Embedding Dimensionality

Overview

Embedding dimensionality refers to the number of dimensions in a vector embedding, directly impacting storage, performance, and accuracy.

Common Dimensions

Small (384-512)

• Examples: MiniLM models
• Pros: Fast, low memory
• Cons: Less information capacity
• Use: Resource-constrained, high-throughput

Medium (768-1024)

• Examples: BERT-base, many SBERT models
• Pros: Good balance
• Cons: Moderate resource needs
• Use: General purpose

Large (1536-2048)

• Examples: text-embedding-3-large, BGE-large
• Pros: Rich representations
• Cons: Higher costs
• Use: Quality-critical applications

Very Large (3072-4096)

• Examples: Gemini Embedding 2 (3072)
• Pros: Maximum capacity
• Cons: Significant resource requirements
• Use: Specialized, high-accuracy needs

Matryoshka Embeddings

Variable-size embeddings:

• One model, multiple dimensions
• Truncate without retraining
• Example: Gemini Embedding 2 (3072/1536/768)

Tradeoffs

Higher Dimensions

Advantages:

• More information
• Better accuracy
• Richer semantics

Disadvantages:

• More storage
• Slower search
• Higher memory
• Increased costs

Selection Criteria

• Task complexity
• Performance requirements
• Budget constraints
• Scale of deployment

Pricing

Dimensionality itself is a model property, affecting infrastructure costs.