An open-source macOS application called SCM is offering local search across photographs and individual scenes inside video files. The project combines visual embeddings with filename matching, optical character recognition and speech transcripts, while its developer says media and inference remain on the user's Mac without an account or cloud upload.
SCM begins a query with a fast filename-keyword pass before applying a vision model to indexed media. Visual results are ranked through similarity between text and image embeddings, with additional controls intended to reduce weak or repetitive matches. Result tiles identify why an item matched and can expose a score breakdown, giving users more context than a single unexplained relevance ranking.
Video search operates at the scene level rather than treating each file as one object. The application uses FFmpeg to detect shot boundaries, samples frames according to a user-selected density and stores a representative image for each segment. A matching result can therefore open a video at the relevant timecode. The project also caps the scenes returned from an individual video and uses a threshold intended to suppress low-confidence results.
The application maintains separate literal-search paths for visible and spoken text. Tesseract OCR records words and their locations so matched terms can be highlighted inside an image. Whisper transcription supports exact phrase and word matching in dialogue, and opening a result seeks to the corresponding line. Because these paths do not depend on semantic embeddings, the repository says they remain usable while the vision engine is unavailable or still warming up.
SCM supports four switchable vision models through ONNX Runtime. Changing the active model triggers re-embedding of the library, with filename search available during that background work. OCR language packs are optional downloads; English is always enabled, and 35 additional languages can be selected. The default configuration includes Simplified and Traditional Chinese, Japanese and Korean support.
A separate, opt-in chat feature runs through a local `llama.cpp` process bound to the loopback interface. It answers questions from evidence already extracted from the library, such as OCR text, dialogue or filenames, and attaches clickable citations to its responses. The feature does not download or run until enabled in settings, according to the project documentation.
The current repository identifies version 0.2.4 installers and targets Apple Silicon systems running macOS 12 or later, with Homebrew presented as the easiest installation route. The extensive feature list is supplied by the developer rather than independently evaluated, so performance and retrieval quality will depend on hardware, model choice and the user's media collection. Still, SCM illustrates how several established local tools can be assembled into one private search layer for personal visual archives.
