Limits and quotas

Every limit that can change what the gallery does, in one place. Most readers never meet any of them; the one worth understanding in advance is the daily model allowance behind the search box, because it is the only limit whose effect lasts longer than a minute.

Rate limits

All of these are per client address, per minute, and all of them return a plain sentence rather than a silent failure.

What	Limit	On refusal
Questions to the search box	20 / minute	429 — “Too many questions in a row — give it a minute.”
Dashboard requests	6 / minute	429 — “That is a lot of requests in a row — give it a minute.”
Bundle downloads	20 / minute	429 — “Too many downloads in a row — give it a minute.”
View beacons	60 / minute	429 with an empty body

The duplicate check and the clarifying question on the request form share the request limit. Neither returns an error when refused — the duplicate check shows no cards and the clarifier shows no question, because both are asides beside a form that works without them.

These are smoothers, not accounting

Cloudflare documents the rate limiting binding as “intentionally designed to not be used as an accurate accounting system” — its counters are per-location and converge asynchronously. Measured against this site, it let 65 requests from one address through in a minute without a single refusal. That is fine for what it is for. It is not a spending limit, which is why the spending limit is a separate mechanism.

Content limits

Where	Limit
A question to the search box	800 characters
A dashboard request	10 to 1,200 characters
An email address on a request	254 characters
Draft length before a clarifying question is offered	40 characters
Dashboards per API listing page	60 by default, 100 maximum
Results the duplicate check shows	3

The answer engine’s own caps

Limit	Value	What happens at it
Searches per question	3	The search tool is withheld, not merely discouraged, and the model is told to answer from what it already has
Model turns per question	4	The turn budget ends and the reply is written
Wall clock for one model call	40 seconds	The call is aborted and the question returns a 503
Answer length	1,200 tokens	The reply is cut. Citations are written last, and the gallery keeps whatever was written before the cut
Results from one search	1 to 12, default 5	—
Results from a category search	The whole tagged set, up to 60	A question about a category is a question about a set, so it is not cut to the default

The daily model allowance

The search box is unauthenticated and calls a language model, which means an unmetered visitor could write the account’s inference bill. So there is a hard ceiling on what all of the day’s questions together may spend.

The ceiling	$1 per UTC day
In Cloudflare’s unit	About 90,900 neurons, at the published rate of $0.011 per 1,000
A measured question	About 18 neurons
Questions that buys	Roughly 1,230 per day
Checked	Before the model runs
Recorded	After it runs, whether it succeeded or not

The unit is the neuron because that is what Cloudflare bills and what each completion reports back — so the ledger is a record of money rather than an estimate from token counts, and it stays true across a price change or a model swap.

Why 1,230 questions rather than 780

Two models answer, and which one you get depends on how much of the day is left:

	Model	When
Primary	DeepSeek v4 Flash	Until 75% of the day’s allowance is spent
Economy	GLM 4.7 Flash	For the remaining 25%

Over a 30-question set they scored the same on recall — each found every dashboard it should have, and each refused every question the gallery has no answer for. So the choice was never about coverage. DeepSeek is more precise (95% against 85%) and far steadier at the slow tail (11.8 seconds at the 95th percentile against 32.3); GLM costs about a third as much to run.

Taking both means the allowance buys ~583 DeepSeek questions and then ~650 GLM ones instead of 780 DeepSeek ones, and the end of a busy day degrades to a slightly less precise answer rather than to no answer. Which model runs is read once per question, from the same spend figure the ceiling check already fetched — so the two cannot disagree, and one question is answered by one model the whole way through.

What happens once the allowance is spent

The search box returns a 429 and says so:

Search has used up today’s allowance and is resting until tomorrow — the filters below still work.

Everything that is not the model keeps working. The categories, the filter rail, every dashboard, every permalink and every download are unaffected, and the duplicate check on the request form drops to text-only matching rather than refusing. The degraded state is the site minus one feature, not a broken page.

The allowance resets at 00:00 UTC.

The ledger counts failures too

What a call consumed is recorded in a finally block, so a question that timed out or failed still counts against the day. A ledger that counted only successes would be one you walk past by making the model fail.

Caching

Resource	Policy
Dashboard images and thumbnails	Immutable for one year — a tightened redaction publishes at a new revision, so the old URL is simply never requested again
A dashboard permalink’s HTML	60 seconds, then revalidated
The sitemap	10 minutes
An export bundle	no-store — never cached, because two downloaders must not receive the same identifiers

Next steps

Ask about the dashboards What the search does with a question before any of these caps apply.

Public HTTP API Which endpoint each of these limits sits in front of.

Troubleshooting Telling a rate limit apart from a spent allowance — both are a 429.