Research: the assistant searches, reads sources and writes a report with links

1How to turn research on

1. Click “+” by the input field

   In the main chat at gibrid.ai/chat there is a plus on the left of the input field. The menu has Attach files, Media library, Image, My documents, Research and Video in Studio.

2. Choose “Research”

   The menu closes, and a chip “Research · ≈ N G, at most 20 G” appears in the input area. If the estimate hasn't loaded yet, it will say “Research · up to 20 G”. Hover hint: “Next message: many sources and a cited report. Click to cancel.”

3. Write the topic and send

   Describe what to find out: “Compare three note-taking apps: prices, platforms, pros and cons”. The more precise the question, the better the report.

4. Wait for the report

   First the search and page reading go on, then the assistant writes the report. A card with the list of sources stays under the reply.

2What the assistant does

• Sub-questions. The assistant splits the topic into several narrow search queries (no more than six) and searches them at the same time using Brave search.
• Choosing sources. Up to 10 different pages are taken from the results, in turn from each sub-question, to cover different sides of the topic. No more than two from one site.
• Reading. Pages are opened at the same time through a protected loader: only ordinary sites over http and https, internal and local addresses are blocked, every redirect is checked again, and waiting is time-limited. Only text and HTML are read: an excerpt of up to 2,600 characters is taken from each page, no more than 26,000 in total.
• If a page didn't open (it needs a login, the site is closed, it is a file such as a PDF), a short snippet from the search results is used. The assistant is expected to say plainly what the sources don't cover.
• The report. First a short conclusion, then sections. Every fact has a source number in square brackets [1], [2], and at the end there is a “Sources” list with links. In Auto mode a strong model writes the report, and if you chose a model by hand, that model does.

The search and page reading themselves are not charged separately: you pay for the model's work, that is, for reading the excerpts and writing the report. If you need a file, ask after the report: “Put the report into a Word document”. Read about documents on the page about documents in chat.

3What it costs and where the 20 G cap is

The chip shows an estimate before the run. It is calculated for a strong model and includes two rounds of the model's work with a request that holds the excerpts from the sources, plus a report of about 2,800 tokens. Because of the plan's markup the estimate differs from person to person: from about 4.1 to about 7.5 G.

4How to cancel the mode

1. Click the “Research” chip

   It is clickable: it has a cross, and on hover the hint says “Click to cancel”.

2. Send the message as an ordinary one

   The chip disappears and the question goes into the chat without research.

If you have already sent the message, the mode has switched itself off: it applies to one message. To do another research, choose “+” → Research again.

5Limitations

• The report is built only from the pages that could be read. Closed sites, paid articles and pages that need a login or scripts are not read.
• Files (PDFs, documents, archives) behind links are not read: only a short snippet from the search is used.
• No more than ten sources and no more than two per site are taken; an excerpt from a page is short (up to 2,600 characters), so long materials are not read in full.
• It is not a guarantee of truth: check important facts using the links in the source list.
• The mode is available only in the assistant's main chat and only for signed-in accounts.

Related