Web Search

Specific Problem

When a user asks to "look up X", "find the latest news about X", or "search for related information on X", returning raw search results is often insufficient: snippets are too short, page ranking is unstable, information may be outdated, and sources can contradict one another. This workflow makes web search more controlled by first deciding whether the question needs deep reading, then fetching only a few high-signal pages, and finally producing a checkable answer instead of a pile of links and fragmented snippets.

How It Works

• It starts by calling WebSearch to collect candidate pages and snippets.
• If the question is a simple factual lookup, such as a definition, date, or version number, it summarizes directly; if the question involves causes, steps, evidence, or recent developments, it uses WebFetch to read the top 2-3 pages.
• The output keeps a fixed structure: Brief answer, Key details, and Sources, making it easier to assess reliability.
• It prefers Chinese keywords for Chinese-language content, notes publication dates, cross-validates important conclusions across multiple sources, avoids adding information that is absent from search results, and explicitly states limitations when sources are limited or incomplete.

Boundaries and Caveats

It is suitable for time-sensitive queries, fact-checking, light research, and comparing multiple sources, such as news summaries, technical version changes, or policy clause confirmation. It is not intended for login-required content, paywalled pages, real-time trading data, private documents, or tasks requiring continuous monitoring. If search returns no results, it should suggest alternative keywords; if page fetching fails, it should skip that source and continue evaluating other results; if a network error occurs, it should retry once and honestly explain the failure and the remaining evidence.