Tencent Cloud Opens Upgraded Yuanbao Search to Developers, Targeting Agent Era

On October 20, 2026, Tencent announced a major upgrade to Yuanbao Search and made it available to enterprises and individual developers through Tencent Cloud via API, MCP, or Skill integration.

The upgrade covers four areas: cognition, algorithms, architecture, and content. The search model now works with agent models, with agents handling task understanding and decomposition while the search system provides signals such as page authority, publish time, and ranking satisfaction to help agents evaluate results. Feedback from agent citations is also fed back into search optimization. Tencent rebuilt the ranking pipeline using large language models, distilling capabilities into lighter models, and introduced Gist Token compression to reduce web page representation to around 100 tokens, combined with KV Cache reuse to lower inference costs. At the architecture level, the system fuses inverted and vector indexes to support both keyword and semantic retrieval at billion-scale.

According to official figures, the vector index scale exceeds 150 billion entries, average retrieval latency is about 8 milliseconds, unit inference cost is reduced by 90%, and end-to-end speed on the Yuanbao app improved by more than 1 second. The search service covers 157 authoritative sources across 13 industries, including Tencent Maps, Tencent Yidian, and data from partners such as JD.com, Ctrip, Meituan, and VariFlight. Tencent also applies an authoritative-source-first strategy with low-quality filtering to reduce hallucinations.

Ma Jin, Vice President of Tencent Cloud and head of Yuanbao Search R&D, said search is shifting from serving humans in the 1.0 era to serving intelligent agents in the 3.0 era, and Tencent aims to make Yuanbao Search a basic infrastructure for the AI industry. More than 50 Tencent products, including Yuanbao, WorkBuddy, ima, and Tencent Video, are already using Yuanbao Search.