Cohere Launches Rerank 3: A High-Precision Model for Enterprise Search and Retrieval Augmented Generation

Today, Cohere is excited to introduce Rerank 3, a new foundation model designed to significantly enhance enterprise search and Retrieval Augmented Generation (RAG) systems. This model is compatible with any database or search index and can be integrated into existing applications with minimal effort-just a single line of code.

Key Technical Improvements

• 4k Context Length: Rerank 3 supports up to 4,000 tokens in context, which is crucial for handling longer documents and improving overall search quality.
• Multi-Aspect Data Support: It excels at searching through complex, semi-structured data like emails, invoices, JSON documents, code, and tables.
• Multilingual Coverage: The model supports over 100 languages, making it a powerful tool for global organizations with multilingual data sources.
• Improved Latency and Lower TCO: Rerank 3 is optimized to reduce latency and lower the total cost of ownership (TCO) without compromising on performance.

How It Works

Rerank 3 combines high-precision semantic reranking with generative models to ensure that only the most relevant information is fed into the generative model. This approach enhances response accuracy while keeping latency low, especially when dealing with large datasets.

Enhanced Enterprise Search

Enterprise data often includes complex and semi-structured formats, which traditional search systems struggle to handle effectively. Rerank 3 addresses this by:

• Ranking Complex Data: It can rank multi-aspect data (e.g., emails) based on all relevant metadata fields, including recency.
• Semi-Structured Retrieval Accuracy: Demonstrates significant improvements in semi-structured retrieval accuracy, as measured by Recall@5 on datasets like TMDB-5k-Movies, WikiSQL, nq-tables, and Cohere's annotated datasets.

Code Retrieval Capabilities

Rerank 3 also excels in code retrieval, which can be particularly useful for engineering teams. It supports:

• Proprietary Code Repositories: Enhances productivity by retrieving relevant code snippets from an enterprise’s proprietary repositories.
• Large Documentation Corpuses: Improves the efficiency of searching through extensive documentation.

Code evaluation accuracy based on nDCG@10 on datasets like Codesearchnet, Stackoverflow, CosQA, Human Eval, MBPP, and DS1000 shows marked improvements.

Multilingual Performance

For global organizations, multilingual data retrieval has historically been a challenge. Rerank 3 addresses this by:

• Strong Multilingual Support: Offers robust performance in over 100 languages.
• Simplified Retrieval: Makes it easier for non-English speaking customers to access and utilize enterprise data.

Multilingual retrieval accuracy based on nDCG@10 on the MIRACL dataset highlights its strong multilingual capabilities.

Implementation Details

To integrate Rerank 3 into your existing systems, you can use a single line of code. This makes it easy to boost search performance or reduce the cost of running RAG applications with minimal impact on latency. The model is designed to be compatible with any database or search index, ensuring seamless integration.

Benchmarks and Performance

• Semi-Structured Retrieval Accuracy: Higher Recall@5 scores on datasets like TMDB-5k-Movies, WikiSQL, nq-tables, and Cohere's annotated datasets.
• Code Evaluation Accuracy: Improved nDCG@10 scores on Codesearchnet, Stackoverflow, CosQA, Human Eval, MBPP, and DS1000.
• Multilingual Retrieval Accuracy: Stronger nDCG@10 scores on the MIRACL dataset.

Conclusion

Rerank 3 is a significant step forward in enterprise search and RAG systems. Its ability to handle complex, multi-aspect data, support for over 100 languages, and improved latency make it a valuable addition to any organization's toolkit. Whether you're dealing with large code repositories, extensive documentation, or multilingual data sources, Rerank 3 is designed to enhance your search capabilities and reduce costs.