Vicinity 1
Find whatyou meant.
Vicinity finds your to-dos, notes, events, contacts and conversations by what they mean, not only the words in them. Searches run on your device, and work offline.
Embeddings
A to-do becomes 384 numbers.
Vicinity reads the text and writes down 384 numbers that stand for what it means. Each bar is one of them, worked out in your browser. Type anything over the to-do and watch the numbers change.
Loading Vicinity 1, 34 MB…
Meaning space
Together, they’re one point.
The numbers are scaled to length 1, so they’re a direction: one point in a space of 384 dimensions. Things that mean the same thing point the same way, so they land together. The flowers gather with the scarf Mum loved and the telescope Eli wants, though none of them says “present”.
Drawn here in three of the 384, for a small made-up account.
Search
Search finds the nearest points.
Your search becomes a point too, on your device. The things closest to it come first.
Nothing close in this small account.
Training
Made on one Mac, in about 53 hours.
Vicinity started from gte-small,1 an open model under the MIT licence, and was trained further for what people keep in Milieu. Everything ran on one M3 Max, from the first row written to the finished model.
- Keywords
- estimate felling customer
- Another way
- send that family a price for removing their big tree
- Question
- did i ever send a prise for that tree job?
Rows written by an open model, Qwen3.6,3 plus about 100,000 public question–answer and title–abstract pairs. Trained without any of your data.
Results
Tested on data it never saw.
307 made-up people and their 6,670 rows were kept out of training. Before training, 96 searches were written by hand. The bars show how often each model finds what you meant in its top five.
Milieu blends in a little of the words, too. Meaning finds what you were thinking of, and the words find exact things, like invoice INV-2041.
| Keywords alone | 0.550 |
| Vicinity alone | 0.663 |
| Blended, 90% meaning and 10% keywords | 0.674 |
On science benchmarks it’s a little worse than gte-small (SciFact 0.696 against 0.727, NFCorpus 0.339 against 0.348). That’s the price of being a specialist.
Size
Small enough to keep everything.
Each thing is kept as 384 numbers of one byte each, and that costs nothing in quality. The first 256 or 128 numbers work on their own too, and so does a single bit of each, at a price.
What Milieu keeps: vicinity-1/384/int8