Semantic File Search
Search your Mac using your memory.
Traditional search engines require you to remember the exact filename. Dhito uses on-device vector embeddings to search your files by meaning, concept, and context.
Why keyword search misses files you know you have
Spotlight builds a lexical index: it maps the literal characters in your query to the literal characters in your filenames and file contents. If there is no character match, the file does not come back — however obvious the connection is to you.
That single design decision explains most of the frustration. Search "client contract signature terms" and nothing appears, because the file is called Agreement_Acme_Corp_2026.pdf and a keyword index has no reason to connect "contract" with "agreement". A photo saved as DSC_0982.jpg is invisible unless you tagged it by hand. A two-hour recorded briefing is matched on its filename alone, while everything of value sits inside the audio.
Semantic search removes the requirement that you guess the words the file uses. You describe what the file is about, and the match happens on meaning.
How semantic search works on your Mac
Dhito runs a local embedding model (BGE-small-en-v1.5) over your files as it indexes them. The model converts each document into a vector — a list of numbers that positions the text by meaning rather than by spelling. Related ideas end up close together in that space, so "contract" and "agreement" land in nearly the same place even though they share no letters.
When you search, your query goes through the same model and becomes a vector too. Dhito then compares it against the index and ranks files by distance. Nothing is matched character by character, which is why a query can find a document that contains none of the words you typed.
The whole pipeline runs on your Apple Silicon. The model, the index, and your files all stay on the machine — there is no account to create for search to work, and no query leaves the device.
Natural Language Queries
Stop guessing if the file was named 'Q3_Report_Final' or 'Project_Update'. Just type 'that presentation about third quarter revenue' and Dhito will find it instantly.
Conceptual Understanding
Dhito understands synonyms and related concepts. Searching for 'automobile maintenance' will surface a PDF called 'Car_Repairs.pdf', something traditional keyword search can never do.
AI Vision for Images
If you have a screenshot of a funny tweet or a photo of a whiteboard, Dhito's local vision models can 'see' the contents. Search for 'photo of the glass conference room' and find your unlabelled images.
Common questions
How is semantic search different from Spotlight?
Spotlight matches the literal characters in your query against filenames and file contents, so it fails when your words differ from the file's words. Semantic search compares meaning instead, using a local AI model, so a query like "policy on remote work" can surface a file called Employee Handbook 2026.pdf.
Does semantic search replace Spotlight entirely?
It covers more of it than it used to. Dhito is built for the job Spotlight was never designed for — finding files when you remember what something was about but not what it was called — and it launches apps as well, matching them by what they are for as well as by name, so “messaging app” finds WhatsApp, Telegram and Messages. It also answers calculations, conversions and time zones inline. What it does not reach is Mail, Calendar, Contacts, System Settings panes and web suggestions, so many people keep Cmd + Space for those and give Dhito a second shortcut.
Do my files or search queries get uploaded anywhere?
No. The embedding model, the index, and the search all run locally on your Mac. Files and queries never leave the machine, and Dhito works with no network connection at all.
What file types can Dhito search by meaning?
Over 100 extensions in total. Documents and text by their contents — PDF, Word, RTF, ODT, EPUB, plain text, Markdown, CSV and around 33 code formats; images by their visual contents using a local vision model — JPEG, PNG, HEIC, HEIF, WebP, PSD, TIFF, BMP, ICO, AVIF and ten camera RAW formats; and 17 video and 14 audio containers by their transcripts. All of it is indexed into the same semantic space, so one query searches across every kind of file at once. The gap worth checking first: Excel, PowerPoint, Pages, Keynote and Numbers are not indexed.
How long does the first index take?
It depends on how much you point it at and which Apple Silicon chip you have. Indexing runs in the background and is incremental after the first pass, so new and changed files are picked up as you work rather than requiring a full rebuild.