中文EN RSS

DeepSeek以低价模型闻名 现又进军搜索API市场挑战对手DeepSeek, known for its low-priced models, now expands into the search API market to challenge rivals


AI free token/AI免费积分

DeepSeek不满足于用低成本模型挑战各大前沿AI实验室,这家中国AI公司如今又将目标瞄准了搜索API市面。

周六,该公司宣布借助两个新API——Search和Fetch——向外部开发者开放其内部研发的搜索引擎,使开发者可以检索实时网络搜索结果和网页内容,并将其用于自己的应用和智能体中。

这两项服务每1000次成功请求收费0.75美元,开发者可以独立于DeepSeek的AI模型使用这些服务。

此次发布还伴随着DeepSeek-V4.1-Flash-1017的推出,这是该公司9月发布的V4.1 Flash模型的小版本更新。

智能体为何需要网络搜索

大语言模型是基于截至特定时间点收集的数据开展训练的,这意味着它们并不总能回答关于近期事件、更新文档或其他动态信息的难题。借助兼容实时网络搜索,模型可以从互联网检索当前信息,从而补充其回答内容。

对于构建AI智能体的开发者来说,这一点尤为关键。例如,一个编程智能体可能需要查阅最新的软件文档、调查最近披露的漏洞,或检查某个库是否已更新。

此前,使用DeepSeek模型的开发者有几种方式来达成这一能力。他们可以将应用连接到第三方搜索服务,或借助DeepSeek的模型API使用其现有的搜索能力。

实际上,当DeepSeek在8月开源其智能体框架(Harness)时,网络搜索就是其能力之一。该框架兼容可互换的搜索给出商,涵盖Exa、Perplexity以及DeepSeek自身。不过,依据该框架的文档说明,借助框架访问DeepSeek的原生搜索需要额外的模型请求,每次搜索都作为对DeepSeek模型的单独调用,这会带来额外的延迟和Token成本。

新的API旨在简化这一流程。开发者现在可以直接调用DeepSeek的Search端点来获取排序后的网络搜索结果,涵盖标题、URL、摘要和概述,而Fetch端点则以Markdown格式检索单个网页的内容。

这两个端点都不需要单独的模型请求,这意味着开发者也可以将DeepSeek的搜索能力集成到由竞争对手AI模型驱动的应用中。

DeepSeek研发自有引擎

在过去两年的大部分时间里,DeepSeek一直依赖外部给出商来兼容网络搜索能力,同时也在构建自己的搜索引擎。

据中国商业媒体《每日经济新闻》2025年3月报道,DeepSeek已集成了总部位于杭州的中国AI搜索初创公司博查(Bocha)的搜索API。博查首席工艺官翁柔颖指出,早在该聊天机器人今年年初人气暴涨之前,DeepSeek就已在使用其服务。

与此同时,DeepSeek 2025年2月的隐私政策也承认其依赖第三方搜索服务,并以微软的Bing为例,说明其外部API给出商之一。

在幕后,该公司也一直在为自建搜索基础设施打基础。今年1月,彭博社报道称,依据DeepSeek发布的招聘广告,该公司正在招募工程师,以构建一款多语言AI搜索引擎。据报道,其规划的能力不仅限于文本搜索,还将涵盖图像和音频。

快进约九个月后,该公司表示其内部研发的搜索引擎(名为DeepSeek Search)如今每天处理数亿次请求,并维护着一个包含数千亿网页的索引。

依据其10月17日发布的更新日志,该引擎兼容多语言搜索,并使用过滤机制来识别搜索垃圾信息、低质量AI生成内容,以及操纵搜索排名的行为。

DeepSeek对其底层架构,以及如何构建和维护索引披露的信息相对较少。当下也不清楚该公司是否已全面替换其所有外部搜索供应商。

AI搜索市面不断增长

搭建一个搜索引擎绝非易事。它需要爬取网页、维护网页内容索引并对结果开展排序的基础设施,同时还要随着网络变化不断更新信息。多年来,这意味着希望为应用获取原始搜索结果的开发者,大多会求助于谷歌和微软这两家具备自有网络规模索引的公司。

不过现在,越来越多的公司开始投资自建搜索基础设施,并日益将AI应用和智能体作为重点考量对象。例如,Brave运营着一个包含超过400亿网页的独立索引,该公司表示该索引借助自有爬虫及其他发现机制维护。与此同时,AI搜索公司Exa也开发了自己的爬取和索引基础设施。

Perplexity同样投资建设了专有的搜索基础设施,涵盖自建的爬取和索引平台。另一竞争者Parallel表示,其已从零开始搭建搜索基础设施,涵盖爬取、索引、查询处理和排序等组件。

Cloudflare也以另一种方式进入这一市面。本月早些时候,该公司借助其AI Gateway推出了Web Search API,使开发者可以借助单一接口访问多家给出商的搜索结果。该服务当下处于测试阶段,默认使用Ceramic作为给出商,并按各给出商公布的API费率向客户收费,不加收额外费用。

而DeepSeek维度,正在售价上与现有巨头展开激烈竞争,尽管严格来说它并非市面上最便宜的选项。为AI智能体给出基于关键词的网络搜索服务的Ceramic.ai,其按需付费计划每1000次查询收费0.25美元。

尽管如此,DeepSeek每1000次成功搜索收费0.75美元,相较于许多其他服务仍具竞争力,涵盖Parallel的Turbo和Fast搜索服务,以及Perplexity的Fast Search API,这两者均为每1000次搜索收费1美元。与此同时,Exa的Instant Search请求每1000次收费4美元,Brave则为每1000次搜索收费5美元。另一个值得关注的参与者是Tavily,这是一家为AI智能体给出搜索API的公司,于今年2月被欧洲AI基础设施公司Nebius收购。Tavily的按需付费计划中,基础搜索每1000次收费8美元。

值得注意的是,各服务在检索能力、结果数量限制和表现维度存在差异,因而售价并不完全具有可比性。但这些数据为评估DeepSeek进军搜索基础设施领域给出了有用的参考基准。

这些替代思路的浮现,恰逢谷歌和微软相继撤下了长期给出的搜索API服务。微软于2025年8月停用了其Bing Search API,而谷歌则关闭了其Custom Search JSON API对新客户的开放,现有使用者须在2027年1月1日之前完成迁移。两家公司仍借助各自的AI平台给出搜索能力,涵盖为Gemini给出的Grounding with Google Search,以及为Microsoft Foundry智能体给出的Grounding with Bing Search。这些服务允许在各自平台上运行的AI模型检索网络信息,并生成带引用来源的答案。

像DeepSeek这样的独立API,则让开发者可以直接获取搜索结果,并自由挑选搭配使用的AI模型和应用。

开发者反应及细则

此次公告在开发者中引发了关于定价,以及新API相较于现有搜索集成思路能否带来实质提高的广泛讨论。在Reddit上,一位使用者特别对售价表示欢迎,称:"我认为没有其他搜索给出商的售价能接近DeepSeek给出的水平。"

另一些人则质疑,对于已经借助第三方工具访问网络搜索的使用者来说,这项发布最终能带来什么改变。一位Reddit使用者指出,他们已经在使用兼容网络搜索的Hermes AI智能体配合DeepSeek。另一位使用者回应称,对于那些已经规格了自己的搜索工具的使用者来说,新API可能难以带来立竿见影的好处。

Hacker News上也浮现了类似的讨论,一位开发者质疑,既然存在像SearXNG这样免费的自托管替代思路,为何还需要付费搜索API。另一位评论者则指出,企业有额外的需要,涵盖"可靠的延迟表现、数据处理保障",以及借助搜索查询暴露敏感公司信息的潜在风险。

对于正在权衡是否使用DeepSeek API的企业来说,最后一点尤其值得关注。DeepSeek的隐私政策指出,其在中国收集、处理和存储个人数据,并且通常只要账户保持开通状态,就会一直保留使用者输入内容。不过,Search和Fetch的文档并未说明搜索查询具体的保留期限。

这些讨论凸显了个人开发者与企业在成本、可靠性以及处理大量请求能力维度存在的不同需要。尽管DeepSeek的定价确实具有竞争力,但仍有一些实际限制值得关注。

依据其API文档,每个账户的Search和Fetch请求上限均为每秒5次,不过单次搜索最多可返回50条结果而不会增加请求费用。另外,DeepSeek仅对返回结果的请求收费,因而失败的调用、空搜索和超时请求均不收费。

当下,DeepSeek暂时对借助DeepSeek Harness、Claude Code和Codex(在使用其API时)发出的服务端工具调用免收Search和Fetch费用。但这仅是临时性安排,将于10月31日结束,此后将恢复标准的按请求收费模式。依据DeepSeek的公告,直接调用独立API已开始计费。

The New Stack
大语言模型搜索APIDeepSeek

本文来源:https://www.zhiding.cn/ai-infrastructure/2026/1026/3203481.shtml
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

DeepSeek, not content with challenging major frontier AI labs with its low-cost models, is now targeting the search API market.

On Saturday, the company announced that it is opening up its internally developed search engine to external developers via two new APIs—Search and Fetch—allowing developers to retrieve real-time web search results and web page content for use in their own applications and agents.

These two services are priced at $0.75 per 1,000 successful requests, and developers can use them independently of DeepSeek's AI models.

The launch also comes alongside the rollout of DeepSeek-V4.1-Flash-1017, a minor version update to the V4.1 Flash model released by the company in September.

Why Agents Need Web Search

Large language models are trained on data collected up to a specific point in time, which means they cannot always answer questions about recent events, updated documentation, or other dynamic information. By supporting real-time web search, models can retrieve current information from the internet to supplement their responses.

This is particularly crucial for developers building AI agents. For example, a coding agent might need to consult the latest software documentation, investigate recently disclosed vulnerabilities, or check whether a library has been updated.

Previously, developers using DeepSeek models had several ways to achieve this capability. They could connect their applications to third-party search services, or use DeepSeek's existing search capability via its model API.

In fact, when DeepSeek open-sourced its agent framework (Harness) in August, web search was one of its capabilities. The framework supports interchangeable search providers, including Exa, Perplexity, and DeepSeek's own. However, according to the framework's documentation, accessing DeepSeek's native search via the framework requires additional model requests, with each search acting as a separate call to the DeepSeek model, which introduces extra latency and token costs.

The new APIs aim to simplify this process. Developers can now directly call DeepSeek's Search endpoint to obtain ranked web search results—including titles, URLs, snippets, and overviews—while the Fetch endpoint retrieves the content of individual web pages in Markdown format.

Neither endpoint requires a separate model request, meaning developers can also integrate DeepSeek's search capabilities into applications powered by competing AI models.

DeepSeek Develops Its Own Engine

For much of the past two years, DeepSeek has relied on external providers to support web search capabilities, while simultaneously building its own search engine.

As reported by the Chinese business media outlet National Business Daily in March 2025, DeepSeek had integrated the search API of Bocha, a Chinese AI search startup based in Hangzhou. Bocha's Chief Technology Officer Weng Rouying noted that DeepSeek had been using its services long before the chatbot's popularity surged early this year.

Meanwhile, DeepSeek's privacy policy from February 2025 also acknowledged its reliance on third-party search services, citing Microsoft's Bing as an example of one of its external API providers.

Behind the scenes, the company has also been laying the groundwork for building its own search infrastructure. In January this year, Bloomberg reported that, based on job postings published by DeepSeek, the company was hiring engineers to build a multilingual AI search engine. Reportedly, its planned capabilities would extend beyond text search to include images and audio.

Fast forward about nine months, and the company states that its internally developed search engine (named DeepSeek Search) now handles hundreds of millions of requests daily and maintains an index containing hundreds of billions of web pages.

According to its changelog released on October 17, the engine supports multilingual search and uses filtering mechanisms to identify search spam, low-quality AI-generated content, and attempts to manipulate search rankings.

DeepSeek has disclosed relatively little about its underlying architecture and how it builds and maintains its index. It is also unclear whether the company has fully replaced all of its external search providers.

The AI Search Market Is Growing

Building a search engine is no easy task. It requires infrastructure for crawling web pages, maintaining an index of web content, and ranking results, as well as constantly updating information as the web changes. For years, this meant that developers looking to obtain raw search results for their applications mostly turned to Google and Microsoft, the two companies with their own web-scale indexes.

However, a growing number of companies are now investing in their own search infrastructure, increasingly with AI applications and agents in mind. For example, Brave operates a separate index of over 40 billion web pages, which the company says is maintained using its own crawlers and other discovery mechanisms. Meanwhile, AI search company Exa has also developed its own crawling and indexing infrastructure.

Perplexity has likewise invested in building proprietary search infrastructure, including a self-built crawling and indexing platform. Another competitor, Parallel, states that it has built its search infrastructure from scratch, including components for crawling, indexing, query processing, and ranking.

Cloudflare has also entered this market in a different way. Earlier this month, the company launched a Web Search API via its AI Gateway, enabling developers to access search results from multiple providers through a single interface. The service is currently in beta, uses Ceramic as the default provider, and charges customers at the published API rates of each provider without additional markup.

As for DeepSeek, it is competing fiercely on price with existing giants, although strictly speaking it is not the cheapest option on the market. Ceramic.ai, which provides keyword-based web search services for AI agents, charges $0.25 per 1,000 queries on its pay-as-you-go plan.

Despite this, DeepSeek's price of $0.75 per 1,000 successful searches remains competitive compared to many other services, including Parallel's Turbo and Fast search services and Perplexity's Fast Search API, both of which cost $1 per 1,000 searches. Meanwhile, Exa's Instant Search requests cost $4 per 1,000, while Brave charges $5 per 1,000 searches. Another notable player is Tavily, a company that provides search APIs for AI agents, which was acquired by the European AI infrastructure company Nebius in February this year. Tavily's pay-as-you-go plan charges $8 per 1,000 for basic search.

It is worth noting that differences in retrieval capabilities, result quantity limits, and performance dimensions among these services mean that pricing is not entirely comparable. However, these figures provide a useful benchmark for assessing DeepSeek's foray into the search infrastructure space.

The emergence of these alternative options coincides with Google and Microsoft phasing out their long-standing search API services. Microsoft disabled its Bing Search API in August 2025, while Google has closed its Custom Search JSON API to new customers, requiring existing users to complete migration by January 1, 2027. Both companies continue to provide search capabilities through their respective AI platforms, including Grounding with Google Search for Gemini and Grounding with Bing Search for Microsoft Foundry agents. These services allow AI models running on their respective platforms to retrieve web information and generate answers with cited sources.

Independent APIs like DeepSeek's, on the other hand, allow developers to obtain search results directly and freely choose the AI models and applications to pair them with.

Developer Reactions and Details

This announcement has sparked widespread discussion among developers regarding pricing, and whether the new API can bring substantial improvements compared to existing search integration approaches. On Reddit, a user specifically welcomed the pricing, stating: "I don't think any other search provider's pricing can even come close to what DeepSeek is offering."

Others questioned what ultimate change this release would bring for users already accessing web search via third-party tools. A Reddit user pointed out that they are already using the Hermes AI agent, which is compatible with web search, alongside DeepSeek. Another user responded that for those who have already customized their own search tools, the new API might be difficult to bring immediate benefits.

Similar discussions emerged on Hacker News, where a developer questioned why a paid search API is needed when free self-hosted alternatives like SearXNG exist. Another commenter pointed out that enterprises have additional requirements, including "reliable latency performance, data processing guarantees," and the potential risk of exposing sensitive company information through search queries.

For enterprises weighing whether to use the DeepSeek API, this last point is particularly noteworthy. DeepSeek's privacy policy states that it collects, processes, and stores personal data in China, and generally retains user input as long as the account remains open. However, the documentation for Search and Fetch does not specify the exact retention period for search queries.

These discussions highlight the different needs of individual developers and enterprises in terms of cost, reliability, and the ability to handle large volumes of requests. Although DeepSeek's pricing is indeed competitive, there are still some practical limitations worth noting.

According to its API documentation, the Search and Fetch request limit for each account is 5 requests per second, but a single search can return up to 50 results without incurring additional request costs. Additionally, DeepSeek only charges for requests that return results, so failed calls, empty searches, and timed-out requests are not charged.

Currently, DeepSeek is temporarily waiving Search and Fetch fees for server-side tool calls made via DeepSeek Harness, Claude Code, and Codex (when using its API). However, this is only a temporary arrangement that will end on October 31, after which it will return to the standard pay-per-request model. According to DeepSeek's announcement, direct calls to the standalone API have already started being billed.

The New Stack

Large Language Model Search API DeepSeek

DeepSeek, known for its low-priced models, now expands into the search API market to challenge rivals

33 Engines, Nearly 100,000 GPUs! A Stress Test of Domestic Computing Power for Commercial Rockets

PEC 2026 AI Innovators Conference

"Late" Borne Optical, but "Early" in AI: Borne Optical's AI Push

How to Implement Industrial AI? Weichao Guoke Provides the Answer

33台发动机,近10万张卡!商业火箭的一次国产算力极限验证

PEC 2026 AI创新者大会

“晚”走的伯恩光学,做AI却赶了个“早”

工业+AI如何落地?魏桥国科交出答卷

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry

←上一篇← Prev下一篇→Next →