Add files, whole folders, photos, or pasted text.
Import from Files or your photo library, or share into SearchThing from another app. Everything lands in one searchable library.
For iPhone, iPad & Android
Search your files by meaning.
You remember the idea, not the exact words. SearchThing’s on-device AI finds the passage anyway.
Not released yet. Choose a platform above for a launch notification.
- On-device AI
- $9.99 once on the App Store
- No account
Supported formats include PDF, Word, PowerPoint, Excel, EPUB, Markdown, Plain text, RTF, CSV, JSON, Web pages, Source code, and photos or scans.
AI search by meaning
It finds what you meant, not just what you typed.
Keyword search fails the moment your words differ from the document’s. SearchThing compares meaning, so a search for field glasses finds your note about binoculars.
Search for antifreeze puddle
Car maintenance log#1 resultNoticed a sweet smell after driving and a green drip on the garage floor under the passenger side of the engine bay.
Search for lost baggage
Portugal trip journal#1 result…landed in Lisbon at 9:40 am after maybe two hours of sleep. My suitcase never came off the carousel.
Search for vacation allowance
Onboarding handbook#1 resultPTO: accrues at 1.67 days per month (20 days a year). You can use it once you’ve been here 30 days.
Search for field glasses
Birding log#1 resultBinoculars: 8x42 roof prism, finally worth the money. Counts go into eBird; notes live here.
Real queries from SearchThing’s test set. None shares a word with the passage it finds, and keyword-only search misses every one.
Under the hood
Two searches at once. One ranked answer.
Smart and Meaning searches run through a keyword index and an AI meaning index at once, then fuse the results into one ranking. All of it happens on your device.
- 01
Read
Every document is split into passages along its own structure. Short paragraphs join their section neighbors, never across a heading.
- 02
Understand
EmbeddingGemma turns each passage into a 512-number vector that captures what it means, right on your device.
- 03
Search two ways
In Smart and Meaning, your query runs through a BM25 keyword index and an HNSW meaning index at the same time.
- 04
Fuse and rank
Reciprocal rank fusion blends both lists, then exact cosine similarity rescores the finalists against your query.
- 05
Show its work
Each result can show why it matched: the words it shares with your search and how close it is in meaning.
Three search modes
Choose how literal to be.
Switch modes without losing your place. Smart is the default.
- Smart
Keywords and AI, together.
Finds your words first, then adds passages the AI judges related. They’re labeled, so you always know which is which.
- Meaning
Pure meaning.
Describe the idea in your own words. The AI finds passages about it even when they share none of your words.
- Exact
Just the phrase you remember.
Stays with the words you entered. Built for names, numbers, and quotes.
Follow the evidence
Every answer comes with its source.
Open the original PDF or readable text, see why a passage matched, then pin it or add a note.
- 01
Find the line you need.
Ranked passages from every file, not a list of file names.
- 02
Open the original.
Jump from a result into its source, then step through the other matches.
- 03
Know why it matched.
See the words and meaning behind every result’s rank.
- 04
Everything in one place.
Projects keep research, work, and reference apart.
Screens from the iPhone app, shown with sample documents.
Modern AI, on device
Built on the latest retrieval research.
A 2025 AI model and proven search techniques, tuned to work together on your device.
- Google DeepMind, 2025
EmbeddingGemma
A 300M-parameter model built for retrieval. At launch it ranked first among open embedding models under 500M parameters on the MTEB benchmark. Quantization-aware training keeps it accurate at 8-bit, small enough for a phone.
- Kusupati et al., 2022
Matryoshka embeddings
Meaning vectors are trimmed from 768 to 512 dimensions. The index is a third smaller, and in our tests it put the right source first more often.
- Asymmetric search
Retrieval prompts
Queries and passages are encoded with EmbeddingGemma’s own search prompts, so a short question lines up with the paragraph that answers it.
- BM25 + dense vectors
Hybrid retrieval
Keyword precision and AI recall in one ranked list. Exact names, codes, and numbers still come first when you type them.
- Cormack et al., 2009
Reciprocal rank fusion
Merges the keyword and meaning rankings by position, so neither score scale drowns out the other.
- Malkov & Yashunin, 2016
HNSW vector index
A navigable graph finds the nearest passages fast. Exact cosine similarity then rescores the finalists.
Measured, not guessed
Tested on 181 questions with known answers.
- 85%
- of synonym searches put the right document first in Meaning mode. Keyword-only: 5%.
- 97%
- of test questions found the right document in the top 10. Keyword-only: 52%.
- 0%
- of test questions came back empty. Keyword-only: 22%.
- 95%
- of benchmark questions got the best source first from EmbeddingGemma alone. iOS’s built-in embeddings: 29%.
From SearchThing’s evaluation harness, which replays the app’s real search pipeline with EmbeddingGemma. Test set: 119 questions over 40 documents written for testing, mostly synonyms, paraphrases, and concepts (September 2026). Benchmark: 62 questions over 37 mixed-format documents (August 2026). Keyword-only means Exact mode.
Built for how you collect
Bring anything in. Keep what matters.
Photos become searchable.
Text in your photos and scanned PDFs is recognized on your device, so a picture of a handout is as findable as a typed document.
Pin passages. Add notes.
Keep the lines worth returning to, with your own notes beside them.
Save the searches you repeat.
Come back to a question with one tap.
A widget for your Home Screen.
See what’s ready and jump straight into search.
On iPad
Room to spread out.
A sidebar keeps Library, Search, Saved, and Settings one tap away, with space for your whole document list.
Private by design
Powerful AI that never leaves your device.
Many AI search tools upload your files to a server. SearchThing brings the model to your files instead. Documents, the index, and your history stay on your phone or tablet, and EmbeddingGemma is downloaded once after you accept its terms.
Read the privacy details Read why private search starts with a smaller trust boundary- 0
- accounts to create
- 0
- searches sent to a server
- 0
- ads or subscriptions
- ≈329 MB
- one-time model download. Then search runs on your device.
How does the AI search work?
SearchThing runs Google DeepMind’s EmbeddingGemma model on your device. It turns each passage and each search into a vector that captures meaning, then blends those matches with a classic keyword index so exact words still count.
Are my documents used to train AI?
No. The model runs entirely on your phone or tablet. Your documents, index, and searches are never uploaded, so there is nothing to train on.
When can I get it?
SearchThing launches soon on the App Store for iPhone and iPad and on Google Play for Android. Tap either store button and leave your email, and we’ll send one note on launch day.
Does it need an internet connection?
Only once. The first launch downloads the EmbeddingGemma model (about 329 MB, Wi-Fi recommended) after you accept its terms. Importing and searching then happen on your device.
What does it cost?
$9.99 once on the App Store. No subscription and no account.
Which devices are supported?
iPhone with iOS 17 or later and iPad with iPadOS 17 or later. The Android version is on its way.
Can I try it now?
Yes. SearchThing for the browser searches your PDFs, Markdown, and text files locally in your browser, with nothing to install.
Find the passage that matters.
For iPhone, iPad, and Android.
Not released yet. Choose a platform above for a launch notification.