Microsoft Outlook Email
By sender, topic and date, in your own words.
ARCHIVE CITY · 2:14 A.M.
It was raining. A file had gone missing, and nobody could remember its name.
PRESS ANY KEY TO ENTER ARCHIVE CITYTAP TO ENTER ARCHIVE CITY
FINDER OF LOST EMAILS, FILES AND SETTINGS ON macOS
By sender, topic and date, in your own words.
Inside PDFs, Word and PowerPoint.
Open either from a few words.
New modelTofazzol v2.1
Shlexball 1.16 runs Tofazzol v2.1, a short further training of Tofazzol v2. Before that training we set aside 100 requests about a real work mailbox and scored them once, at the end. Tofazzol v2.1 returned exactly the right emails for 80 of them. Tofazzol v2 managed 66.
Email search on a real mailbox
100 held-back requests about one real work mailbox (n=100)
File search
Findbench 3.1 file searches (n=746)
Synthetic email search
Findbench 3.1 email searches on synthetic mailboxes (n=240)
Every test, side by side
Share of requests answered exactly
| Test | Tofazzol v2.1Shlexball 1.16 · on your Mac | Tofazzol v2Shlexball 1.13 to 1.15 | Gemma 4 31BOur workstation · 30 examples | Gemini 3.8 FlashCloud · 30 examples | Gemini 3.5 Flash-LiteCloud · 30 examples |
|---|---|---|---|---|---|
| Email on a real mailboxHeld-back requests · exactly the right emails | 80% | 66% | Not run | Not run | Not run |
| Email on a real mailboxHeld-back requests · wanted email among the results | 90% | 77% | Not run | Not run | Not run |
| Email on a real mailboxDevelopment requests · exactly the right emails | 83% | 72% | Not run | Not run | Not run |
| File searchFindbench 3.1 · 746 searches | 98.5%735 of 746 | 98.5%735 of 746 | 98.3%733 of 746 | 98.0%731 of 746 | 93.4%697 of 746 |
| Documents320 searches | 99.1%317 of 320 | 99.4%318 of 320 | 100%320 of 320 | 99.7%319 of 320 | 96.9%310 of 320 |
| File names160 searches | 97.5%156 of 160 | 96.9%155 of 160 | 98.8%158 of 160 | 98.8%158 of 160 | 88.1%141 of 160 |
| Photos, video and audio266 searches | 98.5%262 of 266 | 98.5%262 of 266 | 95.9%255 of 266 | 95.5%254 of 266 | 92.5%246 of 266 |
| Email on synthetic mailboxesFindbench 3.1 · 240 searches | 90.4%217 of 240 | 90.4%217 of 240 | 100%240 of 240 | 100%240 of 240 | 97.1%233 of 240 |
| All of Findbench 3.1986 searches | 96.6%952 of 986 | 96.6%952 of 986 | 98.7%973 of 986 | 98.5%971 of 986 | 94.3%930 of 986 |
The 100 held-back requests were generated from one mailbox, not typed by people, so expect a smaller gain on your own mail. We chose Tofazzol v2.1 with other requests from the same mailbox and with Findbench 3.1, so those are development results; only the held-back 100 were scored once.
A short detective game. Ten files with names nobody remembers, three suspects, one stolen pearl. Describe what you are looking for, read what turns up, and name the culprit.
CASE CLOSED
Ten files with names nobody remembered, and you found them by describing what was in them. That is how Shlexball works on a Mac: it searches file names and the text inside them, even inside PDFs and Word documents.
30 free searches, no account. macOS 15 or later. Apple silicon. 16 GB memory.
Take the case home: the ten case files as real PDF, Word, PowerPoint and text documents (zip, 100 KB), to search on your own Mac.
A game: these searches run in your browser over ten fictional files, with simple word matching. Shlexball on your Mac searches the folder you choose, with an AI model that runs on the Mac.
New paperWorking paper
Can a 1.5B model keep up with far larger ones when the job is narrow enough? On 746 file searches replayed through the same search engine, Tofazzol, the model behind Shlexball, passes 735. Gemma 4 31B passes 733 and Gemini 3.8 Flash passes 731, and Tofazzol has about one-twentieth of Gemma's parameters.
Email goes the other way: Gemma and Flash pass all 240 email tests and Tofazzol passes 217. The paper explains how specialization closes the gap on files, and where it stops.
This compares a fine-tuned specialist with general models given 30 worked examples per request, on a benchmark we also used during development. The file differences are small and are not a formal equivalence result.
Spotlight matches names and indexed words, so a file you can describe but not name often slips past it. Shlexball works differently: type what you remember, like “the lease PDF from last spring”, and it searches file names and the text inside PDFs and Word documents to find it. The AI model runs entirely on your Mac, so your files are never uploaded. Start with 30 free searches, no account required.
Search inside PDFs on a Mac Find files by meaning, offline EasyFind vs Shlexball
Find a file when you forgot its name Spotlight not finding files on your Mac
Shlexball's AI runs entirely on your Mac, finding lost files like your GitHub and Stripe recovery codes in seconds — and your files never leave your Mac.
The Shlexball launch film. The four searches in it are real, recorded on an Apple silicon Mac with the model running on the Mac.
YOU> the lease PDF from last spring
SHLEXBALL> Found it. 3 files.
Type what you remember. Shlexball searches file names, dates, and the text inside PDFs and documents, in the one folder you choose. The model runs on your Mac. Shlexball only reads: it never changes, moves or deletes a file.