FileDig

Find your files by what they contain, not what they are called.

Local semantic search for documents and pictures on Windows. Nothing leaves your computer.

Download · Features · Getting started · Build from source

About

Type what you are looking for in plain language, such as "how much is the rent" or "photo of a sunset", and FileDig finds the matching text files, PDFs and images by meaning.

It runs EmbeddingGemma 2 on your own machine through llama.cpp, using an NVIDIA GPU when one is available and the CPU otherwise. Inspired by DigUp for macOS.

Features

• 🔎 Search by meaning across text and images in one index
• 🎯 Exact-match search for names, codes and numbers, merged with the semantic results
• 📄 PDF results open at the page that matched
• 🖼️ Image search by what a picture shows, with a thumbnail grid
• ⚡ GPU acceleration on NVIDIA cards, with automatic CPU fallback
• 🔄 Incremental indexing that only processes new or changed files
• ⏸️ Stop and resume indexing at any time, with a warning while a folder is incomplete
• ⌨️ Global hotkey (Ctrl+Shift+Space) and a tray icon
• 🔒 Private and read-only: files are never changed, moved or uploaded
• 📦 Portable: runs from a folder with no installation

Supported files

Getting started

Requirements

• 64-bit Windows 10 or 11 (the WebView2 runtime is part of Windows 11 and up-to-date Windows 10)
• About 1 GB of free disk space for the model
• Optional: an NVIDIA graphics card with a recent driver and about 4 GB of free VRAM

Install

1. Download FileDig-lite.zip from the latest release.
2. Unzip it anywhere and run FileDig.exe.
3. Click Download model on the start screen. This is a one-time download of about 825 MB.
4. Click Folders, then Add folder, or drag a folder into the window.

Indexing starts straight away, and you can search while it runs.

Turn on GPU acceleration

On a PC with an NVIDIA card, open Settings (the gear next to Folders) and choose Download GPU support (about 626 MB). Indexing becomes several times faster, and the existing index keeps working.

Only an up-to-date NVIDIA driver is needed. The CUDA toolkit is not.

Usage

• Tabs switch between All, Documents and Images.
• Exact match marks results that contain every word you typed.
• Rescan, in the Folders panel, picks up files added, changed or deleted since the last pass. A rescan also runs each time the app starts.
• Closing the window keeps FileDig in the tray. To quit, right-click the tray icon and choose Quit.

Privacy

Searching and indexing never leave your computer. The only network requests FileDig makes are the two optional downloads you start yourself: the model from Hugging Face and GPU support from the llama.cpp releases on GitHub. Both are pinned to exact versions and verified by SHA-256 before use.

There is no account, no telemetry and no update check.

Performance

Measured on an RTX 3060 12 GB with a Core i7-12700:

The model unloads after 5 minutes without use to free VRAM, and reloads in a few seconds on the next search.

runtime\
  llama-cuda\   llama-server.exe + all DLLs from the CUDA zip and the cudart zip
  llama-cpu\    llama-server.exe + all DLLs from the CPU zip
  models\
    embeddinggemma-2-Q8_0.gguf
    mmproj-embeddinggemma-2-Q8_0.gguf

Build from source

You need Rust, Node.js and pnpm.

git clone <this repository>
cd filedig
pnpm install
pnpm tauri dev

The app offers to download the model on first run. To use files you already have, place a runtime folder in the project root, laid out as described in Installing the model and GPU support by hand above.

Portable packages

pwsh scripts\package-portable.ps1 -Lite -Zip   # about 25 MB; model downloaded on first run
pwsh scripts\package-portable.ps1 -CpuOnly     # model included, CPU backend
pwsh scripts\package-portable.ps1              # model, CPU and GPU backends (about 2 GB)

Packages are written to dist-portable\. A copy is portable whenever a data folder sits beside the executable: the index, logs and downloads then go there instead of %APPDATA%.

Configuration

How it works

flowchart LR
    Files[("Your folders<br/>documents and pictures")]

    subgraph App["FileDig"]
        direction TB
        UI["UI<br/>TypeScript in a WebView"]
        subgraph Core["Rust backend"]
            direction TB
            Search["Search<br/>ranks and merges results"]
            Indexer["Indexer<br/>reads files on several CPU cores"]
            Engine["Engine<br/>starts and talks to the model"]
            Setup["Downloads<br/>model and GPU support"]
        end
        Llama["llama-server<br/>EmbeddingGemma 2 on GPU or CPU"]
        DB[("SQLite<br/>vectors, text, keyword index, thumbnails")]
    end

    Net(["Hugging Face and GitHub<br/>optional one-time downloads"])

    UI -- "query" --> Search
    Search -- "results" --> UI
    Files -- "read only" --> Indexer
    Indexer -- "passages and images" --> Engine
    Search -- "query text" --> Engine
    Engine <-- "embeddings over local HTTP" --> Llama
    Indexer -- "store" --> DB
    Search -- "look up" --> DB
    Net -.-> Setup
    Setup -. "installs" .-> Llama

• Indexer (src-tauri/src/indexer.rs): walks the folders, splits text into overlapping passages of about 1,200 characters, extracts PDF text page by page, and downsizes images before embedding. Files are read on several CPU cores while the model works on the GPU.
• Engine (src-tauri/src/engine.rs): runs llama-server as a hidden helper on a free local port. Text and images are embedded into the same 768-dimensional space, so a text query can match either.
• Storage (src-tauri/src/db.rs): one SQLite file holds the files, passages, vectors, thumbnails and an FTS5 keyword index. Vectors are compared by brute force in memory.
• Search (src-tauri/src/lib.rs): the semantic and keyword rankings are merged with reciprocal rank fusion, and the best passage per file is shown.
• Downloads (src-tauri/src/setup.rs): fetches the model and GPU backend with resume and checksum verification.
• UI (src/): plain TypeScript and CSS, built with Vite.

The index is stored at %APPDATA%\com.rupesh.filedig\index.db. Deleting that file resets it.

Limitations

• Windows only for now.
• No live file watching: changes are picked up on startup or with Rescan.
• Scanned PDFs without a text layer are skipped, and there is no OCR for screenshots.
• Audio, video and source code are not indexed.
• Results are limited to the top 48 matches.
• Text files over 5 MB, images over 50 MB and PDFs over 100 MB are skipped.
• Tested with libraries of a few hundred files. Very large libraries (hundreds of thousands of items) will use a lot of memory, because every vector is held in RAM.

Acknowledgements

• DigUp, the macOS app that inspired this one
• EmbeddingGemma 2 by Google DeepMind (Apache 2.0)
• llama.cpp (MIT)
• Tauri (MIT / Apache 2.0)

License

MIT © 2026 Rupesh Sreeraman