Milvus Marketplace

A data retrieval development assistant based on Claude Code Skills.

Core Philosophy

Focus: Data Retrieval Domain

We specialize in the data retrieval vertical:

• Vector search, semantic search
• RAG, knowledge base Q&A
• Image search, recommendation systems
• Milvus / Zilliz Cloud

Two-Layer Solution Architecture

┌─────────────────────────────────────────────────────────┐
│              Scenario Plugins (6 plugins)               │
│   rag-toolkit, retrieval-system, multimodal-retrieval,  │
│   rec-system, memory-system, data-analytics             │
│   = Pre-built solutions = AI era caching mechanism      │
└─────────────────────────────────────────────────────────┘
                           ↑
                     Match / Combine
                           ↑
┌─────────────────────────────────────────────────────────┐
│                      core plugin                         │
│       Methodology (pilot) + Atomic operators            │
│              (embedding, chunking, ...)                 │
└─────────────────────────────────────────────────────────┘

Core Ideas:

1. Scenario plugins = AI cache: Pre-built scenario solutions act like cache - use directly if matched, no need to build from scratch
2. core = methodology + operators: Universal development methodology + composable atomic capabilities
3. Ship 60% first, iterate to 80%: Deliver working solutions quickly, then optimize based on feedback

Scenario Classification Principles

Scenarios are classified by architectural differences, not by industry or model:

• Different architecture = different scenario: Fundamentally different processes and code structures
• Only model/parameters differ = vertical applications within same scenario: Covered through configuration tables

plugins/retrieval-system/skills/
├── semantic-search/          # Category: architecture definition
│   ├── SKILL.md              # Generic workflow + model selection table
│   └── verticals/            # Subcategory: vertical application guides
│       ├── legal.md          # Legal search
│       ├── academic.md       # Academic papers
│       └── ecommerce.md      # E-commerce search

Workflow

User describes requirement
      │
      ▼
   pilot activates
      │
      ├─→ Clarify data and query
      │
      ├─→ Can match a scenario?
      │       ├─ Yes → Use pre-built solution
      │       └─ No → Combine core operators
      │
      ├─→ Generate code → User tests
      │
      └─→ Collect feedback → Iterate

Installation

1. Add Marketplace

/plugin marketplace add zilliztech/milvus-marketplace

2. Install Plugins

# Core tools (required)
/plugin install core@milvus-marketplace

# Install scenario plugins as needed
/plugin install rag-toolkit@milvus-marketplace        # RAG solutions
/plugin install retrieval-system@milvus-marketplace   # Text search
/plugin install multimodal-retrieval@milvus-marketplace # Image/video/multimodal
/plugin install rec-system@milvus-marketplace         # Recommendations
/plugin install memory-system@milvus-marketplace      # Chat memory
/plugin install data-analytics@milvus-marketplace     # Duplicate detection, clustering

3. Start Using

Simply describe what you want to build:

"Help me build a document Q&A system"
"I want to implement semantic search"
"Build an image search application"

The pilot will automatically activate, clarify requirements, and help you orchestrate the toolchain and generate code.

Plugin Overview

core - Core Capabilities

Type	Skill	Purpose
Controller	pilot	AI application navigator - understands requirements, orchestrates tools, delivers code
Operator	embedding	Text/image vectorization
Operator	chunking	Document chunking
Operator	indexing	Milvus index management
Operator	data-ingestion	Batch data import
Operator	rerank	Search result reranking
Operator	pdf-extract	PDF text extraction
Operator	vlm-caption	Image captioning (VLM)
Environment	local-setup	Local Milvus deployment

Scenario Plugins - Pre-built Solutions (6 Plugins, 17 Skills)

retrieval-system - Text Search

Skill	Architecture	Vertical Applications
semantic-search	embedding → vector search	Legal, academic, news, e-commerce, code, patents
hybrid-search	vector + BM25 keyword + score fusion	E-commerce, legal, academic
filtered-search	vector search + scalar filtering	E-commerce, recruitment, real estate
multi-vector-search	multi-vector field joint search	Products, papers, resumes

rag-toolkit - RAG / Q&A

Skill	Architecture	Vertical Applications
rag	retrieval → LLM generation	Enterprise KB, product docs, policies, customer service
rag-with-rerank	retrieval → rerank → LLM	Legal consulting, medical Q&A, financial reports
multi-hop-rag	multi-round retrieval (results guide next round)	Complex research, fact-checking, troubleshooting
agentic-rag	Agent autonomously decides retrieval strategy	Smart assistants, research agents

multimodal-retrieval - Image / Video / Multimodal

Skill	Architecture	Vertical Applications
image-search	CLIP embedding → vector search	E-commerce visual search, fashion, face, logo, design assets
text-to-image-search	VLM captioning → text embedding	Stock images, surveillance retrieval, medical imaging

rec-system - Recommendation

Skill	Architecture	Vertical Applications
item-to-item	Item vector similarity	Similar products, related articles, similar videos
user-to-item	User vector + item vector + recall	Personalized recommendations, feeds, job matching

data-analytics - Detection / Analytics

Skill	Architecture	Vertical Applications
duplicate-detection	Batch vector comparison + threshold	Plagiarism detection, content deduplication, resume deduplication
clustering	Vector clustering analysis	Topic clustering, user segmentation, anomaly detection

memory-system - Conversation / Memory

Skill	Architecture	Vertical Applications
chat-memory	Conversation vectorization + time decay	Long-term assistants, customer service memory, game NPCs

Note: multimodal-rag and video-search are also in multimodal-retrieval plugin.

Tech Stack (Fixed)

No need to choose - we've already decided for you:

Purpose	Technology	Reason
Language	Python	Best ecosystem, seamless integration with Milvus/ML
External APIs	FastAPI	Simple and efficient, built-in Swagger docs
Data Processing	Ray	Faster than Spark, GPU support
Environment Management	uv	100x faster than pip
Vector Storage	Milvus / Zilliz Cloud	Professional vector database

Quick Start Examples

Semantic Search

I have 50,000 news headlines and want to do semantic search

The pilot will:

1. Confirm data (text, 50k items)
2. Confirm query (semantic search)
3. Auto-select: BGE model + Milvus Lite + HNSW index
4. Generate complete code

Document Q&A

I have 200 PDF documents and want AI to answer questions about them

The pilot will:

1. Confirm data (PDF, 200 files)
2. Confirm query (Q&A)
3. Auto-select: PDF extraction + chunking + BGE + RAG
4. Generate complete code

Image Search

100,000 product images, users upload images to find similar ones

The pilot will:

1. Confirm data (images, 100k)
2. Confirm query (image-to-image search)
3. Auto-select: CLIP model + Zilliz Cloud + HNSW
4. Generate complete code

Directory Structure

milvus-marketplace/
├── .claude-plugin/
│   └── marketplace.json              # Marketplace definition (7 plugins)
├── plugins/
│   ├── core/                         # Core capabilities
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       ├── pilot/                # Controller + methodology
│   │       ├── embedding/            # Atomic operators
│   │       ├── chunking/
│   │       ├── indexing/
│   │       ├── rerank/
│   │       ├── ray/
│   │       └── local-setup/
│   ├── rag-toolkit/                  # RAG solutions (4 skills)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       ├── rag/
│   │       ├── rag-with-rerank/
│   │       ├── multi-hop-rag/
│   │       └── agentic-rag/
│   ├── retrieval-system/             # Text search (4 skills)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       ├── semantic-search/
│   │       ├── hybrid-search/
│   │       ├── filtered-search/
│   │       └── multi-vector-search/
│   ├── multimodal-retrieval/         # Image/video/multimodal (4 skills)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       ├── image-search/
│   │       ├── text-to-image-search/
│   │       ├── video-search/
│   │       └── multimodal-rag/
│   ├── rec-system/                   # Recommendations (2 skills)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       ├── item-to-item/
│   │       └── user-to-item/
│   ├── memory-system/                # Long-term memory (1 skill)
│   │   ├── .claude-plugin/plugin.json
│   │   └── skills/
│   │       └── chat-memory/
│   └── data-analytics/               # Data analysis (2 skills)
│       ├── .claude-plugin/plugin.json
│       └── skills/
│           ├── duplicate-detection/
│           └── clustering/
└── README.md

Installation Options

# Default: user-level, available to all projects
/plugin install core@milvus-marketplace

# Project-level: team sharing (committed to git)
claude plugin install core@milvus-marketplace --scope project

# Local-level: current project only (gitignored)
claude plugin install core@milvus-marketplace --scope local

Managing Plugins

# View installed
/plugin

# Update marketplace
/plugin marketplace update milvus-marketplace

# Uninstall
/plugin uninstall core@milvus-marketplace

# Remove marketplace
/plugin marketplace remove milvus-marketplace

Contributing

Contributions of new scenario solutions or core operators are welcome.

• New scenarios: Add under plugins/<category>/skills/ (e.g., plugins/rag-toolkit/skills/)
• New operators: Add under plugins/core/skills/
• New vertical applications: Add under the corresponding scenario's verticals/