​

Embeddings

Embeddings are dense vector representations of content that enable semantic search and similarity matching. Mixpeek provides state-of-the-art embedding models for different types of content.

​

Available Models

​

Text Embeddings

• General Purpose: Optimized for broad text understanding
• Domain-Specific: Specialized for particular industries
• Multilingual: Support for multiple languages
• Cross-Modal: Text-to-image alignment

​

Image Embeddings

• Visual Features: Capture visual characteristics
• Object-Based: Focus on object recognition
• Scene Understanding: Capture scene context
• Style-Based: Represent artistic styles

​

Video Embeddings

• Temporal Features: Capture motion and changes
• Frame-Level: Process individual frames
• Scene-Level: Understand scene composition
• Action-Based: Represent activities

​

Keyword Embeddings

• SPLADE: Sparse lexical model

​

Model Selection

​

Factors to Consider

1. Use Case

   • Search requirements
   • Similarity matching
   • Classification needs
   • Cross-modal applications

2. Performance

   • Accuracy
   • Speed
   • Resource usage
   • Latency requirements

3. Data Characteristics

   • Content type
   • Language
   • Domain
   • Scale

​

Available Models

Model ID	Modality	Dimensions	Description
text	Text	768	General purpose text embedding, multilingual
image	Image	1024	Visual feature embedding
video	Video	1536	Temporal feature embedding
multimodal	Multiple	1024	Cross-modal embedding
keyword	Text	Variable	SPLADE sparse lexical model

​

Best Practices

​

Preprocessing

• Text Normalization: Clean and standardize text
• Image Resizing: Standardize dimensions
• Quality Control: Filter low-quality inputs
• Format Conversion: Ensure compatible formats

​

Optimization

• Batch Processing: Generate embeddings in batches
• Caching: Store frequently used embeddings
• Dimension Reduction: Use when appropriate
• Quality Thresholds: Filter poor embeddings