ai/embeddinggemma repository overview

⁠Embedding Gemma

Embedding Gemma is a state-of-the-art text embedding model from Google DeepMind, designed to create high-quality vector representations of text. Built on the Gemma architecture, this model converts text into dense vector embeddings that capture semantic meaning, making it ideal for retrieval-augmented generation (RAG), semantic search, and similarity tasks. With open weights and efficient design, Embedding Gemma provides a powerful foundation for embedding-based applications. The GGUF format version is provided by Unsloth.

⁠Intended uses

Embedding Gemma is designed for applications requiring high-quality text embeddings:

• Semantic search and retrieval: Excellent for building search systems, document retrieval, and RAG applications that need to find semantically relevant content.
• Text similarity and clustering: Generate embeddings for measuring text similarity, document clustering, and content deduplication tasks.
• Classification and downstream tasks: Use embeddings as input features for various NLP classification tasks and machine learning pipelines.

⁠Characteristics

⁠Available model variants

¹: VRAM estimated based on model characteristics.

⁠Use this AI model with Docker Model Runner

First, pull the model:

docker model pull ai/embeddinggemma

Then run the model:

docker model run ai/embeddinggemma

To generate embeddings using the API:

curl --location 'http://localhost:12434/engines/llama.cpp/v1/embeddings' \
--header 'Content-Type: application/json' \
--data '{
    "model": "ai/embeddinggemma",
    "input": "Your text to embed here"
  }'

For more information on Docker Model Runner, explore the documentation⁠.

⁠Considerations

• Context length: The model supports up to 2K tokens. Longer texts may need to be chunked for optimal performance.
• Language support: Primarily trained on English text, performance on other languages may vary.
• Embedding dimension: The model produces 768-dimensional embeddings suitable for most downstream tasks.
• Normalization: Embeddings are normalized by default, making them suitable for cosine similarity calculations.

⁠Benchmark performance

⁠Links

• Embedding Gemma Model Card⁠
• Unsloth GGUF Version⁠
• Gemma Model Family⁠
• Gemma Terms of Use⁠