jina-grep: Adding Semantic Search Capabilities to Grep on Apple Silicon

jina-grep is an open‑source command‑line tool that adds semantic search to traditional grep on Apple Silicon by executing MLX models locally, eliminating the need for external vector‑database services.

Working Modes

Pipeline mode : pipes the output of grep into jina-grep for semantic re‑ranking.

grep -rn "error" src/ | jina-grep "错误处理逻辑"

Standalone mode : performs natural‑language search directly over files. jina-grep "内存泄漏" src/ Zero‑shot classification : assigns the most relevant label to each line of text.

jina-grep -e "数据库" -e "错误处理" -e "数据处理" src/

Performance on Apple M3 Ultra

v5‑small model (677 M parameters): single‑query latency 7 ms, peak throughput 25.3 K tokens / s.

v5‑nano model (239 M parameters): single‑query latency 2.9 ms, peak throughput 98.7 K tokens / s.

Technical Details

The project is built entirely on the MLX framework and does not depend on PyTorch or the transformers library. Model checkpoints are fetched on demand from HuggingFace, and the tool supports automatic batching of up to 256 inputs per request. A lightweight server runs in the background and is invoked through the CLI.

Installation

git clone https://github.com/jina-ai/jina-grep-cli.git && cd jina-grep-cli
uv venv .venv && source .venv/bin/activate
uv pip install -e .

Requirements: Python 3.10+, Apple Silicon Mac, Apache‑2.0 license.

Real‑World Example

Developers report that the tool can replace roughly half of existing search workflows. For instance, locating code that implements an "exponential backoff retry mechanism" traditionally requires exact keyword matches; jina-grep can retrieve semantically similar implementations even when the exact phrase is absent.

Repository

https://github.com/jina-ai/jina-grep-cli