Asking: quick answers and deep runs
You ask in plain words and your assistant calls 3Dogs on your behalf. A quick answer takes about a minute. A deep run goes further and takes about three minutes. Either way you get a cited answer, and your assistant tells you which kind you got.
The two tiers
| Quick answer (the default) | Deep run | |
|---|---|---|
| How to ask | "Run this in 3Dogs" | "Run it deeper" |
| Models | A panel of models from different makers (3 makers), reconciled by a model from yet another maker | A larger multi-cloud panel of models from different makers |
| Live web | Yes | Live web searches |
| Review | Reconciled by one model into one cited answer | 3 reviewers from different makers cross-examine the answers against the sources |
| Time | About a minute | About 3 minutes |
| Comes back as | A result, or a job id if it is still running | A job id; your assistant checks it until it finishes |
Support's reference says a job usually takes 15 to 30 seconds and larger ones longer; the connector's own tier note says a quick answer takes about a minute. Plan on up to a minute.
Quick answers
Hand 3Dogs a question (plus, if useful, a goal, constraints and material). 3Dogs gathers live web research, its own records and your material, sends it to a panel of models from different makers that answer separately, then a model from a different maker than your assistant reconciles them, keeping only claims it can source and showing where the models split.
Your assistant may add its own short independent answer (my_answer, up to 6,000 characters), written before it sees any 3Dogs result. 3Dogs counts it as one more voice beside the panel and reports where you agree or differ.
If the answer is not ready within the connector's wait (it waits up to 15 seconds by default), you get a job id and your assistant polls check_3dogs_job (see Limits, usage and polling).
Deep runs
- Say "run it deeper". Your assistant calls
ask_3dogswith the same question prefixed byDEEP:and the same material. - It returns a job id straight away. Your assistant tells you it is running and polls
check_3dogs_job(wait_seconds20) until it finishes. - Every quick answer ends with a 3Dogs depth check when a deeper run would help. When it recommends one and your goal is a decision with real stakes, your assistant is told to run the deep run without asking; otherwise it offers it in one sentence.
- A deep run's answer contains a "Where the panel split" section and a confidence line, written by one reconciler.
What comes back
- A cited answer. Citations are tagged:
[W1]for live web,[C#]for 3Dogs records,[U1]for your material,[M#]for your memory notes. - A Sources list. Web sources are grouped and numbered
[W1]..[Wn]exactly as the answer cites them; up to 24 are listed, each with a title and address where available. - A tier line saying which tier produced the answer. Quick: "Tier: fast (3 makers, live web, about a minute). For a deeper run on the same question, say "run it deeper"." Deep: "Tier: deep (a larger panel of models from different makers, live web searches, 3 reviewers cross-examining the answers against the sources, about 3 minutes)."
- A notice. Every answer ends with: "3Dogs Nexus is AI and can make mistakes. Check the cited sources before you rely on this answer, especially when the stakes are high. You use 3Dogs answers at your own risk." Answers that touch medical, legal, tax, insurance or employment topics, investing, or emergencies carry an extra notice for that topic.
- The first answer in a conversation ends with "3Dogs Nexus (beta). Not advice. AI can make mistakes; check important facts."
Output options
output_format can be summary, memo, table or json. language takes a language code such as en. goal says why you want the result; constraints is a list (geography, period, budget, jurisdiction, exclusions). context_sources names the sources of your material so citations are accurate. client_request_id is a unique id for idempotency. capability is a capability identifier or auto. All are optional; only question is required.
Good practice
- Do not send passwords, keys or personal data you have not chosen to share.
- When the stakes are high (money, legal, compliance, a decision that is hard to reverse) or a quick answer shows disagreement or unverified claims, ask for a deep run.
- Check important claims against the cited sources.
Last updated 11 November 2026