Distance Metrics

Distance metrics define how vectors are compared for similarity. Choosing the right metric is critical for accurate search results and depends on your embedding model.

Available Distance Metrics

Qdrant supports four distance metrics:

// From lib/segment/src/types.rs:306-315
pub enum Distance {
    Cosine,     // Cosine similarity
    Euclid,     // Euclidean distance  
    Dot,        // Dot product
    Manhattan,  // Manhattan distance
}

Cosine Similarity

Measures the angle between vectors, ignoring magnitude:

from qdrant_client import QdrantClient, models

client = QdrantClient("localhost", port=6333)

client.create_collection(
    collection_name="semantic_search",
    vectors_config=models.VectorParams(
        size=384,
        distance=models.Distance.COSINE
    )
)

Formula:

cosine_similarity(A, B) = (A · B) / (||A|| * ||B||)

Score Range: -1 (opposite) to 1 (identical) Characteristics:

Normalized: Vector magnitude doesn’t affect similarity
Order: Higher scores are better (maximization)
Symmetric: cosine(A, B) = cosine(B, A)

Cosine is the most common metric for semantic search and text embeddings. Most embedding models (OpenAI, Cohere, Sentence Transformers) are optimized for cosine similarity.

When to Use Cosine

✅ Use Cosine when:

Working with text embeddings (BERT, GPT, etc.)
Your embedding model outputs normalized vectors
You care about direction, not magnitude
Using pre-trained models (OpenAI, Cohere, etc.)

❌ Don’t use Cosine when:

Vector magnitude contains important information
Your model was specifically trained for Euclidean distance

Example: Text Similarity

# Vectors with same direction but different magnitudes are similar
vec_a = [1.0, 2.0, 3.0]     # magnitude = 3.74
vec_b = [2.0, 4.0, 6.0]     # magnitude = 7.48

# Cosine similarity = 1.0 (identical direction)

Euclidean Distance (L2)

Measures straight-line distance between vectors:

client.create_collection(
    collection_name="image_search",
    vectors_config=models.VectorParams(
        size=512,
        distance=models.Distance.EUCLID
    )
)

Formula:

euclidean_distance(A, B) = sqrt(Σ(Ai - Bi)²)

Score Range: 0 (identical) to ∞ (very different) Characteristics:

Magnitude-sensitive: Vector length affects distance
Order: Lower scores are better (minimization)
Symmetric: euclid(A, B) = euclid(B, A)

Euclidean distance considers both direction and magnitude. It’s commonly used for image embeddings and when vector magnitude is meaningful.

When to Use Euclidean

✅ Use Euclidean when:

Vector magnitude is meaningful
Working with image embeddings (ResNet, EfficientNet)
Your model was trained with Euclidean distance
You need geometric distance in embedding space

❌ Don’t use Euclidean when:

Vectors have inconsistent magnitudes
You only care about direction

Example: Spatial Distance

# Vectors with similar direction but different magnitudes are different
vec_a = [1.0, 2.0, 3.0]
vec_b = [2.0, 4.0, 6.0]

# Euclidean distance = 3.74 (significant difference)

Dot Product

Computes the product of corresponding elements:

client.create_collection(
    collection_name="recommendation",
    vectors_config=models.VectorParams(
        size=256,
        distance=models.Distance.DOT
    )
)

Formula:

dot_product(A, B) = Σ(Ai * Bi)

Score Range: -∞ to ∞ Characteristics:

Unnormalized: Magnitude matters significantly
Order: Higher scores are better (maximization)
Symmetric: dot(A, B) = dot(B, A)
Efficient: Fastest to compute

Dot product is sensitive to vector magnitude. Longer vectors will have higher scores even if less similar in direction.

When to Use Dot Product

✅ Use Dot Product when:

Your vectors are pre-normalized
Working with binary or categorical features
Your model was specifically trained for dot product
Maximum computational efficiency is needed

❌ Don’t use Dot Product when:

Vectors have inconsistent magnitudes
You need normalized similarity scores

Relationship with Cosine

For normalized vectors (||v|| = 1), dot product equals cosine similarity:

# If vectors are normalized
import numpy as np

vec_a = [1.0, 0.0, 0.0]
vec_b = [0.707, 0.707, 0.0]

dot_product = np.dot(vec_a, vec_b)  # 0.707
cosine_sim = np.dot(vec_a, vec_b) / (np.linalg.norm(vec_a) * np.linalg.norm(vec_b))  # 0.707

# They're identical for normalized vectors!

If your model outputs normalized vectors, use Dot Product instead of Cosine for better performance.

Manhattan Distance (L1)

Sum of absolute differences:

client.create_collection(
    collection_name="specialized",
    vectors_config=models.VectorParams(
        size=128,
        distance=models.Distance.MANHATTAN
    )
)

Formula:

manhattan_distance(A, B) = Σ|Ai - Bi|

Score Range: 0 (identical) to ∞ (very different) Characteristics:

Axis-aligned: Measures distance along axes
Order: Lower scores are better (minimization)
Symmetric: manhattan(A, B) = manhattan(B, A)
Robust: Less sensitive to outliers than Euclidean

When to Use Manhattan

✅ Use Manhattan when:

Your features are independent/axis-aligned
You want robustness to outliers
Working with grid-like data
Your model was trained with Manhattan distance

❌ Don’t use Manhattan when:

Most embedding models (rarely optimized for L1)
You need standard semantic similarity

Manhattan is less common in vector search but useful for specialized applications where features are independent.

Score Ordering

// From lib/segment/src/types.rs:342-347
pub fn distance_order(&self) -> Order {
    match self {
        Distance::Cosine | Distance::Dot => Order::LargeBetter,
        Distance::Euclid | Distance::Manhattan => Order::SmallBetter,
    }
}

Metric	Better Score	Range
Cosine	Higher (maximize)	-1 to 1
Euclidean	Lower (minimize)	0 to ∞
Dot Product	Higher (maximize)	-∞ to ∞
Manhattan	Lower (minimize)	0 to ∞

Qdrant automatically handles score ordering. Top results always represent most similar vectors regardless of metric.

Common Embedding Models

Match your distance metric to your model:

Model	Dimensions	Distance
OpenAI text-embedding-3-small	1536	Cosine
OpenAI text-embedding-3-large	3072	Cosine
Cohere embed-english-v3.0	1024	Cosine
Cohere embed-multilingual-v3.0	1024	Cosine
Sentence Transformers (all-MiniLM-L6-v2)	384	Cosine
Sentence Transformers (all-mpnet-base-v2)	768	Cosine
CLIP ViT-B/32	512	Cosine
ResNet-50	2048	Euclidean
BGE-large-en-v1.5	1024	Cosine
E5-large-v2	1024	Cosine

Always check your model’s documentation for the recommended distance metric.

Performance Comparison

Metric	Computation Speed	Memory	Use Case
Dot Product	Fastest	Lowest	Pre-normalized vectors
Cosine	Fast	Medium	General semantic search
Euclidean	Medium	Medium	Image embeddings
Manhattan	Medium	Medium	Specialized applications

Threshold Filtering

Set minimum similarity thresholds:

# Cosine: only return results with similarity > 0.7
results = client.search(
    collection_name="my_collection",
    query_vector=[0.1, 0.2, ...],
    score_threshold=0.7,  # For Cosine: higher is better
    limit=10
)

# Euclidean: only return results with distance < 5.0
results = client.search(
    collection_name="image_collection",
    query_vector=[0.1, 0.2, ...],
    score_threshold=5.0,  # For Euclidean: lower is better
    limit=10
)

// From lib/segment/src/types.rs:357-361
pub fn check_threshold(&self, score: ScoreType, threshold: ScoreType) -> bool {
    match self.distance_order() {
        Order::LargeBetter => score > threshold,
        Order::SmallBetter => score < threshold,
    }
}

Threshold direction depends on the metric:

Cosine/Dot: Use high thresholds (e.g., 0.7) to filter for similar items
Euclidean/Manhattan: Use low thresholds (e.g., 5.0) to filter for similar items

Choosing the Right Metric

Decision Flow

Check your model’s documentation - Use the recommended metric
Are vectors normalized?
- Yes → Use Dot Product (fastest) or Cosine (more intuitive)
- No → Continue to step 3
Does magnitude matter?
- No → Use Cosine
- Yes → Use Euclidean
Special requirements?
- Need outlier robustness → Manhattan
- Maximum speed → Dot Product (with normalization)

Quick Reference

For text embeddings (default)

Use Cosine. Nearly all text embedding models are optimized for cosine similarity.

For image embeddings

Use Euclidean or Cosine. Check your model’s documentation. ResNets often use Euclidean, CLIP uses Cosine.

For maximum speed

Use Dot Product if your vectors are pre-normalized. It’s computationally cheaper than Cosine.

For normalized vectors

Use Dot Product. It’s equivalent to Cosine but faster.

Vector Normalization

Normalize vectors for consistent Cosine/Dot Product behavior:

import numpy as np

def normalize_vector(vector):
    """Normalize vector to unit length"""
    norm = np.linalg.norm(vector)
    if norm == 0:
        return vector
    return vector / norm

# Example
vector = [3.0, 4.0, 0.0]
normalized = normalize_vector(vector)
print(normalized)  # [0.6, 0.8, 0.0]
print(np.linalg.norm(normalized))  # 1.0

# Batch normalization
vectors = np.array([
    [3.0, 4.0, 0.0],
    [1.0, 0.0, 0.0],
    [0.5, 0.5, 0.707]
])

# Normalize along each row
norms = np.linalg.norm(vectors, axis=1, keepdims=True)
normalized = vectors / norms

Many embedding models (OpenAI, Cohere) return pre-normalized vectors. Check before normalizing!

Best Practices

Always use your model's recommended metric

Embedding models are trained with specific metrics. Using a different metric will degrade results.

Normalize vectors when using Dot Product

Dot Product is sensitive to magnitude. Normalize vectors first unless magnitude is meaningful.

Test with your actual data

Benchmark different metrics with your queries to find what works best in practice.

Consider Dot Product for speed

If using Cosine with normalized vectors, switch to Dot Product for better performance.

Be consistent

Use the same metric for indexing and querying. Mixing metrics will produce incorrect results.

Collections

Learn how to configure distance metrics in collections

Vectors

Understand vector types and preprocessing

Indexing

Explore how HNSW indexes work with distance metrics

Points

Learn how points store vectors for comparison

Getting Started

Core Concepts

Working with Data

Advanced Features

Deployment

Operations

Distance Metrics

Available Distance Metrics

Cosine Similarity

When to Use Cosine

Example: Text Similarity

Euclidean Distance (L2)

When to Use Euclidean

Example: Spatial Distance

Dot Product

When to Use Dot Product

Relationship with Cosine

Manhattan Distance (L1)

When to Use Manhattan

Score Ordering

Common Embedding Models

Performance Comparison

Threshold Filtering

Choosing the Right Metric

Decision Flow

Quick Reference

Vector Normalization

Best Practices

Collections

Vectors

Indexing

Points

Getting Started

Core Concepts

Working with Data

Advanced Features

Deployment

Operations

​Available Distance Metrics

​Cosine Similarity

​When to Use Cosine

​Example: Text Similarity

​Euclidean Distance (L2)

​When to Use Euclidean

​Example: Spatial Distance

​Dot Product

​When to Use Dot Product

​Relationship with Cosine

​Manhattan Distance (L1)

​When to Use Manhattan

​Score Ordering

​Common Embedding Models

​Performance Comparison

​Threshold Filtering

​Choosing the Right Metric

​Decision Flow

​Quick Reference

​Vector Normalization

​Best Practices

​Related Concepts

Collections

Vectors

Indexing

Points

Available Distance Metrics

Cosine Similarity

When to Use Cosine

Example: Text Similarity

Euclidean Distance (L2)

When to Use Euclidean

Example: Spatial Distance

Dot Product

When to Use Dot Product

Relationship with Cosine

Manhattan Distance (L1)

When to Use Manhattan

Score Ordering

Common Embedding Models

Performance Comparison

Threshold Filtering

Choosing the Right Metric

Decision Flow

Quick Reference

Vector Normalization

Best Practices

Related Concepts