Vector Retrieval Service for Milvus:Cluster management and business verification through Skill

The alibabacloud-milvus-manage Skill lets you manage Milvus clusters and validate search capabilities using natural language commands. You can create clusters, configure public network access, manage collections, and run scalar, vector, BM25 full-text, multimodal, and cross-modal search — all without writing code.

Prerequisites

• Install the alibabacloud-milvus-manage Skill. Get it from one of these sources:

  • GitHub

  • ClawHub

  • skills.sh

• Install the Skill in your client:

  • Install Skill in OpenClaw

  • Install Skill in Claude Code

• Configure the Alibaba Cloud Command Line Interface (CLI) and authenticate your identity. Run aliyun configure list to verify your current configuration.

• Have a Virtual Private Cloud (VPC) and a vSwitch ready. The Skill prompts you to select existing network resources during execution.

• Create the Milvus service-linked roles (SLRs). On first use, create both AliyunServiceRoleForCloudMilvus and AliyunServiceRoleForMilvus.

Supported scenarios

The alibabacloud-milvus-manage Skill covers eight end-to-end scenarios across two categories:

Scenario 1: Cluster management

Create a cluster

Send a natural language command to create a Milvus instance. The Skill automatically selects a VPC and vSwitch, configures component specifications, and calls the create API.

Example Skill command:

/alibabacloud-milvus-manage Create a Milvus instance of version 2.6 in Hangzhou zone J with password <custom_password>

Result:

// Skill automatically completes VPC/VSwitch selection, component spec configuration, then calls the create API
// Response:
{"success": true, "data": {"instanceId": "c-xxxxxxxxxxxx****", "orderId": "28456242743****"}}

// Creation parameters:
// - Region: cn-hangzhou, Zone: cn-hangzhou-j
// - Version: 2.6, Spec: standalone_pro, 4 CU
// - Billing method: PayAsYouGo

View instance status

After creation, check the instance's running status, endpoints, and specification details.

Example Skill command:

/alibabacloud-milvus-manage Check the creation status

Result:

// Skill calls the list API and detail API to query instance status

Instance list:
c-xxxxxxxxxxxx**** | milvus-skill-test | status=running | v=2.6

Instance details:
- Internal endpoint: c-xxxxxxxxxxxx****-internal.milvus.aliyuncs.com
- Port: 19530
- Total CUs: 4
- Instance status: running (changes from creating to running in about 3 minutes)

Enable public network access

Enable public network access and add your local IP to the allowlist for development or testing.

Example Skill command:

/alibabacloud-milvus-manage Enable public network access for the cluster and add my local IP

Result:

// Skill automatically retrieves local IP and calls updatePublicNetworkStatus API

Public network access enabled:
- Public endpoint: c-xxxxxxxxxxxx****.milvus.aliyuncs.com:19530
- Allowlist: <local_public_IP>/32

Query collections

Connect to the instance and inspect all collections, including sample data.

Example Skill command:

/alibabacloud-milvus-manage Query all collections and show 3 records for each

Result:

// Skill connects to the instance using pymilvus and queries the collection list

Connection: MilvusClient(uri='http://c-xxxxxxxxxxxx****.milvus.aliyuncs.com:19530', token='root:***')

Collection list: [] (new instance has no collections, as expected)

Scenario 2: Service validation

The service validation scenarios confirm that your Milvus cluster handles the full range of search types — from basic scalar and vector search through BM25 full-text search and multi-modal retrieval. Each scenario uses a single Skill command. The Skill handles collection creation, data insertion, embedding generation, and search execution automatically.

Text import and search

Import text data and validate scalar search, vector search (approximate nearest neighbor, or ANN), and hybrid search. The Skill creates a collection backed by AI Function for automatic embedding generation, inserts records from your file, and runs all three search types.

Example Skill command:

/alibabacloud-milvus-manage I have a <data_file_path>/articles.jsonl
file where each line contains title, content, and embedding fields. Import it into the article_db collection;
create it if it doesn't exist. After importing, randomly construct scalar search and vector search tests based on the file content

Replace <data_file_path> with the actual path to the JSONL file. Each line must be a JSON object containing title and content fields.

Result:

// Skill automatically creates article_db collection (with AI Function: text-embedding-v3, 768 dimensions)
// Inserts 10 article records (only text is provided; embeddings are generated automatically)

Collection created: article_db
  Fields: id, title, content, embedding
  AI Function: text-embedding-v3 (dim=768)

Data inserted: 10 records

All three search types run automatically: scalar search (exact or fuzzy match on field values, such as title like "%vector%"), vector search (ANN search by semantic similarity), and hybrid search (vector search combined with a scalar filter).

// Scalar search: title like "%vector%"
Result: found "Introduction to Vector Databases"

// Vector search: "What is machine learning and deep learning"
Results:
  top1: "Summary of Machine Learning Algorithms"       (distance=0.7125)
  top2: "Natural Language Processing Technologies"       (distance=0.5164)
  top3: "Introduction to Vector Databases"         (distance=0.4779)

// Hybrid search: "database technologies" + filter='content like "%vector%"'
Results:
  top1: "Introduction to Vector Databases"         (distance=0.6001)
  top2: "Milvus Best Practices"     (distance=0.5276)

Full-text search and hybrid search

Validate BM25 full-text search alongside semantic vector search in a single collection. Milvus 2.6 AI Function supports configuring both BM25 and an embedding model simultaneously — insert raw text once, and both sparse (BM25) and dense (embedding) vectors are generated automatically.

Example Skill command:

/alibabacloud-milvus-manage I want to do Chinese full-text search in Milvus. Create a new collection with title
and content text fields, use BM25 for full-text search, and add text-embedding-v3 for semantic search with
1024 dimensions. After creation, insert all data from the file, then test vector search, scalar search, full-text + embedding
hybrid search, and full-text + scalar + vector hybrid search using content from the file

Result:

// Skill creates hybrid_search_test collection
// Configures both BM25 Function (chinese analyzer) and text-embedding-v3 AI Function (1024 dimensions)
// Inserts 10 article records (only text is provided; BM25 sparse vectors and embedding dense vectors are generated automatically)

Collection created: hybrid_search_test
  BM25 Function: content_sparse (chinese analyzer)
  AI Function: text-embedding-v3 (dim=1024)

Data inserted: 10 records

Five search validations run automatically, covering all major retrieval strategies:

• Vector search — semantic retrieval using text-embedding-v3

• Scalar search — fuzzy match on field values

• BM25 full-text search — keyword retrieval using the BM25 algorithm

• Full-text + embedding hybrid search — BM25 and vector search fused with RRFRanker

• Three-way hybrid search — BM25 + scalar filtering + vector ANN, fused with RRFRanker

  // Vector search: "How to build a semantic search system"
  Results:
    top1: "Introduction to Vector Databases"         (distance=0.5575)
    top2: "Milvus Best Practices"     (distance=0.5402)
  
  // Scalar search: content like "%Milvus%"
  Results: found "Introduction to Vector Databases" and "Milvus Best Practices"
  
  // BM25 full-text search: "vector database semantic search"
  Results:
    top1: "Introduction to Vector Databases"         (BM25 score=13.3037)
  
  // Full-text + embedding hybrid search (RRFRanker): "deep learning and neural networks"
  Results:
    top1: "Summary of Machine Learning Algorithms"       (score=0.0328)
  
  // Three-way hybrid search (BM25 + scalar filter + vector ANN, RRFRanker):
  //   BM25="index retrieval vector", filter='title contains Milvus or vector', ANN="database indexing and retrieval technologies"
  Results:
    top1: "Introduction to Vector Databases"         (score=0.0325)
    top2: "Milvus Best Practices"     (score=0.0325)

Image import and search

Import images into Milvus and validate text-to-image search. The Skill uses the qwen3-vl-embedding multimodal embedding model to generate embeddings automatically. Sample images are available from here.

Example Skill command:

/alibabacloud-milvus-manage Import all images from the <image_directory_path>/banana
directory, then try searching with "yellow banana"

Replace <image_directory_path> with the actual path to the image directory. The Milvus server must be able to access image URLs — upload images to an OSS bucket in the same region and use signed URLs.

Result:

// Skill creates multimodal_test collection
// Uses qwen3-vl-embedding multimodal model to automatically generate embeddings
// Uploads images to same-region OSS, passes signed URLs to Milvus

Collection created: multimodal_test
  Fields: id, file_path, file_type, description, mm_value, embedding
  AI Function: qwen3-vl-embedding

Data inserted: 6 records (3 banana images + 2 cat images + 1 airplane video)

// Text-to-image search: "yellow banana"
Results:
  top1: banana/img2  (distance=0.5005)
  top2: banana/img3  (distance=0.4953)
  top3: banana/img1  (distance=0.4882)

// Text-to-image search: "cute cat"
Results:
  top1: cat2         (distance=0.5265)
  top2: cat1         (distance=0.4893)

Cross-modal search

Import images and videos together and validate cross-modal retrieval, including image-to-image and image-to-video search.

Example Skill command:

/alibabacloud-milvus-manage First import 2 files from the tiger_cat directory and airplane1.mp4 into the database,
then use any file from tiger_cat to search for similar product images and return the top 3. Then use an image to search for similar videos

Result:

// Image-to-image search: search similar images using cat1 image URL
Results:
  top1: cat1         (distance=0.9996)
  top2: cat2         (distance=0.9476)
  top3: banana/img1  (distance=0.7483)

// Image-to-video search: search using cat1 image URL + filter='file_type=="video"'
Results:
  top1: airplane.mp4 (distance=0.7180)