Tencent has publicly unveiled its internal AI search mechanism for the first time.
An exclusive report has learned that Tencent is about to launch an AI search engine for consumer users, featuring a minimalist interface, full AI integration, enhanced ranking of authoritative sources, and a combined AI Q&A and web search experience.
Just two weeks earlier, Tencent Cloud had already launched "Yuanbao Search" (formerly Internet Search) for business clients. It had quietly operated internally for years and is now being pushed into the spotlight under the banner of AI search.
According to Tencent insiders, this consumer-facing search engine shares the same underlying capabilities as the Yuanbao app's built-in search and Tencent Cloud's "Yuanbao Search" API. In browser and web search scenarios, the three are essentially the same capability productized through different entry points.
Currently, Yuanbao Search offers three access methods: Skill, MCP, and API. With an invitation code, users can make up to 6,000 calls during October and November. The onboarding process is simple: obtain an API key from Tencent Cloud, open a terminal, and issue a curl command to initiate a request.
This raises a question: after more than a decade in the search arena, why has Tencent only now upgraded these capabilities into a cloud service for sale?
To a large extent, Baidu has dominated China's search market for over twenty years. Tencent, after years of meandering, tried SOSO, acquired Sogou, launched WeChat Search, and rolled out the Yuanbao app, but never achieved a substantial breakthrough in the consumer market.
Now, riding the wave of Agent proliferation, Tencent is naturally eager to position itself in business-side search infrastructure.
It is understood that the Yuanbao Search team was assembled in mid-2024 alongside the Yuanbao app, integrating the core technological DNA of QQ Browser and Sogou Search. They restructured Tencent's decade-plus search algorithm and engineering experience into a search infrastructure built for the AI era.
Previously, this capability only served QQ Browser and Sogou: the former was once a super traffic gateway, while the latter held steady in the second tier with considerable scale. Now, with the Agent sector booming, search is shifting from a consumer product entry point to an underlying capability for all AI applications, and demand for standardized search APIs is rapidly expanding.
Against this backdrop, Tencent is no longer content to serve only internal businesses and existing clients. It is capitalizing on the trend to upgrade Yuanbao Search into a cloud service for external monetization.
From this perspective, the opening of Yuanbao Search is less a strategic new frontier and more a pragmatic move to quickly push existing technological assets to market after spotting sector momentum.
Tencent Unveils Internal AI Search Mechanism for the First Time
Ma Jin, Vice President of Tencent Cloud and head of Yuanbao Search R&D, divides search into three eras: 1.0 is where humans enter keywords and filter results; 2.0 is where large models read results and generate answers; 3.0 is where Agents decompose sub-questions and search returns structured results to support reasoning.
Specifically, Search 1.0 was the keyword era, with Baidu and Google at center stage. Tencent spent over a decade trying SOSO and fully acquiring Sogou but never secured a top-tier position in PC and mobile markets.
In the Search 2.0 large model era, Baidu launched Ernie Bot, ByteDance launched Doubao, and OpenAI rolled out ChatGPT with internet access. The deep integration of large models and search became mainstream. Tencent launched the Yuanbao app to compete, binding large model and search capabilities together. But Yuanbao was always positioned as a consumer AI assistant. Even as a chatbot with over 100 million MAU, it still lags behind Doubao and has not established enough say in the search arena.
Stepping boldly into the Search 3.0 Agent era, Tencent no longer insists on confronting consumer entry points head-on. Instead, it is betting on AI search in the business-side infrastructure lane.
Tencent released operational data for Yuanbao Search: a scale of hundreds of billions of indexes, coverage of 13 vertical industries and 157 authoritative sources, and 96% network coverage. It already fully supports over 50 internal core businesses including WorkBuddy, ima, Tencent Financial Technology, and Tencent Video, with hundreds of millions of daily calls.
Ma Jin broke down Yuanbao Search's technical upgrade path for us across four dimensions: cognition, algorithms, architecture, and content.
First, the most intuitive scale data. Official index metrics: 5 billion web pages added daily, 50 billion retrievable documents, and over 150 billion vector indexes. This scale is not the largest in the industry, but Chinese index coverage reaches 95%, and key vertical domains approach 100%. This means Yuanbao Search is not blindly chasing scale but weighting quality content and vertical domain inclusion.
Next, the repeatedly emphasized 157 authoritative sources, with partners covering Ctrip, Meituan, JD.com, Dianping, Autohome, and other leading platforms. This point is somewhat weak, because such business partnerships are not exclusive. Contracts Yuanbao Search can secure today, others can negotiate tomorrow.
What is truly worth highlighting is proprietary content: Yuanbao Search has integrated Tencent Video, Tencent Medipedia, and Tencent Stock, with future possibilities of further connecting the WeChat ecosystem. That is the value others cannot easily replicate.
More critically, when responding to "how to handle inaccessible web pages," Ma Jin admitted: "If a website has anti-crawling measures, we do not include it. This respects their intellectual property." In other words, truly high-value content protected by anti-crawling mechanisms cannot be obtained by Yuanbao Search through general retrieval either. It can only be accessed through individual partnership negotiations.
So the core technical logic of Yuanbao Search remains a data flywheel + Agent collaboration + Trace Debug. Ma Jin describes it as "using AI to find AI, using AI to upgrade search." Essentially, it lets large models assist in labeling, cleaning, and iterating search data. This is an industry-standard direction that nearly all leading search teams are pursuing. It is not a unique technological differentiator for Tencent.
Discussions with two technical experts revealed that Yuanbao Search's algorithm upgrade adopts a three-stage training paradigm: Label Agent annotation, Teacher Model learning, and Student Model distillation. This pipeline is a standard distillation process in large model ranking, already deployed in academia and industry for over two years. The industry consensus is that the real barrier in distillation lies in the base quality of the Teacher Model, the scale and depth of labeled data, and the balance between precision and latency in the Student Model. Yuanbao Search did not provide specific evidence on these details at launch.
It is understood that Tencent Search actually saves over 30% in costs using Gist technology. The principle is caching the gist token KV Cache of documents, so when different queries recall the same document, it can be directly reused. Even with a long-tail query distribution, recalled and ranked documents still show clustering effects. In Yuanbao Search's hybrid ranking scenario, document hit rate is about 85%, yielding considerable savings.
However, this effect depends heavily on cache hit rate. Web pages cannot be fully stored, only popular pages get cached, and long-tail pages still require real-time computation. Officials have not disclosed specific hit rates, and search requests naturally follow a long-tail distribution. This means real average cost savings are likely below the theoretical 30% ceiling.
Additionally, official metrics for recall, coverage, latency, and satisfaction all come from Tencent's internal evaluations, lacking public benchmarks and reproducible test sets.
Take response speed as an example. Officials claim the core link response time after the upgrade is under 300ms (previously around 400–500ms). In our view, in multi-turn Agent invocation scenarios, latency gets multiplied. Compressing single-call time is key to end-user experience.
In actual testing, under real network conditions, five consecutive calls fluctuated between 345–485ms (averaging about 420ms), with a 60% difference between peak and minimum. For infrastructure supporting high-concurrency Agents, latency fluctuation is far more harmful than average latency. As for performance in cross-region and burst traffic extreme production scenarios, there is currently no public data.
Even if these results are usable for a complex retrieval foundation, for enterprise clients pursuing extreme stability, how to smooth out long-tail latency and withstand extreme traffic is the next challenge Yuanbao Search must prove to the market.
In other words, Tencent Cloud's quietly developed Yuanbao Search essentially rebuilt the mature pipeline accumulated in the traditional search era with large model thinking, and validated it in Tencent's real internal businesses.
It is understood that several leading large model companies and external enterprises have already integrated it. But when asked for specific names, the Yuanbao Search team declined to disclose, citing "confidentiality."
Fortunately, when responding to "unavoidable technical bottlenecks," Ma Jin said something honest: "It is actually a process of trade-off... This is more about balancing content, computing power, time consumption, and cost." This statement precisely captures the industry status quo: search development today has long lost any secret recipe. What competes is engineering execution, cost control, and source resources.
Who Is the Target This Time?
Bringing our view back to the Search 3.0 era, the industry is dividing into two clear paths: one is the integrated route binding models and search, represented by OpenAI, Google, Baidu, and ByteDance. They emphasize that search is one of the core capabilities of large models, with deep binding and collaborative optimization. Clients choosing one large model can directly use the accompanying search without additional integration. Integration costs are low, experience is consistent, and it is more suitable for small and medium clients and integrated scenarios.
The other is the independent, neutral search service route, where Tencent Yuanbao Search sits. The core logic: the future Agent market will inevitably be highly fragmented, with diverse models, frameworks, and scenarios. Clients need a neutral, flexible, unbound search foundation. This path is more suitable for medium-to-large clients with technical capability, customization needs, and emphasis on data sovereignty.
From the current industry landscape, the binding route temporarily has the upper hand, because most developers and enterprises prefer one-stop solutions and are unwilling to spend extra on integrating independent search and debugging links. But the Agent ecosystem is still in its early stages, and no one can predict the final market direction.
In light of this, Yuanbao Search's emphasis on neutrality and openness does offer a new choice for clients who do not want to be bound to a single vendor (such as companies with their own large models, developers needing multi-model compatibility, and government and financial institutions with high data sovereignty requirements). Clients can freely pair any vendor's large model and leave only the search component to Yuanbao Search, avoiding technology stack lock-in.
Moreover, Tencent Cloud itself holds a large number of government, financial, internet, and education clients. In the past, when they needed search capabilities while purchasing cloud services, they might have had to separately engage Baidu. Now Tencent is packaging search APIs directly into cloud service bundles, which will undoubtedly directly impact Baidu's enterprise search business.
Especially in government and financial sectors, clients have extremely high requirements for a service provider's comprehensive and compliance capabilities. Tencent Cloud integrating search capabilities into its existing cloud system significantly lowers clients' migration and decision-making costs.
Of course, OpenAI, Google, Baidu, and ByteDance have advantages in the breadth of general web search, and vertical deep partnerships have not formed barriers. Tencent could have used its own ecosystem to quickly carve out differentiation in vertical scenarios. But in its current state, Yuanbao Search cannot yet pose a substantial threat to Baidu.
Because it has gaps in three areas: product capability, service system, and client positioning.
First, it does not support multi-hop search. Currently, it only provides atomic interfaces. Multi-hop capability may be gradually opened based on Agent capabilities accumulated by WorkBuddy. This means clients who buy the search foundation still have to build their own Agent layer for multi-hop reasoning, which undermines the advertised "out-of-the-box" experience.
Second, calling Yuanbao Search currently only returns titles, summaries, links, and vertical cards. To extract full web page content, clients must integrate again themselves. The complete chain should be search result retrieval → web page content extraction → information processing. Now only the first step is open, and clients must fill in the remaining two. Integration costs are quietly raised.
Third, for websites with anti-crawling measures, Yuanbao Search chooses not to include them. This is impeccable on compliance grounds, but it also means a large amount of high-value content is in the index blind spot. Time-sensitive content can only be updated hourly. True real-time access depends on the not-yet-opened Fetch interface. For scenarios with high real-time requirements, current capabilities are far from sufficient.
A Tencent representative also admitted that the largest client is still Tencent's internal products. This fits Tencent's consistent "eat your own dog food" style: prove it internally first, then push it externally.
The problem is that internal businesses can be custom-adapted. The team belongs to the same company, and problems can be solved by pulling everyone into a group chat. Resources can be tilted as needed. But external clients receive a standardized API. If problems arise, can they only submit tickets and wait in queue?
Moreover, external services must be compatible with different models, different Agent frameworks, and different business scenarios, and must bear SLA commitments. The gap here is far more than opening an API can bridge.
In response, Tencent Cloud told us that demand is currently surging. Compared to mature standardized cloud APIs, large model search APIs still have many differentiated needs to absorb. While the team continues to iterate on products, it will handle differentiated needs of some clients on a 1V1 basis.
More severe tests lie in commercialization: transforming from an internal tool to an external product requires not just technology but a complete service system. How to sign SLAs? How to define fault compensation? How to handle data compliance and isolation? How to implement private deployment for government and enterprises? Is billing granularity flexible enough?
We asked the Yuanbao Search commercialization team about this. They said relatively mature plans are already in place, and many automakers and well-known domestic large model companies have already engaged in deep cooperation.
Others may be willing to show support due to long-term business partnerships. But the quality of a B2B infrastructure ultimately depends on not crashing during peak times, responding to failures, and scaling to meet demand. These require real external clients and time to refine.
Tencent officially lists four target client categories: large model companies, smart car companies, office scenarios, and the education sector.
Take smart car companies as an example. Search is highly focused on vertical scenarios like navigation, weather, and POIs. Yuanbao Search is integrated into Tencent Smart Mobility's solutions for automakers, embedded in AI cockpits. This is currently the most in-demand service for automakers, because in-car scenarios have extremely high requirements for search result credibility.
As for leading clients in office and education sectors, they are either already bound to Baidu and other vendors or prefer to build their own. Small and medium clients are highly price-sensitive.
In summary, Yuanbao Search has a solid technical foundation and a clear differentiation direction, opening the curtain on business-side search infrastructure competition. But its full opening this time plunges it into a fiercely competitive red ocean. The AI retrieval API market is already engaged in a brutal price war. Yuanbao Search entering close-quarters combat is undoubtedly a tough battle.
Comments