Reads broken queries · Runs on your servers · One-time payment

Search you buy once —
and that actually finds things.

Your customers type badly — typos, half a part number, two words glued into one. Dartfind finds it anyway. One payment, your own server, no subscription.

Built for distributors, wholesalers and e-commerce with large catalogs — from 10,000 to 100 million records.

LIVE INDEX — ms
>

Broken query in. Right parts out.

This is the whole product in one screen: a part number typed from memory — wrong letter, no dashes — and the catalog answers anyway.

dartfind.local/search
skf 62o5 2rs
6 resultsin 19 ms
matched with typo: 62o5 → 6205
SKF 6205-2RS/C3
Deep groove ball bearing 25×52×15SKF
In stock · 142 pcs
$8.40
SKF 6205-2RSH
Deep groove ball bearing, sealedSKF
In stock · 96 pcs
$9.10
FAG 6205-2RSR
Deep groove ball bearing 25×52×15FAG
3–5 days
$8.95
NSK 6205DDU
Deep groove ball bearing, sealedNSK
In stock · 58 pcs
$8.75
The letter “o” instead of a zero, no dashes, lowercase — and every result is the right part. Demo catalog shown; the same engine runs your data in the pilot.

One search box. Three things the expensive ones still get wrong.

Pay once

Yours for good

You get the software itself — a compiled program installed on your server, like buying a machine for your floor. One payment covers everything: no renewals, no metering, no support contract. Search is the only part of your stack companies still rent forever. This one you own.

Less infrastructure

An x86 server. That’s the whole stack.

One compiled binary on an ordinary server you already own — nobody on payroll to keep search alive. ~20 ms per query in production today; built for catalogs up to 100 million records on a single server.

Reads broken input

The query that returns zero results elsewhere

“oill fitler corola”, half a part number, two words glued into one — found without synonym lists, dictionaries or months of tuning. We tested nearly thirty distributors: a handful return literally nothing for one wrong letter, and others quietly collapse from 121 results to 2. Broken input is the query a real catalog gets all day. It's the one Dartfind is built for.

We tested search at nearly thirty distributors. Here is what breaks.

In July 2026 we ran the same queries against the on-site search of nearly thirty North American electrical and industrial distributors. Four failure patterns stood out.

Buying a search vendor did not predict good search — some of the strongest catalogs we tested ran no commercial search product at all. Reproduce any of these on your own storefront in thirty seconds. That is what your customer sees before they close the tab. Read the full audit →

Don’t believe it? Free pilot.

On your catalog — or a fake one. In your infrastructure. Live in 2 days. If you like it, you pay. If not, it cost you nothing.

Your dataa catalog export, real or fake — no access to your systems needed
Your infrastructureruns on your server; the data never leaves your network
2 daysfrom export to a working search box you can torture
$0no contract, no card, no strings attached
Start the free pilot Type your customers’ worst queries at it. Then decide.

Side by side.

vs
Dartfind Elasticsearch Algolia Coveo HawkSearch
Pricing One-time payment Subscription + infra, every year Per query, forever, grows with traffic Enterprise SaaS, renewed yearly Annual subscription, renegotiated yearly
Infrastructure An ordinary x86 server Multi-node cluster + engineer Their cloud only Coveo cloud only Their cloud
Hardware needed Any x86 CPU made this millennium Sized cluster, RAM-hungry JVM None of yours — and none of your control None of yours None of yours
Dirty queries Found — no configuration Synonym lists, analyzers, tuning Good, within plan limits Strong, tuned per deployment Rules and synonyms, maintained
Spec queries “20A”, “5W-30”, “3/4 inch” — the spec in the query ranks the results Whatever the analyzers were built to do Token match; spec ranking is rules work ML ranking, tuned per deployment Handled through merchandising rules
Data location Stays on your servers Yours, if self-hosted Leaves your network Leaves your network Leaves your network
Time to live search 2 days Weeks to months Weeks Weeks to months Weeks

Deep dives: Algolia alternative · Elasticsearch alternative · Coveo alternative · HawkSearch alternative · Searchspring alternative · Klevu alternative

Built with rocket and space engineering languages.

Most search products are written in languages made for convenience — Python, Java, JavaScript. Convenient to write, slow to run. Dartfind is written in the languages used where software is not allowed to fail or lag: rockets, avionics, telecom.

RUST

Speed without crashes

The modern systems language used by aerospace, Cloudflare and the Linux kernel. As fast as raw C, but memory-safe — memory leaks and the crashes they cause are impossible by construction, not by testing.

ADA SPARK

The rocket language, mathematically proven

Ada flies on rockets, avionics and air traffic control. SPARK is its strictest form: the code is checked by mathematical proof before it ever runs. Used where a single error costs a vehicle — not a support ticket.

ERLANG + ISPC

Decades without reboot

Erlang runs telecom switches with uptime measured in years — it supervises and heals the system while it works. ISPC squeezes every core of the CPU for the number-crunching itself.

No Python, no interpreters, nothing slow anywhere in the path of your query — and nothing that needs babysitting. Once configured, there is nothing left to maintain: no memory to leak, no cluster to nurse, no search engineer to hire. That’s how ~20 ms stays ~20 ms, year after year.

Field report: auto parts.

CASE — AUTO PARTS RETAIL & TRADE COUNTER STATUS: LIVE IN PRODUCTION

The hardest search environment there is: part numbers.

An auto parts business runs Dartfind as its live counter and catalog search. Customers and staff type what auto parts people type all day: half-remembered part numbers, brand names with typos, two words glued together, a code read off a worn, oily label.

Every query hits the full catalog and comes back in the time it takes to draw the next character. No synonym lists were built, no relevance rules written — the engine reads broken input as-is. It runs on an ordinary machine, index in RAM, alongside the rest of the store’s systems, and keeps working with the internet down.

16,787
products live in the catalog
28,810
unique keys in the index
~20 ms
per query, full catalog
1
ordinary machine, no cluster

Every failed search is an order that went to a competitor.

A customer who types “oill fitler corola” and gets “no results” does not retype. They leave. Your sales reps lose the same way: half a part number from a client, twenty minutes of digging, deal cooling down.

Matching

Built for broken input

Typos, partial article numbers, glued words and wrong layouts are handled by the search itself. Nothing to configure, nothing to babysit.

Footprint

An ordinary x86 server

The index lives in RAM and answers from a single server — any x86 machine with SSE2, which means practically anything made this millennium. 16,787 products run in a live store on this engine today.

Control

Your hardware, your data

Works with the internet cut off entirely. No third-party cloud, no per-query metering, no way for anyone outside to switch it off.

Pay once. Own it.

Search is the only part of your stack you rent forever. Dartfind replaces that line with one payment.

Dartfind license
One payment. Yours for good.

Installed on your hardware, tuned to your catalog, handed over with ranking controls your team can adjust in minutes. No renewals, no meters, no lock-in — and nothing that phones home. Need extra filters, sorting or custom behavior later? We build it as paid work, and it ships to every customer after you.

What search costs today, every year
  • Elastic cluster hosting recurring
  • Elastic subscription tier recurring
  • Search engineer on payroll recurring
  • Algolia per-query fees grows with traffic
  • Lost orders on “no results” uncounted

How the pilot works.

STEP 1

Send a catalog export

CSV or any dump — names, SKUs, attributes. A fake dataset works too.

STEP 2

Get a live search in 2 days

Running in your infrastructure. Type the worst queries your customers send and watch them land.

STEP 3

Like it? Then you pay.

One payment, full deployment, wired to your storefront or CRM. Don’t like it — walk away.

Got it. We’ll get back to you within a day to arrange the pilot.
Something went wrong — please try again.

Questions people actually ask.

What exactly do I get for the one-time payment?+
A compiled binary, configured for your product catalog and installed on your server, plus a sync script that keeps the index fresh and a short manual for your admin. Software used to work this way: you bought a program, installed it, and it ran for years. We do the same. We have no floor of support staff to feed with monthly bills — we prepare the code for your case, hand it over, and it’s yours for good.
How much does it cost?+
The price depends on catalog size and integrations, so we quote per project. The honest frame: full deployment — binary, setup, integration — comes in below what companies typically spend on their search stack in a single year. One payment, less than one year of an Elasticsearch cluster with an engineer, or one year of an Algolia subscription at real volume. From year two onward you’re simply ahead.
Why is Dartfind faster than Elasticsearch and Algolia on a fraction of the hardware?+
Elasticsearch runs on the JVM, which eats gigabytes of memory before doing any useful work — that’s why a typical production deployment is a 3–9 node cluster with an engineer keeping it alive. Dartfind is written in Rust, Ada SPARK, ISPC and Erlang: compiled languages from rockets, avionics and telecom, with no interpreter and no garbage collector anywhere in the query path. Matching runs on compact binary signatures compared through native bitwise CPU instructions, which a processor executes billions of times per second. The result: about 20 ms per query in production, on an ordinary x86 server — and an engine built to hold up to 100 million records on that single machine, where the leaders need a cluster. We’re preparing formal side-by-side benchmarks and would rather publish measured numbers than round ones.
Why does it run on any x86 processor?+
The engine needs nothing newer than SSE2 — an instruction set every x86 CPU has shipped with since the early 2000s. No AVX-512, no GPU, no special hardware. Whatever server you already own almost certainly qualifies.
Does it need an internet connection?+
No. Dartfind runs entirely inside your network and keeps working with the cable physically unplugged. There is no license server, no activation, no update checks, no per-query metering — nothing in the code that calls home. Your catalog never leaves your infrastructure, which also settles most data-residency and security-policy questions before they’re asked.
How can we be sure there’s nothing malicious inside the binary?+
Three ways, from simple to formal. First, your admin can verify it in ten minutes: run the engine and watch its network activity — zero outbound connections, ever. Since the engine needs no internet, you can firewall it off completely, and then whatever anyone imagines inside physically has no way to send a single byte out. Second, the contract carries an explicit warranty: no hidden functions, no data transmission, no remote kill switch, with liability attached. And for large deployments there are additional options we discuss individually. In practice the firewall test convinces most security teams on day one.
How does typo tolerance work? Do I need synonym dictionaries?+
For broken input — no dictionaries at all. Misspelled brand names, half-remembered part numbers, glued words, wrong keyboard layouts: the matching core reads them as they are, because tolerance to noise is a property of the math itself, not a rule you configure. This is the exact place we watched the paid platforms fail: in our July 2026 field test of nearly thirty distributors, a handful returned zero results for one misspelled letter, and others collapsed from 121 results to 2 with no correction offered. In Elasticsearch or a rules-based platform, every one of those cases is a synonym or redirect somebody has to write and maintain. Here they simply land.
Will I have to fire my search team?+
No. Dartfind takes over the mechanical grind — typos, partial codes, broken queries. Your business knowledge still needs people. If in your trade “oil” also means part group 15-86, no engine on earth can know that: someone who knows your catalog sets up those mappings, tunes ranking, decides what “relevant” means for your customers. Your search people get to spend their time on exactly that instead of nursing a cluster.
What if something breaks? There’s no support contract.+
After setup there is very little left to break: Rust makes memory leaks impossible by construction, Ada SPARK code is checked by mathematical proof before it ever runs, Erlang watches the system and restarts components on its own. Deployment includes a stabilization period where we fix anything that surfaces during integration, and critical defects we fix regardless — we want this running as much as you do. We just refuse to bill you monthly for the possibility of a problem.
Can I get custom features later?+
Yes. Extra filters, sorting logic, custom behavior — we build them as paid work. One thing we say upfront: features built this way ship to every customer after you, the same way earlier customers’ projects built what you’re getting. The product keeps improving and nobody pays rent for it.
How does it integrate with my store or ERP?+
Two connections. A sync script feeds your catalog into the engine — from a database or a plain export — and keeps it updated on schedule. A query endpoint answers searches; your storefront or counter software calls it instead of the current one. For most e-commerce platforms and ERP systems this is days of work. The free pilot skips integration entirely: you send an export, we hand you a working search.
How big a catalog can it handle?+
100 million records on a single x86 server, index in RAM. If your catalog is bigger, write to us — that’s the kind of problem we enjoy.