

Haiku

HaikuHaiku

HaikuHaiku

解决方案

Solutions

-

-

AI 代理

AI agents

AI agentsAI 代理

AI agentsAI agents

-

-

代码现代化

Code modernization

Code modernization代码现代化

Code modernizationCode modernization

-

-

编程

Coding

Coding编程

CodingCoding

-

-

客户支持

Customer support

Customer support客户支持

Customer supportCustomer support

-

-

教育

Education

Education教育

EducationEducation

-

-

金融服务

Financial services

Financial services金融服务

Financial servicesFinancial services

-

-

政府

Government

GovernmentGovernment

-

-

医疗健康

Healthcare

Healthcare医疗健康

HealthcareHealthcare

-

-

生命科学

Life sciences

Life sciences生命科学

Life sciencesLife sciences

-

-

非营利组织

Nonprofits

Nonprofits非营利组织

NonprofitsNonprofits

Claude 开发者平台

Claude Developer Platform

-

-

概览

Overview

Overview概览

OverviewOverview

-

-

开发者文档

Developer docs

Developer docs开发者文档

Developer docsDeveloper docs

-

-

定价

Pricing

Pricing定价

PricingPricing

-

-

区域合规

Regional compliance

Regional complianceRegional compliance

-

-

Amazon Bedrock

Amazon Bedrock

Amazon BedrockAmazon Bedrock

Amazon BedrockAmazon Bedrock

-

-

Google Cloud 的 Vertex AI

Google Cloud’s Vertex AI

Google Cloud’s Vertex AIGoogle Cloud 的 Vertex AI

Google Cloud’s Vertex AIGoogle Cloud’s Vertex AI

-

-

控制台登录

Console login

Console login控制台登录

Console loginConsole login

学习

Learn

-

-

博客

Blog

Blog博客

BlogBlog

-

-

Claude 合作伙伴网络

Claude partner network

Claude partner networkClaude 合作伙伴网络

Claude partner networkClaude partner network

-

-

课程

Courses

Courses课程

CoursesCourses

-

-

连接器

Connectors

Connectors连接器

ConnectorsConnectors

-

-

客户案例

Customer stories

Customer stories客户案例

Customer storiesCustomer stories

-

-

Anthropic 工程团队

Engineering at Anthropic

Engineering at AnthropicAnthropic 工程团队

Engineering at AnthropicEngineering at Anthropic

-

-

活动

Events

Events活动

EventsEvents

-

-

插件

Plugins

Plugins插件

PluginsPlugins

-

-

Powered by Claude

Powered by Claude

Powered by ClaudePowered by Claude

Powered by ClaudePowered by Claude

-

-

服务合作伙伴

Service partners

Service partners服务合作伙伴

Service partnersService partners

-

-

初创企业项目

Startups program

Startups program初创企业项目

Startups programStartups program

-

-

教程

Tutorials

Tutorials教程

TutorialsTutorials

-

-

使用场景

Use cases

Use cases使用场景

Use casesUse cases

公司

Company

-

-

Anthropic

Anthropic

AnthropicAnthropic

AnthropicAnthropic

-

-

招聘

Careers

Careers招聘

CareersCareers

-

-

经济未来

Economic Futures

Economic Futures经济未来

Economic FuturesEconomic Futures

-

-

研究

Research

Research研究

ResearchResearch

-

-

新闻

News

News新闻

NewsNews

-

-

负责任规模化政策

Responsible Scaling Policy

Responsible Scaling Policy负责任规模化政策

Responsible Scaling PolicyResponsible Scaling Policy

-

-

安全与合规

Security and compliance

Security and complianceSecurity and compliance

-

-

透明度

Transparency

Transparency透明度

TransparencyTransparency

帮助与安全

Help and security

-

-

可用性

Availability

Availability可用性

AvailabilityAvailability

-

-

状态

Status

Status状态

StatusStatus

-

-

支持中心

Support center

Support center支持中心

Support centerSupport center

条款与政策

Terms and policies

-

-

隐私选择

Privacy choices

Cookie 设置

Cookie settings

我们使用 Cookie 来交付并改进服务、分析站点使用情况,并在你同意的情况下,为你定制或个性化体验并向你推广服务。你可以在此处阅读我们的 Cookie 政策。

We use cookies to deliver and improve our services, analyze site usage, and if you agree, to customize or personalize your experience and market our services to you. You can read our Cookie Policy here.

自定义 Cookie 设置 拒绝所有 Cookie 接受所有 Cookie

Customize cookie settings Reject all cookies Accept all cookies

必要

Necessary

启用安全与基础功能。

Enables security and basic functionality.

必需

Required

分析

Analytics

启用站点性能追踪。

Enables tracking of site performance.

关闭

Off

营销

Marketing

启用广告个性化与追踪。

Enables ads personalization and tracking.

关闭

Off

保存偏好设置

Save preferences

-

-

隐私政策

Privacy policy

隐私政策隐私政策

Privacy policyPrivacy policy

-

-

负责任披露政策

Responsible disclosure policy

负责任披露政策负责任披露政策

Responsible disclosure policyResponsible disclosure policy

-

-

服务条款:商业

Terms of service: Commercial

服务条款:商业服务条款:商业

Terms of service: CommercialTerms of service: Commercial

-

-

服务条款:消费者

Terms of service: Consumer

服务条款:消费者服务条款:消费者

Terms of service: ConsumerTerms of service: Consumer

-

-

使用政策

Usage policy

使用政策使用政策

Usage policyUsage policy

x.comx.com

x.comx.com

LinkedInLinkedIn

LinkedInLinkedIn

YouTubeYouTube

YouTubeYouTube

InstagramInstagram

InstagramInstagram

英语(美国)

English (US)

Claude 开发者平台

Claude Developer Platform

代理

Agents

相关笔记

Improved Web Search with Dynamic Filtering | Claude

  • Source: https://claude.com/blog/improved-web-search-with-dynamic-filtering
  • Mirror: https://claude.com/blog/improved-web-search-with-dynamic-filtering
  • Published:
  • Saved: 2026-02-19

Content

Improved Web Search with Dynamic Filtering | Claude

Meet Claude

Products

Features

Models

Platform

Solutions

Use cases

Industries

Pricing

Resources

Insights

Learn

Tools

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Meet Claude

Products

Features

Models

Platform

Solutions

Use cases

Industries

Pricing

Resources

Insights

Learn

Tools

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Contact sales

Contact salesContact sales

Try Claude

Try ClaudeTry Claude

Blog

Blog

/

Increase web search accuracy and efficiency with dynamic filtering

Explore here

Increase web search accuracy and efficiency with dynamic filtering

Dynamic filtering makes Claude more accurate and efficient on complex web search tasks. Here’s how it works, and how to enable it on the API.

Category

Product announcements

Product

Claude Developer Platform

Date

February 17, 2026

Reading time

5

min

Share

Copy link

https://claude.com/blog/improved-web-search-with-dynamic-filtering

Alongside Claude Opus 4.6 and Sonnet 4.6, we’re releasing new versions of our web search and web fetch tools. Claude can now natively write and execute code during web searches to filter results before they reach the context window, improving its accuracy and token efficiency.

Web search with dynamic filtering

Web search is a highly token-intensive task. Agents using basic web search tools need to make a query, pull search results into context, fetch full HTML files from multiple websites, and reason over it all before responding. But the context being pulled in from search is often irrelevant, which degrades the quality of the response.

To improve Claude’s performance on web searches, our web search and web fetch tools now automatically write and execute code to post-process query results. Instead of reasoning over full HTML files, Claude can dynamically filter the search results before loading them into context, keeping only what’s relevant and discarding the rest.

We’ve previously foundthis technique to be effective across other agentic workflows, and we’ve added tools such as code execution and programmatic tool calling for native support on our API. We’re now bringing these same techniques to web search and web fetch.

Evaluating Claude’s ability to search the web‍

We evaluated web search on Sonnet 4.6 and Opus 4.6 with and without dynamic filtering and no other tools enabled. Across two benchmarks, BrowseComp and DeepsearchQA, dynamic filtering improved performance by an average of 11% while using 24% fewer input tokens.

BrowseComp: Searching the web to find one answer‍

BrowseComp tests whether an agent can navigate many websites to find a specific piece of information that is deliberately hard to find online. Dynamic filtering improved Claude’s accuracy significantly, bringing Sonnet 4.6 from 33.3% to 46.6% and Opus 4.6 from 45.3% to 61.6%. ‍

DeepsearchQA: Searching the web to find many answers‍

DeepsearchQA presents agents with research queries that have many correct answers, all of which must be found via web search. It tests whether an agent can systematically plan and execute multi-step searches without missing any answers. It’s measured by an “F1 score,” which balances precision and recall—capturing both the accuracy of returned answers and the completeness of the search.

Dynamic filtering improved Claude’s F1 score from 52.6% to 59.4% for Sonnet 4.6 and from 69.8% to 77.3% for Opus 4.6.

Token costs will vary depending on how much code the model needs to write to filter context. Price-weighted tokens decreased for Sonnet 4.6 on both benchmarks but increased for Opus 4.6. To better understand your own costs, we recommend evaluating this tool against a representative set of web search queries your agent is likely to encounter in production.

Customer spotlight: Quora

Poe by Quora is one of the largest multi-model AI platforms, giving millions of users access to over 200 models through a single interface. Internal teams at Quora found that Opus 4.6 with dynamic filtering “achieved the highest accuracy on our internal evals when tested against other frontier models,” said Gareth Jones, Product and Research Lead. “The model behaves like an actual researcher, writing Python to parse, filter, and cross-reference results rather than reasoning over raw HTML in context.”

Dynamic filtering in the web search and fetch tools

Dynamic filtering will be turned on by default when using our new web search and web fetch tools with Sonnet 4.6 and Opus 4.6 on the Claude API. For complex web search queries, such as sifting through technical documentation or verifying citations, you can expect similar performance improvements to those shown above.

Here’s how to use it in the API:

{
  model: claude-opus-4-6,
  max_tokens: 4096,
  tools: [
    {
      type: web_search_20260209,
      name: web_search
    },
    {
      type: web_fetch_20260209,
      name: web_fetch
    }
  ],
  messages: [
    {
      role: user,
      content: Search for the current prices of AAPL and GOOGL, then calculate which has a better P/E ratio.
    }
  ]
}

Code execution, memory, and more tools are now generally available

We’re also graduating several tools to general availability to help agents perform better across token-intensive tasks:

  • Code execution: Provides a sandbox for agents to run code during a conversation to filter context, analyze data, or perform calculations.

  • Memory: Store and retrieve information across conversations through a persistent file directory, so agents can retain context without keeping everything in the context window.

  • Programmatic tool calling: Execute complex multi-tool workflows in code, keeping intermediate results out of the context window.

  • Tool search: Dynamically discover tools from large libraries without loading all definitions into the context window.

  • Tool use examples: Provide sample tool calls directly in your tool definitions to demonstrate usage patterns and reduce parameter errors.

Getting started

Improved web search and web fetch—as well as code execution, memory, programmatic tool calling, tool search, and tool use examples—are available now on the Claude Developer Platform. Read our API documentation to get started.

No items found.

PrevPrev

0/5

NextNext

eBook

FAQ

No items found.

Related posts

Explore more product news and best practices for teams building with Claude.

Jan 12, 2026

Cowork: Claude Code for the rest of your work

Product announcements

Cowork: Claude Code for the rest of your workCowork: Claude Code for the rest of your work

Cowork: Claude Code for the rest of your workCowork: Claude Code for the rest of your work

Jan 26, 2026

Your favorite work tools are now interactive connectors inside Claude

Product announcements

Your favorite work tools are now interactive connectors inside ClaudeYour favorite work tools are now interactive connectors inside Claude

Your favorite work tools are now interactive connectors inside ClaudeYour favorite work tools are now interactive connectors inside Claude

Dec 8, 2025

Claude Code and Slack

Product announcements

Claude Code and SlackClaude Code and Slack

Claude Code and SlackClaude Code and Slack

Aug 25, 2025

Piloting Claude in Chrome

Product announcements

Piloting Claude in ChromePiloting Claude in Chrome

Piloting Claude in ChromePiloting Claude in Chrome

Transform how your organization operates with Claude

See pricing

See pricingSee pricing

Contact sales

Contact salesContact sales

Get the developer newsletter

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

SubscribeSubscribe

Please provide your email address if youd like to receive our monthly developer newsletter. You can unsubscribe at any time.

Thank you! You’re subscribed.

Sorry, there was a problem with your submission, please try again later.

HomepageHomepage

NextNext

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Write

Button TextButton Text

Learn

Button TextButton Text

Code

Button TextButton Text

Write

Help me develop a unique voice for an audience

Hi Claude! Could you help me develop a unique voice for an audience? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Improve my writing style

Hi Claude! Could you improve my writing style? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Brainstorm creative ideas

Hi Claude! Could you brainstorm creative ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Learn

Explain a complex topic simply

Hi Claude! Could you explain a complex topic simply? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Help me make sense of these ideas

Hi Claude! Could you help me make sense of these ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Prepare for an exam or interview

Hi Claude! Could you prepare for an exam or interview? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Code

Explain a programming concept

Hi Claude! Could you explain a programming concept? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Look over my code and give me tips

Hi Claude! Could you look over my code and give me tips? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Vibe code with me

Hi Claude! Could you vibe code with me? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

More

Write case studies

This is another test

Write grant proposals

Hi Claude! Could you write grant proposals? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to — like Google Drive, web search, etc. — if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational.

Please execute the task as soon as you can - an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Write video scripts

this is a test

AnthropicAnthropic

© [year] Anthropic PBC

Products

Claude

ClaudeClaude

Claude Code

Claude CodeClaude Code

Cowork

CoworkCowork

Max plan

Max planMax plan

Team plan

Team planTeam plan

Enterprise plan

Enterprise planEnterprise plan

Download app

Download appDownload app

Pricing

PricingPricing

Log in

Log inLog in

Features

Claude in Chrome

Claude in ChromeClaude in Chrome

Claude in Slack

Claude in SlackClaude in Slack

Claude in Excel

Claude in ExcelClaude in Excel

Claude in PowerPoint

Claude in PowerPointClaude in PowerPoint

Skills

SkillsSkills

Models

Opus

OpusOpus

Sonnet

SonnetSonnet

Haiku

HaikuHaiku

Solutions

AI agents

AI agentsAI agents

Code modernization

Code modernizationCode modernization

Coding

CodingCoding

Customer support

Customer supportCustomer support

Education

EducationEducation

Financial services

Financial servicesFinancial services

Government

GovernmentGovernment

Healthcare

HealthcareHealthcare

Life sciences

Life sciencesLife sciences

Nonprofits

NonprofitsNonprofits

Claude Developer Platform

Overview

OverviewOverview

Developer docs

Developer docsDeveloper docs

Pricing

PricingPricing

Regional compliance

Regional complianceRegional compliance

Amazon Bedrock

Amazon BedrockAmazon Bedrock

Google Cloud’s Vertex AI

Google Cloud’s Vertex AIGoogle Cloud’s Vertex AI

Console login

Console loginConsole login

Learn

Blog

BlogBlog

Claude partner network

Claude partner networkClaude partner network

Courses

CoursesCourses

Connectors

ConnectorsConnectors

Customer stories

Customer storiesCustomer stories

Engineering at Anthropic

Engineering at AnthropicEngineering at Anthropic

Events

EventsEvents

Plugins

PluginsPlugins

Powered by Claude

Powered by ClaudePowered by Claude

Service partners

Service partnersService partners

Startups program

Startups programStartups program

Tutorials

TutorialsTutorials

Use cases

Use casesUse cases

Company

Anthropic

AnthropicAnthropic

Careers

CareersCareers

Economic Futures

Economic FuturesEconomic Futures

Research

ResearchResearch

News

NewsNews

Responsible Scaling Policy

Responsible Scaling PolicyResponsible Scaling Policy

Security and compliance

Security and complianceSecurity and compliance

Transparency

TransparencyTransparency

Help and security

Availability

AvailabilityAvailability

Status

StatusStatus

Support center

Support centerSupport center

Terms and policies

Privacy choices

Cookie settings

We use cookies to deliver and improve our services, analyze site usage, and if you agree, to customize or personalize your experience and market our services to you. You can read our Cookie Policy here.

Customize cookie settings Reject all cookies Accept all cookies

Necessary

Enables security and basic functionality.

Required

Analytics

Enables tracking of site performance.

Off

Marketing

Enables ads personalization and tracking.

Off

Save preferences

Privacy policy

Privacy policyPrivacy policy

Responsible disclosure policy

Responsible disclosure policyResponsible disclosure policy

Terms of service: Commercial

Terms of service: CommercialTerms of service: Commercial

Terms of service: Consumer

Terms of service: ConsumerTerms of service: Consumer

Usage policy

Usage policyUsage policy

x.comx.com

LinkedInLinkedIn

YouTubeYouTube

InstagramInstagram

English (US)

English (US)

日本語 (Japan)

Deutsch (Germany)

Français (France)

한국어 (South Korea)

Claude Developer Platform

Agents

📋 讨论归档

讨论进行中…