Dify is enhancing its knowledge base retrieval system. We're phasing out the N-to-1 retrieval strategy on September 1, 2024, and introducing a more flexible multi-path retrieval strategy. We recommend switching to this new approach to boost your application's retrieval efficiency.
Why retire "N-to-1 retrieval"?
Our analysis has uncovered key limitations in the N-to-1 retrieval strategy. This approach restricts searches to a single knowledge base and relies heavily on LLM interpretation of knowledge base descriptions. As a result, it often produces incomplete or inaccurate results, compromising retrieval quality. Feedback from our community supports these findings, driving our decision to move towards a more effective solution.
A Better Solution: Configurable "Multi-path Retrieval"
Our enhanced multi-path retrieval strategy offers:
Optional reranking strategies
Semantic and keyword weighting for optimized retrieval
Integration with reranking models (e.g., Cohere, Jina) for peak performance
We recommend using this new setup for more accurate retrieval.
What you need to do
Dify Cloud Users: Switch from "N-to-1 retrieval" to "multi-path retrieval" in Context > Retrieval Setting. Don't worry if you miss it - we'll automatically update to default multi-path settings on September 1, 2024.
Community and Enterprise Users: Keep an eye out for our post-September 1 release notes. We'll provide migration scripts and details there.
Optimizing "Multi-path Retrieval" with Rerank
Multi-path retrieval in Dify offers two primary configuration options: Keyword & Semantic Weighted Score and Rerank Model selection.
Keyword & Semantic Weighted Score Configuration
Keyword-only (weight: 1): Best for exact matches. It's fast and efficient, especially for large knowledge bases. Use this when your users know precisely what they're looking for.
Semantic-only (weight: 1): Understands the meaning behind queries. It can find relevant info even without exact keyword matches. Great for multilingual content and complex searches.
Custom weight balance: Blend keyword and semantic approaches to fit your needs. Adjust the mix to match your unique business requirements or complex information structure.
Rerank Model
For maximum retrieval precision, we recommend implementing a rerank model. This refines the initial results, significantly enhancing overall accuracy.
For detailed configuration steps and best practices, please refer to our documentation.
Looking ahead
This upgrade kickstarts our journey to enhance Dify's RAG capabilities. We're dedicated to refining our RAG system, prioritizing flexibility and openness to cater to our diverse community and customer needs.
Your insights are crucial as we grow. Join our community and help us shape the future of Dify.