Semlith Files

Added

• A CPU cap. SEMLITH_CPU_CAP, the Limits card's CPU cap row or cpu_cap_percent in settings.json sets the most of this machine's CPU semlith's own process may use while it indexes, from 0 to 100 % of every core. semlith reads its own CPU time and holds a rolling two-second window under the cap by pacing file preparation, the CPU lane, the image lane and the writer; while capped, it also keeps files prepared at once, CPU embedding threads and vector encoding inside the cap's share of the cores, so one batch cannot burst past it. 0 % pauses CPU work and the run card says so. Measured on an M1 with eight threads: 25 % and 50 % caps held 30 of 30 one-second samples under the cap plus five points. The Limits card and semlith accel status show what semlith is using now beside the cap.

Changed

• The CPU lane is always on. semlith accel off cpu refuses, the switch is locked, and a saved "off" from an older release reads as on. The cap is the control instead. Its row says active only while it embeds.
• The switches tell the truth. Each lane reports what was saved and where its state came from beside what runs use, so a page draws the choice that was made and names an environment override. Which lanes can run is read from the machine's hardware: no usable GPU makes the GPU and llama.cpp lanes unavailable, no NVIDIA card and driver makes CUDA and TensorRT unavailable, and a non-Intel CPU makes OpenVINO unavailable. An unavailable or failed lane can be switched off but never on, and the refusal says why; switching a failed lane off clears the failure.

Fixed

• A first index through a remote GPU is 1.39x faster (#207). When a file ran past the writer's 256-chunk window, its hand-over committed the window and nothing reopened the transaction, so every remaining row and the file's whole graph went to disk one autocommit at a time. The writer now keeps them in a transaction. microsoft/vscode (353,398 chunks) from a 4-vCPU EPYC box through a Confidential G4: median of three 437 s against 606 s for rc.5 (808 against 583 chunks a second), the write stage from about 400 s to 215 s (graph 185 to 50 s, rows 97 to 15 s). The store is the same: files, chunks, keyword rows, symbols, edges and images match rc.5's counts. The writer's commit (about 90 s) is now its largest part; [#207] stays open for the rest of the way to 2x.
• images in index -v's write parts now counts only the writer's work, not the image model's own thread.

Install

macOS and Linux

:::sh
curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/semlith/semlith/main/install.sh | sh

Windows

:::powershell
irm https://raw.githubusercontent.com/semlith/semlith/main/install.ps1 | iex

Or download the archive for your platform below, unpack it, and put semlith on your PATH. Intel macOS has no prebuilt binary: ONNX Runtime no longer publishes x86_64 macOS builds.

Verify a download against SHA256SUMS before running it, and its provenance with gh attestation verify <file> --repo semlith/semlith.