How Zilliz’s Two Skills Enable AI to Code with pymilvus and Manage Cloud Clusters

1. Two Skills, Two Roles

Milvus Skill and Zilliz Cloud Skill are independent tools aimed at different stages of a vector‑database workflow. Milvus Skill teaches an AI agent to use the pymilvus Python SDK for data operations, while Zilliz Cloud Skill teaches the agent to manage cloud resources through the zilliz-cli command line.

A side‑by‑side diagram (Fig. 2) highlights the architectural split: SDK‑based coding versus CLI‑based DBA assistance.

2. Milvus Skill – AI‑Powered Coding Assistant

Modular Design (7 reference files)

collection.md

: collection schema and CRUD vector.md: vector CRUD, hybrid and full‑text search, iterators index.md: index types and management partition.md: partition handling database.md: database management user-role.md: RBAC permissions patterns.md: common patterns such as RAG and hybrid search

Each file isolates a knowledge domain so the AI can load only the relevant context, conserving the limited prompt window.

Supported Data Types

Scalar types include BOOL, INT8/16/32/64, FLOAT, DOUBLE, VARCHAR, JSON, ARRAY. Vector types cover FLOAT_VECTOR, FLOAT16_VECTOR, BFLOAT16_VECTOR, BINARY_VECTOR, SPARSE_FLOAT_VECTOR, and INT8_VECTOR, each suited to different use cases.

Hybrid Search

Hybrid search combines a dense vector (semantic) path and a sparse vector (keyword) path, then merges results with Reciprocal Rank Fusion (RRF) or weighted ranking. The Skill documents the required components:

AnnSearchRequest : request object for each search lane

RRFRanker / WeightedRanker : result‑fusion strategies SPARSE_FLOAT_VECTOR: field for sparse vectors

RRF works by summing the reciprocal ranks of each lane; higher‑ranked documents receive higher scores. WeightedRanker applies a static weight to each lane.

Full‑Text Search with BM25

Since Milvus 2.6+, built‑in BM25 support (via Function and FunctionType.BM25) automatically converts text to sparse vectors, eliminating the need for a separate embedding model.

Search Iterators

For large collections, search_iterator and query_iterator provide cursor‑based pagination, avoiding the limits of a fixed‑size search response.

3. Zilliz Cloud Skill – AI‑Powered DBA Assistant

Command‑Mode Design

The Skill instructs the AI to construct CLI commands of the form:

zilliz <resource> <action> --flag <value>

Examples: zilliz cluster list – list all clusters zilliz collection create --name my_collection – create a collection zilliz backup create --name daily-backup – create a backup

Output can be formatted as json, table, or text depending on the context.

Fourteen Reference Files (Full Lifecycle)

setup.md

: installation, authentication, context setup cluster.md: create/list/modify/pause/resume/delete clusters collection.md: collection CRUD, load/release, alias vector.md: vector operations via CLI index.md: index management database.md: database management (Dedicated clusters only) partition.md: partition handling user-role.md: RBAC (Dedicated clusters only) backup.md: backup creation, restore, export, policies import.md: bulk import from cloud storage (Parquet/JSON/CSV) billing.md: usage queries, invoices, payment methods monitoring.md: cluster status, statistics, load project-region.md: project and region management job.md: asynchronous task tracking

Security Design – Three Rules

Confirm destructive operations (e.g., cluster deletion) with the user before execution.

Execute credential‑sensitive commands in the user’s terminal, not via AI relay.

Never expose API keys or passwords in logs or chat output.

Pre‑condition Checks (Three‑Step Flow)

Verify zilliz-cli is installed; install if missing.

Check user login status; guide authentication if needed.

Ensure a cluster context is set; prompt selection if absent.

4. Why Skill Instead of MCP?

Skill acts as a prompt loader: a folder with a SKILL.md manifest and optional reference files that inject domain knowledge into the AI’s context. MCP (Model Context Protocol) provides a standardized tool‑calling API via a running server process.

Comparison (Fig. 5):

Essence : Skill = knowledge injection; MCP = protocol‑based tool calls.

Deployment cost : Skill requires only files; MCP needs a server.

Suitable scenario : Skill teaches AI to use existing tools; MCP adds new capabilities.

Zilliz maintains both paradigms, ensuring coverage regardless of the ecosystem’s direction.

5. Practical Scenarios

Scenario 1 – Building a RAG Application

Prompt: “Create a document‑retrieval collection, 768‑dim vectors, BM25 full‑text, fields title, content, embedding.” The AI reads collection.md and patterns.md and generates Python code that defines the schema, adds a BM25 function, and creates the collection.

Scenario 2 – Managing Zilliz Cloud Clusters

Prompt: “Create a Serverless cluster in us‑east‑1 and a 768‑dim collection.” The AI checks the environment, then runs a sequence of zilliz CLI commands (e.g., zilliz cluster create, zilliz collection create).

Scenario 3 – Backup and Data Migration

Prompt: “Set a daily backup retaining 7 days for the production cluster.” The AI follows backup.md. For export, it uses import.md to generate commands that export a collection to S3.

Scenario 4 – Hybrid Search Optimization

Prompt: “Upgrade my search to dense + sparse hybrid with RRF.” The AI consults vector.md for AnnSearchRequest and RRFRanker, then emits the combined search code.

6. Ecosystem Positioning

Zilliz’s AI‑Agent toolchain spans MCP servers, Skills, and Plugins. Notable projects include claude-context (5.9k Stars) that stores code in a vector DB and reduces token usage by ~40 %, and DeepSearcher (≈5k Stars) showcasing multi‑round knowledge retrieval.

The company’s strategy is clear: regardless of whether the AI‑Agent ecosystem adopts MCP, Skill, or Plugin, the vector‑database layer must have a compatible integration.

Conclusion

Clear division of labor: Milvus Skill handles SDK‑level coding; Zilliz Cloud Skill handles CLI‑level operations.

Fine‑grained modularization (7 + 14 files) maximizes prompt efficiency.

Robust security: destructive‑action confirmation, credential protection, and pre‑flight checks.

Strategic ecosystem placement: Zilliz covers both Skill and MCP paradigms, ensuring future‑proof AI‑Agent support.

Developers using AI agents for vector‑database tasks should consider installing both Skills to streamline RAG development and cloud‑cluster management.