Skip to content
Build Perch
Browse documentation

Web & research

Short answer: The AI’s research tools cover live web search (DuckDuckGo by default, or Brave with a key), full-page and PDF reading, Wikipedia lookups, arXiv paper search, GitHub code/issue search, and image search (Openverse or Brave), all built in, with nothing to connect except an optional Brave key. They work in modes like Assistant by default.

What research tools does the AI have?

Six tools cover live, external research, bundled into the External Research skill:

  • web_search: snippet-only web search (title, URL, short snippet) for anything time-sensitive: news, prices, current events, anything past the model’s training cutoff. With Brave as the provider, results also carry a published date, and the AI can restrict a search by page age (past day, week, month, year, or a date range); DuckDuckGo results say so plainly when no date filter applied.
  • web_fetch_page: reads a document in full when a snippet isn’t enough: web pages (returned as clean text with headings and links intact), PDFs, and plain-text or JSON files. Long documents are read in ~3,000-word installments the AI can page through. If a link is dead (404, site down), the tool automatically falls back to an archive.org snapshot and labels the result with its capture date.
  • web_fetch_then_query: for long pages (Wikipedia articles, RFCs, papers), embeds the page on-device and returns only the sections relevant to the AI’s question, instead of the whole body.
  • search_wikipedia: encyclopedia lookups for stable factual topics (history, science, geography, biographies), preferred over web_search for anything unlikely to have changed.
  • search_arxiv: scientific-paper search (computer science, machine learning, physics, math, and more). Every result includes the paper’s PDF link, which the AI can read directly with web_fetch_page.
  • search_github: code, issue, and repository search on GitHub; see the GitHub page for token details.

Answers built from these tools cite their sources: numbered references like [1] in the reply are clickable links to the pages the AI actually used, and they keep working after you close and reopen the app.

Image search is separate, part of the Images & Vision skill: image_search finds existing images on the web so the AI can attach one, crediting its license and creator, rather than generating something new.

Which search engine does it use?

Web text search defaults to DuckDuckGo. Switch it to Brave Search under Chat Configuration → Settings → Internet Search (the “Internet text provider” setting) for richer results, at the cost of needing an API key. Image search defaults to Openverse. Every result carries a license (CC-BY, CC0, and so on). Switching image search to Brave returns a broader pool pulled from the general web, tagged “license unknown” with a reminder to verify at the source, since Brave doesn’t return licensing metadata. Wikipedia and GitHub search need no key at all.

How do I add a Brave key?

  1. Get an API key from Brave Search’s developer site; the Brave Search row in Integrations links to sign-up.
  2. Paste it into that row and save.
  3. Switch “Internet text provider” and/or “Internet image provider” to Brave wherever you want it.

One key powers both Brave text and Brave image search; web_search keeps working through DuckDuckGo even if you never add one.

Good to know

  • URL-fetch tools are withheld on untrusted, unattended runs. A workflow or agent triggered by an incoming email, Discord message, or file change can’t call web_fetch_page or web_fetch_then_query, closing off a path where injected content could exfiltrate data through a crafted link. Manually-run agents and interactive chat keep full access.
  • Automated fetches honor robots.txt. Agent and workflow runs respect a site’s robots.txt rules; fetches you direct in an interactive chat behave like your browser.