Cohere’s faster query model barely dents retrieval quality in its tests

Cohere introduced Embed 5 on Wednesday, offering teams the choice to index data using Embed 5 Pro and query it with Embed 5 Fast, a faster and more cost-effective option that shares the same embedding space. While the faster query model may slightly compromise retrieval quality, Cohere's testing indicates the impact is minimal. Across 40 datasets encompassing text, images, fused documents, and parsed documents, Embed 5 Fast achieved a retrieval score of 96.6 when used for both indexing and queries, compared to 98.4 when Pro was used for indexing and Fast for queries.

Embed 5 supports six vector dimensions ranging from 256 to 2,048, along with float32, int8, and binary formats, with int8 providing a good balance between reduced memory and storage needs and nearly full-precision retrieval quality. The model can handle text, images, and fused text-image inputs across over 100 languages, with Cohere reporting an average Pro score of 82.3 for its fused text-image evaluation, slightly higher than Fast's 81.2 and significantly higher than Google's Gemini Embedding 2 at 81.2.

For parsed-PDF evaluation, Embed 5 Pro scored 84.8, while Fast came in at 83.4, outperforming Google's Gemini Embedding 2 at 80.8. In ViDoRe V3, Pro averaged 85.8, with Fast at 84.5, surpassing Voyage 4 Large at 83.6, Gemini Embedding 2 at 80.8, and Cohere's previous Embed 4 at 77. While Cohere's Embed 5 Pro and Fast models are accessible through various platforms, production RAG and agent systems should still conduct their own benchmark tests to ensure optimal performance on their specific corpus and query distribution.