Deepset and Mixedbread have taken an innovative leap by introducing a revolutionary open-source German/English embedding model called deepset-mxbai-embed-de-large-v1. The tool aims to correct the imbalance in the AI landscape, where English-speaking markets dominate. Based on the intfloat/multilingual-e5-large model, it is fine-tuned using over 30 million pairs of German data to enhance natural language processing (NLP) capabilities.
Particularly designed for retrieval tasks, this model sets new performance benchmarks for open-source German embedding models. The measurement of its effectiveness relies on NDCG@10, a metric employed for assessing the precision of ranking results against a list arranged in the ideal order. The deepset-mxbai-embed-de-large-v1 model exhibited an average performance of 51.7 on this metric, surpassing other models such as multilingual-e5-large and jina-embeddings-v2-base-de.
The focus of the development process is on optimizing storage and inference efficiency. The developers use two innovative techniques for this purpose — Matryoshka Representation Learning (MRL) and Binary Quantization.
MRL lowers the output dimensions count of the embedding model without any significant loss in accuracy. This method modifies the loss function to emphasize crucial data in the initial dimensions, enabling the truncation of the later dimensions to boost efficiency.
Binary Quantization, on the other hand, converts float32 values into binary values that drastically reduce memory and disk space usage. This technique can maintain high performance during inference, making the model powerfully resource-efficient.
For its practical application, users can easily incorporate the deepset-mxbai-embed-de-large-v1 model with the Haystack framework through SentenceTransformersDocumentEmbedder and SentenceTransformersTextEmbedder components. Mixedbread allows for smooth integration via MixedbreadDocumentEmbedder and MixedbreadTextEmbedder. To use this model with Haystack’s Sentence Transformers Embedders, users need to install the ‘mixedbread-ai-haystack’ and export their Mixedbread API key to ‘MXBAI_API_KEY’.
In conclusion, Deepset and Mixedbread have built upon the success of the German BERT model and believe that their new cutting-edge embedding model will enable the German-speaking AI community to create innovative products, particularly in areas like retrieval-augmented generation (RAG). This product could serve as a valuable tool for developers and researchers working on German language tasks.
