Skip to main content
Private access is opening up — request an invite
askFinz
How it works

We don't rentour search results.We read the web.

Most AI products answer your question by paying someone else for search results. askFinz built its own index instead — harder to build, far cheaper to run, and the one thing every answer depends on is ours rather than rented.

4.4 KB
Per stored page
everything needed to find and rank it
1.7
Passages per page
each one separately searchable
47
Collections
each filed as the kind of thing it is
read pathpage/
a page arrives/
 ├── L1/
 │    ├── handles      ~2 in 3
 │    ├── speed        milliseconds
 │    └── cost         negligible
 ├── L2/            only if L1 could not
 │    ├── handles      ~1 in 3
 │    └── cost         higher
 └── L3/            only if L2 could not
      ├── handles      <0.1%
      └── cost         highest

kept afterwards/
 ├── the meaning, not the keywords
 ├── citable back to the source
 └── a few KB per page on disk
cheap firstescalate rarely
Reading a page

Three stages. Almost nothing reaches the third.

Pages differ enormously in how hard they are to read. Rather than treat them all the same, each one starts cheap and only escalates if it has to.

How a page gets read3 stages · automatic · no queue-jumping
L1First pass

≈ 2 in 3

Most pages are read here, in milliseconds, at almost no cost.

only if needed →
L2Second pass

≈ 1 in 3

A substantial minority need more work to read properly, and escalate automatically.

only if needed →
L3Assisted

< 0.1%

The stubborn few — handled with the heaviest tools, only where it is warranted.

A page moves to the next stage only when the one before it could not read it properly. That is the whole cost model: the heaviest machinery runs on well under a percent of what we read, so reading all of it stays affordable.

What it costs the sites we read

Most sites never notice us, because there is very little to notice.

Starting cheap is usually described as our saving. It is just as much the site's. Every page starts at the lightest method that can read it, and only the small remainder that genuinely needs more gets more — so a publisher sees ordinary requests spread out over time rather than a burst that shows up on their dashboard as a problem to deal with.

A fraction of the weight

The heaviest machinery runs on well under a percent of what we read. Every page starts at the cheapest method that can do the job, which is most of why reading the whole web is affordable — for us and for the sites.

Paced, not hammered

Reading is spread across time and across the machines doing it. No site sees a flood, because nothing about how we work rewards going fast at one publisher.

Read once, reused

A page read once answers questions for everyone. Nobody's search sends fresh traffic at a publisher, which is the difference between an index and a tool that fetches on demand.

We take a check at face value

We do not buy captcha-solving services. In practice the question comes up rarely — the overwhelming majority of what we read never puts a check in front of us, and only a fraction of a percent ever reaches L3 at all. Where a site does ask, we treat that as the site's decision to make. Publishers who want their material read can simply tell us, and it gets read properly rather than argued with.

Every machine reading is accounted for

Reading runs on infrastructure we operate and can account for, every machine enrolled and identifiable to us. None of it is anonymous capacity bought by the hour from a broker. That is why we can say exactly what the fleet does, show it working in the open, and answer for how it behaves.

Being light is the point rather than a side effect. A reader gets a faster answer, a publisher gets left alone, and we get to read far more of the web for the same money.

The difference

Owning the index changes what we can offer.

We read the web ourselves

Most AI products buy their web results from someone else's index. askFinz runs its own — the reading, the ranking and the storage are all ours.

Nothing is rented

No per-query bill to a third-party search provider, and no ceiling set by their rate limits. What it costs to answer a question is something we control.

It cannot be taken away

Providers change terms, raise prices and close APIs. A product built on a rented index inherits every one of those decisions. Ours doesn't.

It fits what we're asked

Because the index is ours, it can hold whatever turns out to matter — new subjects, new formats, new collections — rather than only what a supplier chose to sell.

Where the work happens

Spread out, on purpose.

The work is shared across many machines rather than concentrated in one place. That keeps it cheap, keeps it close to the pages being read, and means no single failure stops it.

Our own index servers

The core of the fleet, always on and always ours.

Every surface askFinz runs on

The extension, the desktop app, devices running askFinz OS — each able to lend a hand.

Machines close to the page

Reading happens near the source rather than routing everything through one place.

Partner installations

Publishers who want their own material read well can host a node themselves.

Connecting…
61 GB
Memory pooled
across every machine reading
10
Machines online
dedicated to reading
54
Processor cores
spare capacity, pooled
530k
Pages read by the fleet
these machines' own tally
483 GB
Web pulled down
read, cleaned, filed
1167/s
Shared card throughput
across 2 shared graphics cards

You can watch it running at live status, which reads from the same system this page does.

What it buys you

Reading ahead of time means answering now.

Because pages are read and understood before anyone asks, a question does not wait on the web. These are the live numbers from the running system.

Connecting… · 25 queries served
486 ms
Cold query
fresh, never seen before
<1 ms
Cached query
popular / repeat
4%
Cache hit rate
served without re-search
0%
GPU-embedded
with CPU fallback

query → vector on the GPU pool · payload-indexed SafeSearch filter · deterministic result cache

How a page becomes an answer

seen → read → understood → answerable

Four steps, and no waiting in between.

A traditional index rebuilds on a schedule, so what it knows is always a little out of date. Here the last step happens the moment the first one does — a page seen now is answerable now.

01

Seen

A page is reached — by the crawler working through the web, or by the extension as somebody reads it.

02

Read properly

Menus and clutter come off. Documents are read all the way through. Pages that only assemble in a browser get built first.

03

Understood

What the page means is worked out and stored — not just the words on it, so a question can match an idea rather than a keyword.

04

Answerable

It becomes searchable immediately, and any answer drawn from it cites the page it came from.

Questions

Is the index available on its own, as an API?

Not today. The index powers askFinz Search and the apps built on it, and is not currently sold or licensed as a standalone feed. That is a decision rather than a limitation — if there is enough industry interest, opening it up is something we would look at.

Why does storage per page matter so much?

It is the difference between an index that can afford to grow and one that cannot. Costs scale with every page held, so a few kilobytes either way decides whether reading the whole web is a business or a bonfire.

Does more machinery mean slower answers?

No — the opposite. Reading is spread out and done ahead of time, so by the time you ask a question the work is already finished.

How the index works — askFinz