Deep Research
Deep Research is for questions that outlive a chat turn. It proposes a plan and waits for your approval, sends parallel researchers after each sub-question, checks every claim against the passage it rests on, and comes back with a cited report you can export into a workspace.
At a glance
| Handle | @deep-research |
|---|---|
| Availability | Enabled on demand under Build → Capabilities |
| Where you use it | The Research page in the main menu, or chat (the Orchestrator can hand it a whole research goal) |
| Tags | background |
| Model tier | advanced |
| Autonomy | trusted — the only built-in above supervised: it may write without pausing, but a destructive call still parks |
| Tools | web_search, web_fetch, search_knowledge, code_execution, generate_document, generate_image |
| Skills | research-method, source-triage, contradictions, report-writing, and three playbooks |
| Schedulable | Yes — a weekly brief is an automation |
Asking a question
| Sources | Web + internal, Web only, or Internal only (your organisation's knowledge, under your own access) |
|---|---|
| Depth | Quick — up to 3 sub-questions, a broad survey · Standard — 3–6 sub-questions · Exhaustive — 5–8 sub-questions, read closely |
| Playbook | Optional method for a familiar shape of question: market sizing (triangulate top-down and bottom-up), literature review (surveys first, studies weighed by design), vendor comparison (one fixed criteria matrix) |
| Workspace | Where the research, its evidence and its report live. Defaults to your Personal workspace — it is never picked silently |
Your research below the form lists every run with its date, researchers finished, cost and status.
How a run proceeds
The run view shows the phases as a strip — Scope · Plan · Research · Consolidate · Write & verify · Deliver — with budget meters for cost, searches, tokens and time underneath.
-
Scope
If the question is ambiguous, it asks you a few short questions in a form (a date range, a region, what "large" means). If nobody is there to answer, it states the assumptions it made instead. -
Plan — and you approve it
You see the sub-questions it will research, each with what an answer must contain, plus the estimated cost, searches and time. Approve, edit the sub-questions, or cancel — right in the chat or on the run page. If the estimate is over what your organisation lets you approve alone, an administrator approves the spend from their inbox. -
Research, in parallel
Each sub-question gets its own researcher. Researchers search broad then narrow, prefer primary sources, and record notes — each a claim tied to the exact passage it came from. A note whose passage is not in the source it cites is refused. -
Consolidate, then follow up once
Findings are merged and gaps or contradictions identified; up to one round of follow-up questions (markedf1,f2…) goes back out to researchers. -
Write & verify
The report is written only from the recorded notes. Every sentence is then checked against its passage: an unsupported sentence is removed, a partly supported one is rewritten or hedged, and the original success criteria are re-checked on the final text. -
Deliver
The report arrives with its Evidence and Verification sections and a computed confidence for each finding.
Watching the researchers
The report tabs are Report, Evidence (every note and the passage behind it), Sources, Verification (what the checker removed or rewrote, and why), and Plan & trace.
Confidence is computed, not claimed
- High — two or more independent primary sources (different domains) and the verifier found it supported.
- Medium — one primary source (an internal system of record counts), or two independent secondary sources.
- Low — a single secondary source, a partially supported verdict, or a note caught in an unresolved contradiction.
Taking the report with you
Export to Word, PowerPoint or Markdown into any workspace's Documents, with
numbered references, or Open in editor to keep working on it
with the Document Author. Citations
in the report render as [src:N] pills that open the source.
Administration
Under Admin → Governance → Research an org admin sets:
- whether the organisation uses the research engine;
- the cost threshold above which a plan needs approval, and how far over it a requester may approve alone (30% by default) — beyond that, an admin approves the spend;
- the limits every run stops at: cost, searches, tokens and time. A run that reaches one delivers what it has and says what is missing — it never simply fails.
An eval pack for the engine can be installed from Build → Capabilities; its planner and verifier cases run without invoking any researcher.
The classic run
With the research engine off, Deep Research runs as a single
background run: it writes research-plan.json
first, works the steps with a checkpoint after each (a worker that dies
resumes from the last one), reports progress to your inbox, accepts steering
at the next step boundary, and synthesizes a cited answer — marked
budget-limited if it hits its one-hour wall clock rather than truncating.