Query Decomposition

소개

The Query Decomposition skill enhances AI retrieval systems by breaking down compound or multi-hop questions into distinct, manageable concepts. By identifying independent topics within a single query, it allows for parallelized retrieval across vector databases and subsequent result fusion using Reciprocal Rank Fusion (RRF). This approach solves the common 'missing concept' problem in RAG, where a single-pass search might overlook documents related to secondary topics in a complex question, making it essential for building robust, production-ready AI search experiences.

주요 기능

• Compatibility with HyDE and reranking patterns for maximum precision
• Parallel retrieval architecture to maintain low-latency performance
• Heuristic detection for sub-millisecond identification of complex queries
• LLM-powered decomposition into 2-4 independent searchable sub-queries
• 29 GitHub stars
• Implementation of Reciprocal Rank Fusion (RRF) for robust result blending

사용 사례

• Handling multi-hop questions that require information from disparate documents
• Improving recall for comparative queries like 'Technology X vs Technology Y'
• Optimizing RAG pipelines for complex technical documentation and troubleshooting