Xpert Plugin: Milvus Vector Store

Overview

This package exposes the Milvus vector store integration for the XpertAI platform. It wraps the official @zilliz/milvus2-sdk-node client with an opinionated LangChain-compatible adapter that promotes frequently used metadata fields, handles hybrid-search ready schemas, and registers a server-side strategy via @xpert-ai/plugin-sdk.

Key Features

• Registers a global VectorStoreStrategy named milvus, ready to be consumed by the XpertAI agent runtime.
• Provides an improved LangChain Milvus adapter that stores raw metadata safely while promoting filterable fields like knowledgeId, documentId, and chunkId.
• Supports sanitized collection names and automatic partition provisioning for each knowledge base.
• Enables configurable hybrid search, analyzer parameters, and credential management through environment variables.
• Ships with lifecycle hooks (onStart, onStop) and structured logging for observability.

Installation

To use the plugin inside an XpertAI deployment: add this plugin to the PLUGINS environment variable when starting the XpertAI system, and it will be loaded automatically:

PLUGINS=@xpert-ai/plugin-dify

Configuration

The plugin relies on NestJS ConfigService to resolve Milvus connection details. Set the following environment variables (or corresponding config entries) in your host application:

| Variable | Description | Default | | -------- | ----------- | ------- | | MILVUS_URI | HTTP or gRPC endpoint of your Milvus instance, e.g. http://localhost:19530 | http://127.0.0.1:19530 | | MILVUS_USER | Username for Milvus authentication (if required) | null | | MILVUS_PASSWORD | Password for Milvus authentication (if required) | null | | MILVUS_TOKEN | Token-based auth string for Milvus Cloud or managed deployments | null | | MILVUS_DATABASE | Target database name | default | | MILVUS_ENABLE_HYBRID_SEARCH | Enables scalar + vector hybrid search features | true | | MILVUS_ANALYZER_PARAMS | JSON string describing analyzer params, e.g. {"type":"chinese"} | null |

Metadata Filtering

The adapter promotes several metadata fields (enabled, knowledgeId, documentId, chunkId, parentChunkId, model) for efficient filtering. Additional metadata is serialized into a JSON column, so arbitrary attributes remain queryable via hybrid search.

Deletion Helpers

The wrapped vector store overrides delete() to accept either LangChain-style filters or direct chunk ID lists. Internally it builds Milvus filter expressions like chunk_id in [...].

Deployment

• npx nx release -p @xpert-ai/plugin-milvus runs package tests.
• npx nx run @xpert-ai/plugin-milvus:nx-release-publish --access public --otp=<one-time-password-if-needed> publishes to npm.

Requirements

• Node.js 20+
• Milvus 2.4+ (2.5+ recommended to leverage hybrid search)

Additional Resources

• Milvus documentation: https://milvus.io/docs
• XpertAI platform: https://xpertai.cloud