As industry competition shifts toward underlying infrastructure, search is evolving from an application feature into the "water, electricity, and gas" of the AI era, driving large-scale implementation.

Article author and source: GeekPark

Over the past year, the battlefield in the AI industry has been quietly shifting.

As agents rapidly penetrate real-world business scenarios such as office work, content creation, and finance, industry consensus is quietly shifting: the upper limit of AI experience is determined by its ability to connect with the real world.

Policy updates, market fluctuations, industry trends, niche data... Information from the real world is constantly updating every second. Even large models have inherent limits to their knowledge base; to generate reliable conclusions and accomplish complex tasks, they must integrate real-time, authoritative external information.

Search is the core interface through which large models connect with the real world. Yet it remains a common bottleneck in current AI deployments: traditional search relies on keyword matching and cannot understand the nuanced intentions behind natural language, while low-quality content often triggers model hallucinations. Building a complete retrieval system from scratch, however, entails immense engineering challenges—including web crawling, parsing, indexing, and maintenance—that most teams cannot afford.

In fact, Tencent has been quietly building this foundational capability for some time. Over the past year, Tencent has been highly active in AI applications, launching more than 50 products such as WorkBuddy, the AI knowledge assistant ima, and the Tencent Video AI assistant, covering core scenarios including office work, content, and finance. However, few have noticed that the underlying search infrastructure enabling these products’ web retrieval and information access capabilities has long remained behind the scenes.

On October 20, this "hidden card" was officially unveiled, marking a comprehensive upgrade to Yuanbao Search, which is now available to individual developers and enterprises via API through Tencent Cloud.

This search capability, which has already supported over 50 core products within Tencent, accumulated trillions of indexed entries, integrated 157 authoritative sources, and handles hundreds of millions of daily queries, is now being officially upgraded as Yuanbao Search—a public AI infrastructure open to the entire industry.

Coming from behind the scenes to the forefront, the opening of Yuanbao Search fills a long-missing, real-world business-validated piece of information infrastructure needed for the large-scale deployment of AI.

I. An Unrevealed Search Card

The evolution of search has always been closely tied to how the internet organizes information and allocates traffic. Each generational shift in the industry has redefined the role of search.

In the early days of PC internet, it was an era of directory navigation, where Yahoo organized websites through manual categorization, making search the internet’s “map.” Later, Google ushered in the era of keyword search with its PageRank algorithm, turning search into the primary gateway for internet traffic. The core logic was “people finding information,” returning lists of web pages through keyword matching. Advertising became the dominant business model—those who controlled search controlled the distribution of traffic.

With the advent of the mobile internet era, super apps have fragmented the open internet, giving rise to information silos. The entry-point value of general search has been diluted, and search capabilities have gradually become embedded features within vertical products—searching for products on e-commerce platforms, videos on content platforms, and chat histories on social tools. Search has evolved from an independent traffic product into a foundational capability serving each ecosystem, with new forms such as voice search and local search emerging as well.

As the AI wave sweeps in, search is undergoing its third fundamental transformation: its target audience has shifted from "humans" to "agents," and its core objective has changed from "returning web links" to "delivering trustworthy, structured, and directly actionable information."

“The users facing search have shifted from humans in the 1.0 era to agents in the 3.0 era. This is a fundamental change in the nature of the users,” said Ma Jin. Large models inherently suffer from three key shortcomings: hallucinations, insufficient knowledge timeliness, and missing domain-specific terminology—search is the core solution to addressing these issues.

Among these, real-time capability is one of the most core values distinguishing AI search from traditional static large models. Traditional large models have knowledge frozen at their training cutoff date and are inherently delayed when facing the rapidly changing real world; in contrast, AI-native search synchronizes with the latest updates across the web, enabling agents to continuously perceive external changes—whether it’s real-time market fluctuations, recent regulatory amendments, or immediate industry developments—all instantly transformed into actionable information inputs for the agent.

For agents, search has become the core sense for perceiving the external world; it is this fundamental shift in requirements that has brought search back to its central position as infrastructure rather than merely an application-level feature.

But from today’s industry perspective, the supply of search infrastructure adapted for the AI era far lags behind demand.

Traditional search APIs are rooted in the logic of "people finding information" and still rely primarily on keyword matching. When faced with complex queries involving long-form semantics and strong intent, their recall accuracy often drops significantly. General web search results are cluttered with excessive SEO content, promotional articles, and outdated information; feeding these directly to large models easily induces hallucinations, leading AI to generate incorrect conclusions.

The industry has yet to find a permanent solution to the flood of low-quality AIGC content. Ma Jin noted that this is essentially a cat-and-mouse game: it requires models to identify characteristics of individual pieces of content, as well as targeted governance of websites and accounts that frequently post low-quality material. However, a more effective approach is to increase the weight of high-quality source content at the source, ensuring that trustworthy content is prioritized in retrieval.

Choosing to build a search system from scratch presents an astonishingly high barrier: combating web crawlers, cleaning and parsing web content, storing and indexing hundreds of billions of data points, and iteratively optimizing ranking models—all requiring a full team of specialized search engineers, along with sustained computational power and operational investment. For the vast majority of startups and small-to-medium companies, this is nearly an impossible task. On one side is the rigid demand from Agent implementations for trustworthy search capabilities; on the other is a supply side marked by high barriers and poor adaptability. The gap in AI search infrastructure has become a universal bottleneck across the industry.

Tencent has actually been quietly addressing this gap.

In the public’s perception, Tencent’s AI strategy is often divided into two layers: the bottom layer consists of the HunYuan large model, while the top layer includes user-facing applications such as WorkBuddy and Tencent Video AI. However, few notice the crucial intermediate infrastructure layer between the model and applications—common capabilities like Yuanbao Search and vector databases—that connect to the real world, process external information, and provide stable data inputs for all upper-layer AI applications.

It is designed for AI developers and enterprises, functioning more like an “information power station” behind AI applications—delivering structured, high-reliability information to large models and agents to support the stable operation of upper-layer services. Ma Jin also clearly defined the underlying positioning of Yuanbao Search in an interview: “We will not create a standalone consumer-facing search interface; instead, we will operate behind various AI products, providing higher-quality and more trustworthy search sources.”

For a long time, this hidden card remained behind the scenes, serving only Tencent’s internal business lines. But as big tech companies’ AI competition shifts from “competing on model parameters” to “competing on deployment capabilities,” the intermediate infrastructure has become the new decisive factor—and this long-hidden search card is finally ready to be revealed.

II. First, enable the full internal workflow: a search foundation proven at scale

Previously, AI search products in the industry were mostly at the "plugin level"—adding internet connectivity to large models, still fundamentally relying on "users posing requests, with search filling in information."

But to truly enter the Agent era, search must become a standalone, production-grade infrastructure: capable of continuously ingesting authoritative information, adapting to AI’s semantic understanding, and supporting large-scale, highly stable concurrent requests.

Yuanbao Search initially began by meeting the online search needs of the Yuanbao app. As AI products within Tencent were rolled out en masse, it gradually evolved into a universal AI-friendly search capability, integrating with over 50 core internal products across areas such as office productivity, content, finance, and tools. Its daily search volume reaches hundreds of millions, with an index scale in the hundreds of billions, covering 13 major vertical industries and 157 authoritative sources—all refined through continuous real-world business applications.

This "internal validation first, then external output" model gives Yuanbao Search a natural advantage in engineering maturity.

In the office agent scenario, after WorkBuddy integrated Yuanbao Search, it completely shed the heavy burden of maintaining its own search system. Previously, to enable the agent with web research capabilities, the team had to build an entire engineering system including web crawlers, indexing, and ranking—resulting in high development costs and difficulty ensuring the reliability of returned information. After integrating Yuanbao Search, authoritative information is directly incorporated into the agent’s workflow, with every conclusion traceable to official sources. The team no longer needs to expend effort maintaining data sources or crawling pipelines, allowing them to focus entirely on task planning and delivery experience for the agent.

In the content and entertainment scenario, after integrating Tencent Video’s AI assistant, information coverage increased by 20% directly. Previously, due to limitations of the internal content library, updates on new releases, popularity rankings, and cast and crew activities outside the platform were delayed, and retrieval performance was suboptimal when users made vague, plot-based requests. After integrating Yuanbao Search, real-time information from across the web is synchronized instantly, enabling precise matching of works even based on vague plot descriptions, directly driving growth in both average AI search usage per user and secondary retention.

Search and Agent do not have a one-way calling relationship, but rather a bidirectional, co-evolutionary collaboration. Ma Jin explains: “Agents must learn to break down queries that are better suited for retrieval, while search must provide agents with characteristic signals such as authority level, publication time, and ranking satisfaction. In turn, the final cited results and quality feedback from agents are fed back to search to drive continuous optimization.” This bidirectional alignment directly improves efficiency: after the upgrade, the average number of search calls required by internal agents to complete similar tasks dropped from 8 to around 5—higher result quality means agents no longer need to repeatedly retrieve and test, significantly improving task completion efficiency.

The full-scenario deployment across dozens of business lines has effectively subjected Yuanbao Search to continuous, multi-dimensional stress testing and scenario adaptation, allowing this capability to bypass the demo stage and directly achieve production-grade stability and compatibility.

In practical use, the official platform provides three modes: direct API, MCP, and SkillHub.

To get started, developers only need to enable Yuanbao Search on Tencent Cloud to immediately receive a free monthly quota of 2,000 requests.

For developers accustomed to direct API connections, simply create and copy a plaintext key on the 'Service API Key' page in the Tencent Cloud console, then configure it in your local environment variable (export TENCENTCLOUD_WSA_APIKEY='sk-your key') to quickly complete authentication and API calls.

In standard mode, based on our search query, 10 results will be returned, each including a title, date, site, URL, and score.

In addition, the official platform offers a rich set of configuration options, including Freshness (time range), Industry (vertical domain), Mode (result type), Site (include/exclude sites), and Cnt (number of results), effectively addressing the limitation of traditional search adjustment methods. For example, setting Freshness to d1 locks results to news from the past day, while setting Industry to gov restricts searches exclusively to government and party sources, completely preventing low-quality marketing accounts from misleading large models.

Using the MCP approach is even simpler: just map the previously obtained key into the configuration file of Cursor or Codex, enter the official remote MCP address https://api.wsa.cloud.tencent.com/Mcp, and after launching, confirm that SearchPro is lit up in the tool list. You can then directly use natural language in the editor to command the Agent to access the internet.

If you use SkillHub to connect, it becomes even simpler: after installing the skill package via the command line, a SKILL.md guide and script are automatically generated locally.

At this point, simply tell the editor’s AI, “Help me search for the latest...,” and the Agent will automatically recognize the intent and perform the call.

Additionally, I ran another comparative test with the search keyword "latest policy release." Since this type of content typically requires monitoring the most recent official sources, I added the filters Industry=gov and Freshness=d7 to retrieve the latest policy updates.

The results returned consist solely of official government sources such as the official website of the Chinese government, the Ministry of Commerce, and Xinhua News Agency, with no involvement of self-media or marketing accounts.

Overall, the core strength of Yuanbao Search lies in its engineering maturity: it is lightweight to integrate, delivers accurate understanding, provides reliable data sources, and responds quickly—addressing developers’ core need for reliable search capabilities without having to build them from scratch. It is a true production-ready search infrastructure.

Ma Jin said that Yuanbao Search will continue to evolve in three directions: first, achieving self-iteration and self-optimization through a data flywheel, enabling the search model to continuously undergo closed-loop evolution; second, continuously expanding high-quality multilingual content across the web—currently covering both Chinese and English, with further expansions planned; and third, deeply specializing in vertical domains, thoroughly refining search capabilities in finance, healthcare, law, and other niche areas.

III. The competition in infrastructure has just begun in the Agent era

Throughout the development cycle of the technology industry, the focus of competition will always trickle down the technology stack.

In the era of cloud computing, corporate competition centers on computing and storage resources—those who can provide stable, low-cost computing power gain control of the industry. In the early stages of the large model era, industry attention focused on model parameters and generation quality, with parameter scale being the most critical competitive metric.

As AI enters the agent deployment phase, the focus of competition shifts further downward—those who can connect AI to the real world more stably and at lower cost will be best positioned to enable scalable applications above.

A clear industry trend is emerging: AI search is evolving from an optional plugin into the "water, electricity, and gas" of the agent era, becoming a foundational infrastructure indispensable to all intelligent applications.

Shortly before Yuanbao Search’s official announcement, DeepSeek also officially launched its proprietary Search API, providing real-time web retrieval capabilities for agents and RAG applications. From model developers to cloud providers, domestic players are aggressively investing in search infrastructure—confirming a key insight: opening up search capabilities is an inevitable step toward industry maturity. As large model capabilities converge, competitive differentiation in upper-layer applications will ultimately hinge on the depth and strength of underlying infrastructure.

The path choices of different players also reflect distinct strategies.

Model providers' search services emphasize deep integration with their own models, focusing on lightweight design and low cost; whereas Yuanbao Search, originating from cloud providers, excels in proven engineering stability through large-scale business operations, access to rich authoritative sources in specialized domains, and strong ecosystem resource integration capabilities. Different players can collectively expand the market for search infrastructure, offering developers a more diverse range of choices.

More importantly, it lowers the barrier to AI adoption across the entire industry. By opening up search infrastructure like Yuanbao Search—proven at scale—developers no longer need to reinvent the wheel. Instead, they can focus their limited resources on product logic, scenario innovation, and user experience, which will further spur the creation of AI applications in more vertical use cases and accelerate industry-wide adoption.

The second half of AI will inevitably be a war over infrastructure. Search, as the core gateway connecting large models to the real world, will become a critical battleground in this war.