1. Conduit
  2. Features
  3. Research

Research

Research reports you can check

Approve a short brief, and Conduit searches the web, reads the full pages and writes a report with numbered sources. Every finding is quoted from its page, and the app checks the quote is really there.

  • Quotes checked by the app
  • Weak sources marked
  • Quick, Standard or Deep

How it works

A brief, a search, a checked report

Turn on Research in the composer, type /research, or pick it on Home, and ask your question in your own language. A worked example: the research with Exa tutorial.

  1. Approve the brief

    Research proposes up to six sub-questions, a scope and a depth. You edit them and approve before anything runs.

  2. It reads and checks

    It searches, reads pages and PDFs in full, and pulls out facts with the exact words that support them. A fact counts only if those words are on the page.

  3. You get a report

    A summary, findings per sub-question, open questions and numbered sources, kept as a document in your chat.

The brief

Agree the questions before it starts

Research turns your question into sub-questions, a scope and a depth. Change any of it, then press Start research.

  • Quick: up to 8 searches, 15 pages, 10 minutes.
  • Standard: up to 20 searches, 40 pages, 20 minutes.
  • Deep: up to 50 searches, 100 pages, 40 minutes.

While it runs

Every quote checked against its page

The model pulls out claims, each with the exact words from the page that support it. The app then looks for those words in the stored text of that page. A claim whose quote isn't there is left out, and the report says how many were.

It searches again for whatever is still unanswered, and you can watch the count of searches, pages and claims as it goes. Stop keeps what it found so far.

  • Sources rated for credibility. A weak source is marked, and its claims are attributed to it or left out, never stated as fact.
  • A review pass. Every sentence of the report is checked against the facts it cites before it is finished.
  • Pages can't give it orders. Page text only reaches a model call with no tools, so a page can't make it search, fetch or save anything.

The report

A report with numbered sources

The report opens beside the chat: a summary, findings for each sub-question, open questions, and the sources. The numbered citations and the source list are made by the app, not the model. The brief, findings and report follow the language you asked in.

Write from this report starts a Writing draft that uses its checked facts.

Straight answers

What Research does, and what it doesn't

What it does

  • Read web pages and PDFs in full, up to 20 MB, parsed on your device
  • Keep a claim only if its supporting quote is found on the page, checked by the app
  • Mark weak sources, and attribute or leave out what they say
  • Know today's date, and explain figures that differ by year, edition, definition or units
  • Name the sites it couldn't read when a sub-question goes unanswered
  • Keep the report as a document you can open, export or write from

What it doesn't

  • Run without web search turned on, or in local-only mode
  • Read the whole of a very long page: the model sees the opening plus the parts that match your questions
  • Read scanned PDFs, which have no text layer
  • Use your own Documents as a source yet
  • Guarantee the credibility rating: that is the model's judgement, while the quote check is the app's
  • Finish fast on a slow thinking model: Deep can hit its 40-minute limit

Research questions

How is this different from asking a chatbot with web search?

A chat answer with search may cite a page that doesn't say what the answer claims. Research keeps a finding only if its supporting quote is found in the stored text of the page, and that check is done by the app, not the model. The report also says how many claims were left out because their quotes couldn't be found.

Which search engine does Research use?

A local search backend you choose in Settings → Web search: Exa (the default, free and rate-limited without a key), Tavily, Brave, your own SearXNG, or DuckDuckGo. Search queries go to that service, and page text goes to the chat provider you picked.

Can Research run on a local model?

The model can be local, but Research needs web search, which local-only mode turns off. With local-only mode off, you can pair a local model with a search backend.

Can I turn a report into an article?

Yes. A finished Research card has Write from this report, which starts a Writing draft that prefers the report's checked facts.

Can Research run on a schedule?

Yes, as a step in a workflow. A Research step runs the same research as in chat, with the same quote check, and "Weekly research digest" is a starter for it. A scheduled run asks before it researches the web.

Ask your first research question

Free and open source. Turn on web search, pick Research, and approve the brief.

A release candidate: installers aren't OS code-signed yet, so expect a SmartScreen or Gatekeeper prompt on first launch.