Skip to content
- A docs section for docs.pitomd.com: an overview, install, guides to searching by meaning, getting the model, picking a device, knowing the machine, measuring the speed and shipping a binary, the formats, the API reference and this changelog. Every Release now carries it as
docs-vX.Y.Z.tar.gz, built by bin/docs-bundle.
embed and probe are both 1.0.1; pin them at tag = "v1.0.1". The library is unchanged.
examples/bench.rs times loading the model, a query and 256 lines on your GPU and on your CPU: on an RX 9060 XT over Vulkan, 256 lines take 477 ms, 6.3× faster than the CPU.examples/machine.rs prints what probe finds: the system, the card, the best lane for each kind of work, the adapters and the free disk space.
embed and probe are both 1.0.0; pin them at tag = "v1.0.0".- A new README with clips of the search, the bench and the machine probe, and a speed chart.
- The brand now reads PITO MD.
examples/search.rs: downloads the model into the standard data folder, loads it on the device the probe picks and ranks eight lines against a query.
- The model catalog points at the upstream files on Hugging Face, onnx-community/embeddinggemma-300m-ONNX pinned to one commit, with the same sha256 as before.
- Both crates are MIT licensed, and pito-fetch v0.1.2 is pinned over https.
- The
probe crate joins the workspace in crates/probe, so embed and probe are pinned from one repository at one tag.
- pito-fetch v0.1.1, which polls body reads on ureq 3.4.1 and later.
- On-device EmbeddingGemma-300m through fastembed-rs, on the GPU or the CPU, with a catalog of the model’s files and their sha256.
