網路搜尋

這是什麼

外部網路查詢和擷取能力。

白話解釋

想像派出一名偵察兵替你搜尋網際網路，收集相關頁面並交回乾淨可讀的文字，而不是你自己打開十個瀏覽器分頁。Hermes 可以查詢資料，也可以在知道某個頁面存在後拉取該頁面的完整文字內容。

為什麼重要

用於查詢的 web_search 和用於從頁面拉取結構化內容的 web_extract 涵蓋了兩個相關的工作，而 Hermes 為 Brave、DuckDuckGo、Exa、Firecrawl、Parallel、SearXNG、Tavily 和 xAI 提供了插件後端，因此兩者背後的實際提供者都是可以互換的。研究簡報循環工作流程最直接地依賴這個工具集，因為將分散的來源轉化為可供決策的簡報，始於可靠的檢索。由於其工具集中只有兩個註冊項目，它刻意保持輕量，而不是捆綁不相關的能力。

How it works

Two tools make up the web toolset: web_search, which returns up to five results by default, and web_extract, which pulls the actual content out of as many as five URLs per call (tools/web_tools.py). Long pages get head-and-tail truncated to a configurable character budget — 15,000 by default — with the full text saved to disk for later, so a single huge page doesn’t blow up the response. Which search engine actually runs behind either tool is a config choice (web.backend) among plugin-registered providers, including Exa, Firecrawl, Parallel, Tavily, SearXNG, Brave, DuckDuckGo, and xAI, so switching providers doesn’t change how the tool is called. web_extract also works directly on PDF URLs like arXiv papers, and it blocks any URL that looks like it’s carrying an embedded API key or that points at a private network address.

A concrete example

A founder wants to compare the pricing pages of three competitor SaaS products before a meeting.

1. Hermes runs web_search for each competitor’s pricing page URL.
2. It then calls web_extract on the three discovered URLs, pulling the actual pricing tables and feature lists.
3. The extracted content is returned to the model, which can compare the pricing structures directly rather than guessing from search snippets alone.