HOW IT WORKS
EveryCli combines local lexical and semantic retrieval, served by a Rust daemon that keeps the model ready in memory.
The local daemon
Loading the semantic model for every request would add unnecessary startup time and CPU work. EveryCli keeps it loaded in a local background daemon (a native Rust binary, installed as a Windows service or a systemd --user service on Linux) and communicates through a local TCP socket.
- —The daemon keeps the ONNX Runtime session and model loaded in memory.
- —The client sends the query and displays the result.
- —Measured warm-search latency: single-digit milliseconds of inference, ~130-210ms end-to-end from the client.
Hybrid retrieval
EveryCli combines three signals, not a single search strategy. Lexical matching is fast and precise for exact terms such as git; semantic matching (via ONNX Runtime) understands intent even when the wording differs entirely from the stored command description.
- —A custom lexical scorer handles exact keyword overlap.
- —A multilingual MiniLM model (fine-tuned on the EveryCli corpus, run locally via ONNX Runtime) contributes semantic similarity.
- —A namespace hint (e.g. the word “docker” in the query) adds a bonus to matching results -- it is never a hard filter, so a personal command added via `everycli add` in a different namespace stays reachable.
Local command corpus
Command scenarios live in YAML files under everycli/data/commands/. Each scenario carries descriptions, tags, per-platform command variants, explanations, and optional warnings.