Alibaba's 'zg' tool offers local search for human and AI agents

Summary

Alibaba's zg (zvec-grep) tool is a local-first search layer on top of the zvec vector database that combines ripgrep, BM25, and vector search. It is designed to serve both human CLI users and AI agents querying the same workspace directory. The article walks through the setup, indexing two public domain texts, and a semantic query that correctly surfaces the relevant Sherlock Holmes passage.

Why it matters

AI agents currently consume large amounts of tokens and tool calls to perform grep searches in codebases. A single indexed and ranked search endpoint shared by both humans and agents can significantly reduce this overhead. By keeping files, directories, and local embedding models on the user's machine, it prevents the data leakage and privacy concerns associated with cloud-based RAG pipelines. If the token and latency savings shown in benchmarks hold true in real repositories, zg could become a default layer in agent toolchains.

Highlights

• Combines ripgrep, BM25, and vector search in a single CLI/agent interface, powered by Alibaba's zvec vector database.
• Benchmarks on pylint, matplotlib, and django show significant drops in input tokens, tool calls, and wall-clock time. The most notable improvement in response quality was observed in pylint.
• Local-first by design: files, the index, and the embedding model (potion-retrieval-32m) remain on the user's machine; remote models are only invoked with explicit permission.

Critical take

The benchmarks were self-published by Alibaba. Quality improvements in two of the three test libraries are described as 'marginal.' Furthermore, the practical demo only indexes two plain-text novels, which is far removed from the scattered, multilingual, and multi-thousand-file monorepos where agents actually struggle. Until independent, large-scale evaluations emerge, the token savings claims should be considered promising but unverified.