AI deep research that returns a report with sources
An autonomous agent searches the open web, compares perspectives and writes a structured report with inline citations and a bibliography. It costs 500 credits and requires the Expert plan.
Free credits · No credit card · Instant access What changed recently
What this solves
Deep Research turns one research question into a structured report. An autonomous agent plans the search, reads sources on the open web, and writes the answer with inline citations and a bibliography at the end. It runs inside SmartChat, uses your project sources as extra context, and costs 500 credits on the Expert plan.
How it works
-
Ask one clear question
Open SmartChat, turn on the Deep Research button in the composer and write what you want answered. The narrower the scope, the more useful the report: one question per run, not a list of topics.
-
Follow the run
The agent plans, searches and reads while you wait. A progress panel shows the status, a timer and an expandable log with the agent's own notes. A run takes minutes, not seconds, so leave the tab open.
-
Read the report and keep going
The report arrives in the conversation with sections, inline citations and a bibliography at the end. It is saved to the conversation history, so follow up questions continue from there at the normal message price.
Method
What happens to your file, step by step, and where the result stops being a decision made by a model.
How it works inside
The question is wrapped in a fixed six-point instruction (identify the question, find verifiable sources, present multiple perspectives, separate consensus from debate, add a bibliography, state when data is missing), plus up to 40,000 characters of the project's own sources. That prompt goes to Gemini deep-research-pro-preview-12-2025 through the Interactions API in background streaming mode. Searching, reading and report writing happen inside the provider's agent. Reasoning summaries stream to the log panel as they arrive.
Where AI is used, and where it is not
The model does the substance. Xplore Dados builds the prompt, streams the output and charges; it does not run the search, choose the sources or check the report's citations. The agent decides what to look for and what to cite. There is no structured source list our code can inspect, so nothing here strips a fabricated reference. Only the provider fallback and the billing are deterministic.
What it accepts
- One question in free text, Portuguese or English (the session language selects the prompt wrapper).
- Optionally, the sources of the SmartChat project the conversation belongs to, truncated at 40,000 characters. A conversation outside a project sends the question alone.
- An Expert plan with active or trialing subscription, or an unlimited account. Any other plan is refused with HTTP 403.
- Attachments are recorded in the history row, but the research prompt itself is text only.
What you get back
- One Markdown report streamed token by token and rendered to HTML, with the sections and bibliography the model chose to write.
- Progress updates: the provider status, the agent's reasoning summaries in the log and, on the OpenAI path, a still working signal with the elapsed time every 4 seconds.
- Two rows saved to the conversation history (question and report), so the report reloads with the thread.
- No structured source list and no reference objects. The bibliography exists only as prose inside the report.
What this tool does not do
- It does not verify the references the agent writes. No DOI is resolved and no link is fetched by our code.
- It does not read the PDF loaded in the current conversation. Only project sources are injected, and only when the conversation belongs to a project.
- Flat 500 credits per run, charged after the report finishes and before it is saved. A run that returns no text is not charged. The same question in the same conversation within 120 seconds counts as a retry, not a second run.
- There is no provider choice. Every run starts on Gemini; gpt-5.5-pro is only the automatic fallback when Gemini errors, and a provider sent by the client is ignored.
- A run takes minutes and is bounded by a 15-minute read timeout. Gemini polling gives up after 90 attempts spaced 10 seconds; OpenAI polling after 180 attempts spaced 5 seconds.
- Plan
- Expert
- Cost
- 500 credits
What you get
-
A report, not a chat answer
The agent is asked for a structured research report: clear sections, data tables where they fit, inline citations and a bibliography with full references at the end.
-
Evidence from the open web
The run does its own searching instead of answering from memory. The instructions ask for real, verifiable sources, for multiple perspectives, and for saying when data is unavailable instead of guessing.
-
Your project sources go in
When the conversation belongs to a project, the academic sources in that project are sent as context and cross referenced with what the agent finds online.
-
You can see it working
Status, timer and an expandable log with the agent's notes. When a run falls back to OpenAI, the server also sends a still working signal during long stretches with no text, so a slow run is not confused with a hang.
Changelog
Every line below is a change that actually shipped, dated by the day it went out.
-
Latest
- The 500 credits are now charged before the report is published: if the charge fails, nothing is billed and the research is not saved to the conversation.
- A reconnection or retry of the same research no longer charges twice.
Show 5 earlier updates
-
- Price moved from 30 to 500 credits, to cover the real cost of a run with web searching.
- Gemini became the only provider you run on. OpenAI stays wired in as the automatic fallback when Gemini fails, and the provider selector left the composer.
- When the run falls back to the other provider, the log now says why (quota exceeded, rate limit, timeout) instead of switching silently.
- On the OpenAI path, the progress panel now shows a still working signal from the server during long quiet stretches, both in the log and on the status label. The panel also renders bold and italic in the agent's notes.
-
- The automatic OpenAI fallback moved to gpt-5.5-pro, the replacement for the model that was being retired.
- The credits dashboard started showing the same Deep Research cost that is actually charged.
-
- Deep Research became Expert only: the button no longer appears in the composer on other plans.
- Plan and balance are checked before the run starts, and the message says which of the two is missing.
-
- Progress panel for the run: status, a timer and an expandable log with the agent's notes.
- The wait limit was raised so runs of several minutes are not cut short.
-
- First version of Deep Research inside SmartChat: one question becomes a structured report with inline citations and a bibliography, shown as it is written.
- Academic sources from the project are sent as context and cross referenced with what the agent finds on the web.
- It launched at 30 credits and with no plan restriction. Both changed later.
Questions and answers
Why does it cost 500 credits?
It is the most expensive action on the platform, and that is deliberate. One run is a multi agent task that plans, searches the web, reads pages and writes a long report over several minutes, so the provider cost of a single run is measured in dollars, not cents. The price was 30 credits until July 2026 and was raised to 500 to cover that real cost. Every other action in Xplore Academy is cheaper because none of them do this.
Do I need the Expert plan, or are credits enough?
You need Expert. The button only appears in the composer for accounts with an Expert subscription that is active or in trial. On Starter and Pro it is not shown, and a direct request is refused and you are sent to the credits panel, where the plans are listed. Credits alone do not unlock it.
How long does it take, and what happens if I close the tab?
Runs take minutes, and the server waits up to 15 minutes for the provider before giving up. If you close the tab or lose the connection during the run, the research is interrupted: nothing is charged and nothing is saved to the conversation, so you would need to start over. The credits are only charged when the report is delivered, and if that charge fails the report is not saved either.
Can I trust the sources it lists?
Treat the report as a well organized starting point, not as a verified bibliography. The agent is instructed to use real, verifiable sources, to separate scientific consensus from open debate and speculation, and to list full references at the end, but it can still describe a source imprecisely. Open the links before citing anything in your own work. Deep Research finds and summarizes evidence: auditing a finished reference list is what the Reference Checker does.
How to cite this tool
Used it in your research? Here is the reference, already filled in with the version you are looking at and today's access date.
ABNT (NBR 6023)
LESSA, P. W. B. Deep Research. Versão 2026.08. [S. l.]: Xplore Dados, 2026. Disponível em: https://xploredados.com/en/tool/deep-research. Acesso em: 21 ago. 2026.
APA 7
Lessa, P. W. B. (2026). Deep Research (Version 2026.08) [Computer software]. Xplore Dados. https://xploredados.com/en/tool/deep-research
BibTeX
@software{xploredados_deep_research_2026,
author = {Lessa, Patrick Wendell Barbosa},
title = {Deep Research},
organization = {Xplore Dados},
version = {2026.08},
year = {2026},
url = {https://xploredados.com/en/tool/deep-research},
urldate = {2026-08-21}
}
Start with the free credits
Create an account and test the tools before deciding on a plan.
Create free account