Search supported documents, photos, audio, and video by filename, contents, metadata, or meaning. EmbedCore indexes the folders you choose locally and keeps them current in the background.
Exact when you know it. Semantic when you don't.
Search the way you remember. Use exact names when you know them, natural language when you don't, and filters when you need precision. EmbedCore keeps selected folders indexed in the background and can surface duplicate files without sending your library to a cloud service.
One search surface combines filenames, extracted text, metadata, and local vector indexes. Exact matches can appear immediately while semantic results populate alongside them.
Find filenames, paths, and extracted text immediately when you know the words you're looking for.
Describe what you remember in plain language and retrieve conceptually similar files even when the wording differs.
Narrow results by file type, date, size, modality, and metadata such as BPM or image resolution when available.
Surface exact duplicates and, where supported, near-duplicate content without sending files off-device.
In local mode, your files, embeddings, and metadata stay on your machine. Optional online providers can be enabled explicitly with your own credentials, but they are never required for core local search.
From documents to photos to music to video — EmbedCore modules declare the file types they support and bring those indexed modalities into one search interface.
PDFs, markdown, plain text — semantic search across your entire document library by meaning, not filenames.
EXIF data, visual embeddings, and content analysis — find images by what's in them, not what they're called.
BPM, key, timbre, loudness — search your music by how it sounds. Find similar tracks instantly by timbral similarity.
Container metadata, frame analysis, and temporal embeddings — index and search video content by visual and audio similarity.
When a module uses a compatible multimodal model, EmbedCore can search across modalities from one interface — for example, using text to find related images, audio, or video. Incompatible embedding spaces remain separate instead of being compared as if they were interchangeable.
Modules are standalone executables that communicate via a simple JSON protocol. Write them in any language — Rust, Python, Go, C++ — with zero framework lock-in.
Write modules in any language. If it can read stdin and write stdout, it can be an EmbedCore module.
Each module bundles its own models and dependencies. Install or remove modules without affecting anything else.
No framework assumptions. Modules choose their own ML runtime — ONNX, PyTorch, TensorFlow, or pure DSP.
Filesystem watcher detects new and changed files in real time
Two-tier detection: fast mtime check, then SHA-256 hash for accuracy
User requests first, then new files, then reprocessing jobs
Topological sort ensures modules execute in the right order
Independent modules process files in parallel for maximum throughput
The processing pipeline handles everything — from detecting file changes to orchestrating module execution — so your library stays indexed without you thinking about it.
Filesystem events trigger processing automatically. Add files to a watched folder and they're indexed within moments.
Dependency graph analysis prevents circular module chains via depth-first search before any processing begins.
Core indexing and supported local models run on your hardware. Optional online providers are used only when you explicitly enable them.
Built as a Tauri and Rust desktop app with a local background service. The first releases target modern desktop hardware, with local acceleration where available.
Desktop build optimized first for Apple Silicon, with Finder integration planned for fast search access.
Linux desktop builds for modern distributions, with packaging and acceleration support refined per release.
Windows desktop build using WebView2, with File Explorer integration planned for fast search access.
EmbedCore is being built for people who believe their files — and the intelligence extracted from them — should stay on their machine. Be the first to know when it launches.
No spam. One email when we launch.