Cohere has introduced Embed 5, a new embedding model that provides two distinct tiers: pro and fast. Both tiers operate within a shared embedding space, allowing teams to select the appropriate speed and performance level for their specific use cases without needing to manage separate models or incompatible vector dimensions. This structure supports flexibility in deployment, enabling organizations to balance computational resources against the need for high-quality embeddings across different applications.
Alongside the model launch, Cohere has released a new enterprise retrieval evaluation method. This approach shifts the focus away from generic search averages, which often fail to capture the nuances of enterprise data retrieval. The new metric emphasizes rank consistency, providing a more accurate picture of how well a system retrieves relevant information. By prioritizing rank consistency, the method aims to reflect the actual ordering of results that users encounter, rather than relying on broad, averaged scores that may obscure performance issues in specific queries.
The company advises teams to measure evidence recall directly on their own documents. This recommendation suggests that relying on standard, generic benchmarks may not accurately reflect performance in real-world scenarios. By testing on proprietary data, organizations can better understand how their retrieval systems perform with the specific content they handle. This targeted evaluation helps ensure that the chosen embedding model and retrieval configuration effectively meet the needs of the business, rather than just scoring well on abstract, industry-wide averages.
Source: https://cohere.com/
Your membership also unlocks:
