Building a Resource-Efficient Semantic Search System with Milvus Lite and GTE

This article explores how to implement a lightweight semantic search solution using Milvus Lite, GTE, and Rerank tools through Docker deployment. The system can run on just 1GB of RAM - even on a Raspberry Pi - while supporting local data storage and migration. It’s perfect for local knowledge bases or small-scale applications.

Project Overview

GitHub Repository: https://github.com/BetaStreetOmnis/milvus_lite_server

Key Components

Vector Database

We use Milvus Lite, a lightweight version of Milvus, for efficient storage and retrieval of large-scale vector data in resource-constrained environments.

Embedding Model

For text vectorization, we utilize the nlp_gte_sentence-embedding_chinese-base model, which outperforms both BGE and M3E in our testing.

Rerank Model

To optimize query result precision and improve search relevance, we implement a reranking mechanism based on the gte_passage-ranking_multilingual-base model.

Docker Deployment and Resource Optimization

System Requirements

• Minimum: 1 CPU core, 1GB RAM
• CPU-only operation supported
• Local storage with data migration capabilities

Docker Setup