Qwen introduces zg, a unified local search layer combining ripgrep, BM25 and vector embedding

The Qwen development team unveiled zg (zvec-grep), an open-source local search layer designed to spare agents the high costs of multiple search-tool calls. When the target is a known symbol, ripgrep returns an exact match; when the target is a natural-language behavior description, keyword-based matching tends to miss, forcing the agent to guess terms, read whole files and build context manually – steps that consume tool calls, tokens and wall-clock time. zg provides a single interface that replaces the three approaches – semantic search, BM25 ranking and exact literal search – and is released under the Apache 2.0 license, which permits commercial use.

Installation is performed via npm under the package name **@zvec/zvec-grep** and requires Node.js 22 or newer on macOS, Linux or Windows. The default model runs without a GPU, and the code is hosted by the zvec-ai organization on GitHub.

zg creates a single index for each workspace and then offers four retrieval modes: the hybrid default that blends intent with lexical anchors, **--fts** which performs precise ranking using BM25, **--vector** which finds conceptual similarity without lexical ranking, and **--rg** which performs literal or full-regex matching without an index – a feature useful when the repository has not yet been indexed.