Answers you can actually cite
Most AI answers ask you to take them on faith. Deep Research doesn't. Instead of a single guess from a single model, it runs a real, multi-step investigation - searching the web, reading your knowledge base, weighing what it finds - and then writes one clear answer where every claim links back to the source it came from.
No black box. No "trust me." Click any citation and read the original yourself.
Start a research run ยท See plans that include it
How it works
Ask a genuinely hard question - "AI inference market size by vendor, with the 2024-2025 growth forecast" - and Deep Research doesn't just reply. It plans.
- It breaks the question into research steps and runs them in parallel - web searches, knowledge-base lookups, and reasoning all at once, instead of one slow lookup at a time.
- It gathers and ranks sources. Every source is captured with its URL, the exact passage it pulled, and a confidence score - so the evidence is structured, not just a pile of links.
- It de-duplicates the facts. When five sources say the same thing, they collapse into one finding - ranked higher for the agreement, with all five citations kept.
- It synthesizes one answer with inline citations, so you see the claim and its evidence side by side.
Because the heavy lifting runs in the background (a thorough pass can take 30-60 seconds), it never freezes your chat - you keep working and the report lands when it's ready.
Grounded in the web and in your own material
Deep Research doesn't only read the open web. It blends live web search with retrieval from your own knowledge bases, so an answer can weave together a public statistic, a passage from your internal handbook, and a quote from a primary source - each one cited. That makes it as useful for "what does the market say?" as for "what does our documentation say?" - in the same run.
Honest about confidence, not just confident
A normal chatbot states everything with the same swagger, whether it's certain or guessing. Deep Research carries a confidence signal through to the answer: findings backed by many agreeing sources rank higher, thin or contradictory evidence is flagged, and where the sources genuinely disagree, it shows you the disagreement instead of papering over it. You're reading a researched position, not a polished hunch.
Why it matters
A normal chatbot will confidently invent a statistic. Deep Research can't hide behind confidence, because every sentence carries its receipts. That's the difference between an answer you'd paste into a group chat and one you'd put in a report, a brief, or a paper - or in front of a client, a board, or a regulator.
And it's honest about cost the way the rest of ToRun is: a research run is one line item on your bill, fully itemized - the searches, the model, the tokens - with no surprise "premium feature" surcharge buried somewhere.
What you get
- Every claim is cited
bi-link-45deg- no floating facts. Each statement links to the exact source and passage it came from - click through and verify it yourself. - Multi-step, in parallel
bi-diagram-3- web search, your knowledge base, and reasoning run together as a research plan, so a thorough answer arrives fast. - Web + your own knowledge
bi-globe- blends live web results with retrieval from your uploaded documents, citing both in one synthesized answer. - Sources ranked & de-duplicated
bi-layers-half- when many sources agree, the finding is merged and ranked higher; every citation kept, the noise stripped. - Hallucination-resistant
bi-shield-check- because each fact must point back to real evidence, the answer can't quietly make things up. - Runs in the background
bi-hourglass-split- a deep pass takes 30-60 seconds and never blocks your chat. Keep working - the cited report drops in when it's done. - Transparent cost per run
bi-receipt- one research run, one itemized line on your bill - searches, model, and tokens all visible.
Built on sonar-deep-research
Deep Research is a first-class mode inside ToRun chat - not a bolt-on. Pick it from the mode switcher and ToRun's capability-first routing sends your question to sonar-deep-research, the model purpose-built for cited, multi-step investigation. Follow-up questions reuse what's already been found, so a second ask only pays to research what's genuinely new - and the whole session keeps its sources, so the report stays auditable long after it's written.
