On October 20, Tencent announced that Yuanbao Search is officially open to the public. This web-connected search capability, widely used by Tencent's internal Agents, is offered to developers in three ways: Skill, MCP, and API.
Over the past year, Yuanbao has integrated publicly available web resources with Tencent's premium content, applying large model technology across the entire pipeline of indexing, understanding, and retrieval to build a search engine designed for large models and Agents. Its index now reaches a scale of tens of billions, with hundreds of millions of calls per day, supporting more than 50 core businesses including WorkBuddy, ima, and Tencent Video.
At the source level, Yuanbao Search connects to 157 authoritative sources across 13 vertical industries including finance, healthcare, law, and travel, with financial report coverage exceeding 99% for A-shares, Hong Kong stocks, US stocks, Japanese stocks, and Korean stocks, and it supports minute-level updates for policies, news, weather, and market data. In the retrieval stage, it resolves the true intent behind ambiguities such as colloquial language, mixed multi-language usage, typos, and unclear references, while ranking comprehensively weighs relevance, timeliness, and authority, covering core information with as few documents as possible to reduce duplication and conflicts.
On performance: given the pipeline characteristics of Agents—where a single question is often split into 3 to 5 sub-questions and may invoke more than ten external services—Yuanbao Search adopts at the retrieval layer a fused index architecture combining first inverted index, vector, and fusion efficiency-boosting index types, maintaining a 94% high recall rate while compressing average underlying index retrieval time to 8 milliseconds; the ranking layer compresses documents into about 100 tokens and caches them, reducing core ranking model inference time from 149 milliseconds to 30 milliseconds; the pipeline layer sets time limits for each stage with automatic degradation on timeout, shortening end-to-end response time in the Yuanbao App by nearly 1 second. Integration comes with a monthly free quota of 2,000 calls, embedding retrieval capability directly into Agent workflows, continuing the trend of large model vendors moving from model supply toward engineered toolchains.