Matryoshka Embeddings Example

Attention: The code in this repository is intended for experimental use only and is not fully tested, documented, or supported by SingleStore. Visit the SingleStore Forums to ask questions about this repository.

Demonstrates using Matryoshka-capable embeddings with SingleStore. Shows how a single embedding can be truncated to different dimensions and compares search results at each dimensionality.

Quick Start

# Install dependencies
pip install -r requirements.txt

# Configure database connection
cp config.json.example config.json
# Edit config.json with your SingleStore connection details

# Run the demo
python matryoshka_demo.py

What This Demonstrates

• Generate ONE 768d embedding per document
• Truncate to create 64d, 128d, 256d, 512d, 768d prefixes
• Store each prefix in SingleStore (separate columns)
• Compare search results across dimensions
• Measure overlap with full 768d baseline

Note: This example demonstrates an example scenario with a Matryoshka-trained model on a small literary dataset. Real-world results will vary.

Example Output

Query: "Why does Elizabeth initially dislike Mr. Darcy?"

64d:  overlap 40%
256d: overlap 80%
512d: overlap 100%
768d: overlap 100% (baseline)

How It Works

# 1. Generate ONE full embedding
embedding = model.encode(f"search_document: {text}", ...)

# 2. Truncate to different sizes
embedding_256 = embedding[:256]
embedding_768 = embedding[:768]

# 3. Normalize after truncation
embedding_256 = embedding_256 / np.linalg.norm(embedding_256)

# 4. Store and search at different dimensions

Files

• matryoshka_demo.py - Working demo with Nomic Embed
• pride_and_prejudice.txt - Sample data
• requirements.txt - Dependencies
• config.json.example - Database configuration template

References

• Matryoshka Representation Learning Paper
• Nomic Embed Documentation