Research Desk is for the questions where being roughly right is not good enough — because someone will ask where a number came from, or because an agent is about to act on the answer.
What are the current rate limits and pricing tiers for the major LLM inference providers?
An agent that researches inline burns its context window on raw pages and then acts on whatever it half-remembers. One call returns a compact brief where every number is attributable.
The agent can check a claim before acting on it — and when it hands its output to a person, so can they.
How have the top three project-management tools changed their pricing in the last twelve months?
The deep rung searches six angles, reads long, and then goes back for whatever the first pass could not settle — which is usually the thing someone asks about in the room.
Every line in the deck traces to a page you can open while being questioned.
What regulatory filings has this company made about its supply chain since 2024?
Narrow it to the domains that count — a regulator, a filings database — and everything outside them is excluded at the search, not filtered out afterwards.
A brief with an unverifiable claim in it is worse than no brief. Unsupported claims are dropped and the count is reported.
What are the practical trade-offs between the major vector databases for a 100M-embedding workload?
Three angles surface the vendor claim, the independent benchmark and the complaint thread, and the brief says plainly where they disagree — citing both sides in the same claim.
Disagreement is reported, not averaged into a bland middle.
What has been announced about on-device AI chips in the last 14 days?
Set a recency window and the search never sees anything older. Quick is one search read shallowly, which is the right shape for a question about what just happened.
Cheap enough to run daily; cited enough to publish from.
What approaches have been proposed for reducing hallucination in retrieval-augmented generation since 2024?
Point it at the preprint servers and journals you trust, read long, and let the second pass chase the gaps the first pass named.
Open questions come back as a list — the tool tells you what it could not settle instead of papering over it.
15 credits for a quick look, 35 for the standard brief most people run.