Skip to content
cups

Tools · Local

Find the passage, the picture or the video in your own folder.

Choose a folder, then type a question or describe a picture or video in your own words. Local shows the passages most likely to hold the answer, each with its file and lines, and the pictures and videos closest to what you described. It writes no answer of its own, and your files stay on your computer.

Why this exists

Ordinary search finds the words you type. Often I remember what a note was about but not the words in it. Search that matches by meaning usually means uploading the folder to a service. Local does that matching on my own computer, so I can point it at private notes.

What goes in and what comes out

Goes in
A folder of your own files. Text files, such as notes, Markdown, code and CSV. PDFs and Word documents, read as text. Pictures: PNG, JPEG, GIF, WebP, TIFF, BMP and iPhone HEIC files. Videos: MP4, MOV, M4V, WebM, MKV and AVI files.
First
Local makes an index. It reads each file once and keeps what it finds in a hidden folder inside your folder. Your files are not changed. When files change, you update the index, which takes seconds for a few files.
You type
A question, or what a picture or video shows.
Comes out
Three lists of up to five. “Documents” has passages: those that use the words you typed, and those a model judges close in meaning, even in other words. A passage from a PDF says which page it is on. “Pictures” and “Videos” have the closest of each, with a video shown as one still. A result the model scores no higher than it scores an unrelated question is folded away, one press from view, and a list with nothing close says so.
You can
See the words you typed marked in each passage, open a picture at full size, and show any found file in Finder.

A picture or video comes first when every word you typed is a whole word of its file or folder name. The rest are ranked by what they show. A video is looked at as up to eight stills spread across it, so a short scene in a long video can be missed. Local does not describe a picture, read the words written in one, or listen to the sound of a video.

Nothing is kept except the index. The folder’s path, your questions and the results are not saved or logged.

Sort, which files a pile of photos into the folders I already have, reads the pictures in Local’s index where there is one, and compares them more closely. The route is written up as the playbook A pile of photos.

Models used

EmbeddingGemma 2
By Google. Its one job is to turn your question, each passage, each picture and stills from each video into numbers, so that passages, pictures and videos close in meaning can be found. It writes nothing. About 1.5 GB, Apache-2.0 licence, revision 914f7f89142e33e77833254d9c9b90c3cef7303b. Every file’s size and hash is checked before the model loads.
No model
Matching the words you typed, reading the text of a PDF or Word document, finding a picture or video by its name, and marking your words in a passage are ordinary code.

The model runs on your computer’s CPU. Setup downloads it once; after that, normal use downloads nothing and sends nothing. An iPhone photo is read from a JPEG copy the Mac makes for the moment and does not keep.

What has been checked

On this project’s own files, 593 of them, twelve questions were written down before any search, each with the passage that answers it. Eleven of the twelve were written by an AI agent that had read the files. Matching by words alone had the right passage in its five for 6 of the 12. Matching by words and meaning, which is what Local shows, had it for 9. Three were found by neither.

On 601 of my own notes I asked a few questions of my own, and matching by words and meaning worked much better for me than words alone. I kept no count.

For pictures, I used one folder of 99 varied pictures and 13 descriptions, written with the pictures in view. The right picture was in the five shown for all 13, and first for 12. Pictures that look alike, such as a run of screenshots of one app, have not been tried.

For videos, I made an index of one folder of my own holding 26 videos, and finding a video by what it shows worked for me. I kept no count. One more video in that folder could not be read.

For PDFs, I used one folder of my own holding 25, from 1 to 42 pages, and 23 questions an AI agent wrote from the file names. The right PDF came first for 22, and four unrelated questions showed nothing. For one of the 22 the right PDF was folded away as not close. No scanned PDF and no page set in columns was among them. Word documents have only been tried on one made up for the test.

These are counts from one person’s folders on one computer, not a measure of how often Local finds the right thing.

Known limits

Availability

Not available to download yet. Local runs in the development Kettle app on my Mac, and as its own page. There is no public installer or source download at present.