Doubao Search Steps Out of Doubao: Opening Agent-Grade Search Capabilities to Developers
ByteDance
Moonshot AI
The search capability behind Doubao App is now available as a standalone service for enterprises and developers. It returns structured data including authority level, publish time, summary, and Markdown excerpts, enabling AI agents to directly use the information without browsing full web pages. In benchmarks like SimpleQA and FreshQA, integrating Doubao Search improved accuracy by 70% and 40% over the base model without search.
Doubao Search, the search engine powering the Doubao App, has been released as an independent service for enterprises and developers via API, MCP, and Skills. It returns not only links but also structured fields such as source name, authority level, publish time, abstract, and Markdown excerpts from the original text. This allows agents to assess relevance and reliability without opening each page. Tests show significant improvements: +70% on SimpleQA, +40% on FreshQA, and leading gains on BrowseComp-ZH and Xbench-2505. The service integrates with models like Doubao-Seed-Evolving and provides 500 free searches per month. It handles dynamic queries like new product launches (e.g., Kimi K3 open-weight release), industry research (e.g., NEV policies and trends), and fast-changing domains (e.g., AI coding tools). Key features include authority grading, time filtering, site blacklisting/whitelisting, and vertical domain coverage for finance, commerce, and more. The goal is to make search a reliable, controllable tool layer for agents.
- Сокращения
- API = Application Programming Interface — интерфейс прикладного программирования
- MCP = Model Context Protocol — протокол контекста модели
- NEV = New Energy Vehicle — электромобиль/новое энергетическое транспортное средство
Source: QbitAI 量子位 —
original
