feat: add configurable metric type for Milvus vector store

[xxiaoxiong]

Description

Fixes #6033

Adds a configurable Metric Type parameter to the Milvus vector store node, allowing users to choose between L2, IP, and COSINE distance metrics instead of being locked to L2.

Problem

Current behavior:

• Metric type is hardcoded to MetricType.L2 when creating indexes
• Queries fail with metric type mismatch errors when the collection uses COSINE or IP
• Users cannot configure the metric type for their use case

User report from #6033:

"In the Milvus Node implementation, the metric type is implicitly set to L2 for both upsert and query operations. However, Milvus supports multiple metric types, including L2, COSINE, and IP. This hardcoded use of L2 leads to metric type mismatch errors when querying collections configured with COSINE or IP."

Error Example

Error: Metric type mismatch: collection uses COSINE but query uses L2

Solution

Add a dropdown parameter that lets users select the metric type:

• L2 (Euclidean Distance) - default, maintains backward compatibility
• IP (Inner Product) - for maximum inner product search
• COSINE (Cosine Similarity) - for normalized similarity

Changes

packages/components/nodes/vectorstores/Milvus/Milvus.ts

1. Added Metric Type Input Parameter

{
    label: 'Metric Type',
    name: 'metricType',
    description: 'Distance metric for similarity search. Default to L2 (Euclidean distance)',
    type: 'options',
    options: [
        { label: 'L2 (Euclidean Distance)', name: 'L2' },
        { label: 'IP (Inner Product)', name: 'IP' },
        { label: 'COSINE (Cosine Similarity)', name: 'COSINE' }
    ],
    default: 'L2',
    additionalParams: true,
    optional: true
}

2. Read Metric Type in Both Methods

// In upsert method (line 220-221)
const metricType = (nodeData.inputs?.metricType as string) ?? 'L2'

// In init method (line 309-310)
const metricType = (nodeData.inputs?.metricType as string) ?? 'L2'

3. Pass to MilvusLibArgs

// Set metric type for index creation
milVusArgs.indexCreateParams = {
    metric_type: MetricType[metricType as keyof typeof MetricType]
}

4. Use in Index Creation

- metric_type: MetricType.L2
+ metric_type: this.indexCreateParams?.metric_type || MetricType.L2

Behavior

Before (Hardcoded L2)

// Always L2, no configuration
const resp = await this.client.createIndex({
    collection_name: this.collectionName,
    field_name: this.vectorField,
    index_name: `myindex_${Date.now().toString()}`,
    index_type: IndexType.AUTOINDEX,
    metric_type: MetricType.L2  // ❌ Hardcoded
})

Problems:

• ❌ Cannot use COSINE or IP
• ❌ Mismatch errors with existing COSINE/IP collections
• ❌ No flexibility

After (Configurable)

// User selects metric type from dropdown
const resp = await this.client.createIndex({
    collection_name: this.collectionName,
    field_name: this.vectorField,
    index_name: `myindex_${Date.now().toString()}`,
    index_type: IndexType.AUTOINDEX,
    metric_type: this.indexCreateParams?.metric_type || MetricType.L2  // ✅ Configurable
})

Benefits:

• ✅ User can select L2, IP, or COSINE
• ✅ No mismatch errors
• ✅ Defaults to L2 for backward compatibility
• ✅ Existing flows work without changes

Use Cases

1. COSINE for Normalized Embeddings

When to use: OpenAI embeddings, Sentence Transformers (normalized)

Metric Type: COSINE
→ Measures angle between vectors
→ Range: [-1, 1], higher is more similar
→ Best for normalized embeddings

2. IP for Maximum Inner Product

When to use: Recommendation systems, collaborative filtering

Metric Type: IP
→ Measures dot product
→ Higher values = more similar
→ Best for non-normalized vectors

3. L2 for Euclidean Distance (Default)

When to use: General purpose, spatial data

Metric Type: L2
→ Measures straight-line distance
→ Lower values = more similar
→ Default for backward compatibility

Testing

Manual Testing

Test 1: Create Collection with COSINE

1. Add Milvus node to flow
2. Set Metric Type to "COSINE (Cosine Similarity)"
3. Upsert documents
4. Query the collection

Expected:

• ✅ Index created with COSINE metric
• ✅ Queries use COSINE metric
• ✅ No mismatch errors
• ✅ Similarity scores in [-1, 1] range

Test 2: Create Collection with IP

1. Add Milvus node to flow
2. Set Metric Type to "IP (Inner Product)"
3. Upsert documents
4. Query the collection

Expected:

• ✅ Index created with IP metric
• ✅ Queries use IP metric
• ✅ No mismatch errors

Test 3: Default L2 (Backward Compatibility)

1. Add Milvus node to flow
2. Leave Metric Type as default (or don't set it)
3. Upsert documents
4. Query the collection

Expected:

• ✅ Index created with L2 metric
• ✅ Queries use L2 metric
• ✅ Existing flows work without changes

Test 4: Query Existing COSINE Collection

1. Connect to existing Milvus collection with COSINE metric
2. Set Metric Type to "COSINE"
3. Query the collection

Expected:

• ✅ Queries work correctly
• ✅ No mismatch errors
• ✅ Correct similarity scores

Edge Cases

• ✅ Missing parameter: Defaults to L2
• ✅ Existing index: Uses existing index, doesn't recreate
• ✅ Existing flows: Work without changes (default L2)
• ✅ Type safety: TypeScript enum conversion

Impact

Affected Components

• ✅ Milvus vector store node (upsert)
• ✅ Milvus vector store node (query)
• ✅ Index creation logic
• ✅ Similarity search logic

Breaking Changes

• ❌ None - defaults to L2 for backward compatibility

Backward Compatibility

• ✅ Existing flows without metricType → use L2 (current behavior)
• ✅ No changes required to existing flows
• ✅ No API changes
• ✅ No database schema changes

Benefits

• ✅ Fixes metric type mismatch errors (Milvus metric type is not configurable #6033)
• ✅ Enables COSINE and IP metrics
• ✅ Aligns with Milvus best practices
• ✅ User-friendly dropdown UI
• ✅ Clear option labels with descriptions
• ✅ Maintains backward compatibility

Screenshots

New Parameter in UI

The Metric Type dropdown appears in Additional Parameters section with three options:

• L2 (Euclidean Distance) - default
• IP (Inner Product)
• COSINE (Cosine Similarity)

Documentation

The parameter includes helpful descriptions:

• Label: "Metric Type"
• Description: "Distance metric for similarity search. Default to L2 (Euclidean distance)"
• Options: Clear labels explaining each metric type
• Default: L2 (maintains backward compatibility)

Related Issues

• Fixes Milvus metric type is not configurable #6033
• Addresses user-reported metric type mismatch errors
• Aligns with Milvus documentation on metric types

Type of Change

• New feature (non-breaking change which adds functionality)
• Bug fix (fixes metric type mismatch errors)
• Breaking change

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• Backward compatible (defaults to L2)
• No breaking changes
• Addresses the exact problem described in Milvus metric type is not configurable #6033
• User-friendly dropdown UI
• Clear option labels and descriptions
• TypeScript type safety
• Falls back to L2 if not configured

[gemini-code-assist]

Code Review

This pull request introduces the ability to configure the metric type for Milvus vector stores, adds a 'Select All' functionality to the workspace sharing dialog, and updates the tooltip component to use unordered lists. A suggestion was made to simplify the state logic for the 'Select All' checkbox in the workspace dialog to remove redundant ternary operations.

[gemini-code-assist]

The ternary logic here is redundant. Since allSelected is only true when all rows are shared, and in all other cases (none selected or some selected) the selectAll state should be false, you can simplify this to just setSelectAll(allSelected).

Suggested change

	setSelectAll(allSelected ? true : noneSelected ? false : false)
	setSelectAll(allSelected)