Milvus Plugin for Dify

A plugin that integrates Milvus vector database with the Dify platform, providing vector operations for collection management, data insertion, search, and querying.

中文文档 | English

Features

🗂️ Collection Management

• List Collections: View all available collections
• Describe Collection: Get detailed collection information
• Collection Stats: Retrieve collection statistics
• Check Existence: Verify if collections exist

📥 Data Operations

• Insert Data: Add vectors and metadata to collections (vectors must be precomputed by the caller)
• Upsert Data: Insert or update existing data
• Query Data: Retrieve data by ID or filter conditions (auto-detects the primary-key field)
• Delete Data: Remove data from collections (accepts IDs or filter expressions)

When inserting, provide vectors directly in each entity, for example:

[
  {"id": 1, "vector": [0.1, 0.2, 0.3], "metadata": "doc-1"},
  {"id": 2, "vector": [0.4, 0.5, 0.6], "metadata": "doc-2"}
]

🔍 Vector Search

• Similarity Search: Find similar vectors using various metrics
• Filtered Search: Combine vector similarity with metadata filters
• Multi-Vector Search: Search with multiple query vectors
• Custom Parameters: Adjust search behavior parameters

🔀 Hybrid Search

• Combine results from multiple vector fields in one request and rerank with a strategy.
• Endpoint: /v2/vectordb/entities/hybrid_search (the plugin auto-injects dbName).
• Tool name in Dify: Milvus Hybrid Search.

Parameters (Dify tool form)

• collection_name (string, required)
• searches_json (string, required): JSON array of search objects with precomputed vectors. Each object should include at least:
  • data: list of embeddings (precomputed), e.g. [[0.1, 0.2, ...]]
  • annsField: target vector field name
  • limit: per-route top-K
  • Optional per-route keys supported by REST API: outputFields, metricType, filter, params, radius, range_filter, ignoreGrowing
• rerank_strategy (select): rrf or weighted
• rerank_params (string or JSON object): For rrf use { "k": 10 }; for weighted use { "weights": [0.6, 0.4] }
• Optional top-level: limit, offset, output_fields (comma-separated string or JSON array), partition_names (comma-separated string or JSON array), consistency_level, grouping_field, group_size, strict_group_size (boolean), function_score (JSON)

Built-in validation

• Each search item must have annsField (non-empty) and limit (> 0 integer), and provide data (non-empty array) with numeric vectors.
• If rerank_strategy = weighted, rerank_params.weights must be numeric and its length must equal the number of search routes.
• If top-level limit is provided, ensure limit + offset < 16384 (API limit).
• If output_fields or partition_names are provided as strings, they are split by comma after trimming whitespace.

Example

searches_json = [
  {
    "data": [[0.12, 0.34, 0.56]],
    "annsField": "vector",
    "limit": 10,
    "outputFields": ["*"]
  }
]
rerank_strategy = rrf
rerank_params = {"k": 10}
limit = 3
output_fields = "user_id,book_title"

If you manage tool parameters via code, searches_json and output_fields can also be passed as native JSON structures (list/dict) instead of strings. The plugin will normalize both formats.

Installation & Configuration

Connection Configuration

Configure your Milvus connection in the Dify platform:

• URI: Milvus server address (e.g., http://localhost:19530)
• Token: Authentication token (optional, format: username:password)
• Database: Target database name (default: default)

License

MIT License