zg launches local-first search layer on Alibaba's zvec

The tool's pitch rests on a split that has become familiar in agent work: the same corpus has to be readable by a person typing a question and by a model that would otherwise burn tokens reading and re-reading files. zg's answer is to build one index that supports keyword matching, ripgrep-style exact and regex checks, and semantic ranking, then let either caller query it. The benchmark graphs shown for pylint, matplotlib and django point in the same direction — fewer input tokens, fewer tool calls, less wall-clock time, and answer quality that improved markedly on pylint and slightly on the other libraries.

The walkthrough suggests how it behaves in practice. Indexing two public-domain texts and asking a riddle-like question about unseen marks returned three ranked blocks from the Sherlock Holmes file, tagged as matched by both keyword and vector search, with line ranges and a "fresh" status. The same question routed through OpenCode produced a specific answer naming the story and the animal, with line-number citations.

On privacy, the design keeps files, indexes and local models on the user's machine and only reaches a remote embedding model with permission, which may matter to teams weighing whether search indexes can leave a laptop. Whether the token and time savings hold at repository scale is not shown in the benchmarks presented.