What this is
On a regular i5 + Iris Xe laptop, open-source tool Recall from developer pdstn12 hits 95% accuracy with a 0.17-second response—and it throws filename matching out the window. The core idea: convert PDFs, Office files, scans, screenshots, audio, and video frames into vectors (numeric representations that let machines understand meaning), then index them all uniformly. Ask "when did we decide to freeze hiring" and the tool lands on the exact second of the recording.
Recall uses Google's EmbeddingGemma 2, running via llama.cpp (a tool for executing AI models locally) on the user's machine. CUDA, Vulkan, and CPU are all supported—no internet connection required once the model is downloaded. OCR is handled by Chinese-developed PaddleOCR, with a Rust engine, GTK 4 interface, and MIT license for free use.
Industry view
Supporters see this as the right direction for "local AI"—data never leaves the machine, responses in milliseconds, no big-tech lock-in—making it especially well-suited for privacy-sensitive professions like lawyers, doctors, and journalists.
The criticism is just as direct. First, EmbeddingGemma 2 is a Google model—offline after download, yes, but the training data, biases, and safety reviews baked into the model are all out of your control. "Local" does not mean "autonomous." Second, the project currently lacks audio transcription and installer support, and Wayland on older desktops has no shortcut bindings—making it effectively nonexistent for non-technical users. Third, 95% accuracy sounds decent, but one wrong hit means missing a critical file. That bar isn't tough enough for serious, high-stakes scenarios.
Impact on regular people
For enterprise IT: if your team includes heavy Linux users or people processing large volumes of scans and legal documents, it's worth evaluating whether to deploy on an internal network—but right now it requires dedicated staff to maintain.
For individual professionals: at this stage it's closer to a geek toy. For typical office workers, it makes more sense to wait until a comparable product lands on macOS or Windows.
For the consumer market: "let AI dig through our old hard drives" is a genuine demand, but no cross-platform, out-of-the-box winner has emerged yet.