Agent时代搜索新需求 腾讯元宝搜索优化实测
A recent industry observation points out that automated requests now account for over half of all web traffic, marking the first time AI-driven activity has surpassed human-generated traffic in internet history. Against this backdrop, major AI players have rolled out new search capabilities: one leading model provider has launched its own search API, while another tech giant has announced opening its "元宝搜索" service to the entire industry.
Search functionality has evolved dramatically alongside AI development. For human users, search was about finding and reading relevant links; for large language models, it was about summarizing search results; now, AI agents rely on search to break down complex tasks into sub-questions, gather information, and perform reasoning. The core difference lies in judgment: humans can sift through results to identify credible sources, but agents lack this ability, so every piece of content they access is directly incorporated into their responses—making it critical to address gaps in how search serves agents.
When adapting search for agents, three key challenges emerge. First, results must be usable directly: agents need full text, not just titles or abstracts, which would otherwise require additional, slow, error-prone requests to fetch content and handle anti-crawling measures. Second, results must be trustworthy: agents cannot distinguish between official sources, SEO farms, and marketing content, so unvetted results can introduce misinformation. Third, speed is essential: agents often run multiple sequential searches for a single task, and even a small delay per query adds up quickly for users.
To tackle these issues, the 元宝搜索 service has implemented three targeted optimizations. First, it delivers full text content directly, eliminating the need for extra fetch requests—saving token usage and context space. Second, it integrates 157 authoritative sources across 13 industries, focusing on high-credibility domains like finance, healthcare, and government to boost result reliability. Third, it compresses web pages into 100 tokens each during indexing, so retrieval can reuse cached content, reducing computational overhead that grows with document length. These changes have shown tangible results: internal products using this service reduced the average number of search calls per task from 8 to 5.
A comprehensive test of this service covered 9 categories and 31 search queries, with each returning 10 results adjusted for time and domain parameters. The test evaluated performance across 8 key dimensions. For financial queries, results like stock prices and trading volumes were accurate and timely, reflecting the latest data. For authoritative information, the service consistently prioritized official sources: government portals for policy, national statistics bureaus for economic data, and judicial bodies for legal texts. For common knowledge and encyclopedia-style queries, it cited reputable outlets like state-run news agencies and industry analysis from leading media. In health-related searches, results were mostly correct, though source credibility varied slightly across queries, with some entries from general health platforms instead of top-tier medical authorities. Complex questions about industry connections or regulatory requirements reliably pulled data from government and institutional sources. For video-related searches, the service required parameter tuning, with improved results when restricting to specific platforms. Broad, vague queries about current events were less effective, requiring pre-processing by the agent to refine the question. Life service queries yielded many results but lacked structure and booking functionality, better suited for separate API integrations. For academic and technical content, results were mixed: official documentation was prioritized, but research papers from academic platforms were not always the top result.
Overall, the service excels in three key areas: authoritative sources, timely updates, and full text delivery. These strengths help agents access first-hand information, reduce delays, and cut token consumption. For domestic agent developers, these features can significantly streamline development, though it’s important to note that performance depends heavily on parameter configuration, and some use cases like life services or academic papers may require additional third-party integrations.
The service offers three access methods: a lightweight prompt-based approach for quick integration, a toolchain-compatible option for existing MCP systems, and direct API integration for native product functionality. Pricing is tiered: the basic version costs 14 yuan per thousand queries, the standard version 20 yuan, and the enterprise-level version 26 yuan (with enterprise applications eligible for the premium tier). New users receive 2,000 free queries monthly, and promotional offers can increase this to 6,000 queries with a valid invitation code.
