Deep research

Deep Research (Deep Research) is an independent, multi-round research streamline: the model first sets up a research plan, then performs multiple cyber searches and page readings, cross-validates the assertions, and ultimately generates a copy. ** Complete report with reference to source ** Its tool budget is significantly increased compared to ordinary questions and is suitable for overview questions that require multiple sources of support.

The way to open.

1. In the input box toolbar click the telescope icon "** Researching ** The button, the button lights up as the mouse tail green is opened.
2. If the question is typed and sent, the message goes through the Deep Research process.
3. The switch remains on and the follow-up message continues in the research mode until you turn it off again.

The display of the “Research” button must meet three conditions simultaneously:

Conditionally	by whom to control.
The current model opens up Deep Research	The administrator exposes the Deep Research switch on the model editing page, see Channel and Model
Your user group has research rights.	Administrator configuration in the user group function switch.
Currently no painting.	The study button automatically hides when selecting an image model

info

The mobile end "Study" entry is received in the "+" menu of the input box; the "+" button shows a high bright circle when the study is opened.

Difference from ordinary models

Common question-and-answer models can also search online, but the budget is limited and there is no systematic verification. Deep Research raises the overall single-wheel tool budget:

Tools	The normal upper limit.	The Deep Research Model
Search by Web_Search	16	40
Page crawling (web_fetch)	12	25
Download the image (fetch_image)	16	12
Images are generated (image_generate)	8	4
Python Execution (Python Execution)	16	8
All tools in total.	48	150

In addition to the budget, the process is different: the usual model is "model want to search for"; the research model is "first dismantle the plan, then cover each sub-problem as planned, and finally unify verification and writing".

Explaining the process

One Deep Research is divided into several stages, all of which are driven automatically by the engine:

1. ** Planned **: Divide your questions into 3 to 6 complementary sub-questions, identify the type of research (concept interpretation, contrast, trend, technology, market, decision) and determine the reporting structure. The program will deliberately include at least one counter-or critical perspective query, avoiding information gaps; timely queries automatically take the year, with a technical topic containing at least one English query.
2. ** Extensive search and reading. **: Multiple cycles of web search are performed in parallel around each sub-question, and the entire page is thoroughly read. The pages that have been visited are automatically weighed, and the reading order prioritizes choosing sources with higher credibility and more domain names.
3. ** Coverage of Audit ** Check whether each sub-question has sufficient independent source support, mark evidence-weak or uncovered sub-questions, and supplement a new round of search.
4. ** Cross verification ** Divide the evidence collected into three categories, directly determining the wording in the report:
   • ** has been confirmed ** Two or more independent sources agree, directly stated in the report;
   • ** There is controversy ** Conflict between sources, reports present the views of the parties separately, and do not force mergers;
   • ** not verified ** Only a single source, and the report indicates "by a certain source."
5. ** Writing the report ** The main model stream generates the final report, while writing and rendering.

Source credibility classification

Each source is ranked A to D level 4 by domain name inspired, and is displayed on the source tab of the Progress Panel:

The class	Meaning of
A	Official, Academic and Standard Organizations
B	Autoritative media, research institutions
C	Blogs, communities, and unknown domains
D	Anonymous forum, no signature content

The A/B-grade badges are bright with a high-green tail, and the engine prioritizes the digestion of highly reliable sources in reading order.

The Progress Interface

After sending, the research panel with the telescope icon appears above the response, which is automatically launched and updated in real time during the study:

• ** The title ** Study Theme, Progress Counting N/M (Number of completed sub-questions / total number of sub-questions).
• ** The research plan ** Sub-question list, each pre-existing status mark: the gray circle is for study, the green pulse point is in progress, and the green pair is completed.
• ** The Source (N) **: Double source card column, showing the title, domain name and A to D credibility badges; scratch the failed source to warn the triangle to be marked.

After the report starts output, the panel automatically folds into a row of summary, which can be reopened at any time to view the complete list of plans and sources.

Production: Report with Reference

The final report is presented as a message stream, the structure changes with the type of study (comparing the general overview table with the scenario recommendation table, the trend class indicator table and the short and medium-term outlook), and jointly complies with:

• The beginning is a metadata line (date of study, scope description) and an overview;
• Identification of key numbers;
• The confirmed conclusions are directly stated, the opinions of the parties are disputed, and the single source information is marked.
• Data marking time;
• Clearly define research limits;
• List of reference sources at the end.

A report is a typical AI response: it can be copied, re-generated (re-run a study round, start a new branch), continued query, spell-out conversation, and also included by a conversation sharing link.

Interrupting and running.

• In the course of the research, you can "stop generating" at any time: the content that has been produced is retained; according to the platform's "send instant" billing rules, the part that has been consumed is normally counted with the amount and credits.
• Report on “Rebuilding”** A full round of research. ** (Re-search and verify) to save in parallel branches, with < N/M > Contrast with the old report.

Applicable scenes and time

** Suitable to **:

• A factual overview of the need for cross-source verification (“Major advances in the field of X over the past two years”);
• Comparison options (“A and B scenarios of advantages and advantages respectively”);
• Trends and Market Analysis;
• Hope to get a research task with a verified list of references.

** not suitable **:

• Simple factual questions or chatter (normal mode is faster, one or two searches are possible);
• Pure code, pure computational tasks (delivered to The Python Sandbox);
• Depends only on your own documentation. The knowledge base It is possible).

Time and cost tips

Deep Research performs dozens of searches and scraps, then verifies and writes. ** It takes significantly longer than normal questions, usually minutes. ** During the period, you can "stop generating" interruption at any time. The task model called by the study process and the main model are charged normally, the amount of consumption and credits is higher than the normal conversation, please turn on as required.

The administrator side.

Deep Research relies on two background configurations, and if the average user doesn’t see the “research” button or the search always fails, contact the administrator to check:

1. ** The search back. **: SearXNG (self-deployed) or Serper-compatible search services need to be configured in the background; unconfigured research fails to obtain evidence, and reports are downgraded to unreferenced general answers.
2. ** model and user group **: Open research entry by model, grant research rights by user group, configure methods see Channel and Model。

Common Questions Review

phenomenon	Possibly the cause.
There is no “Study” button.	The current model is not open research, or your user group has no research permissions, or the image model is currently selected
The report does not cite sources.	Administrators don't configure the search backend, research downgrade to ordinary answers
The progress panel stays on the same problem for a long time.	The target site scans slowly or fails, the failed source is marked with a warning triangle, and the engine automatically changes the source continues.
Report significantly shorter than expected	By clicking “Stop Generating” or the evidence of the problem itself is scarce (the “Limit” section of the report will explain)