What is query decomposition in RAG?

Query decomposition is the process of breaking a complex, multi-part user question into smaller, independent concepts that can be searched separately to improve document recall.

How does this skill improve search accuracy?

By searching for individual concepts in parallel, the system captures relevant documents that might not contain all the keywords of a complex query at once, preventing 'misses' in the retrieval phase.

What is Reciprocal Rank Fusion (RRF)?

RRF is a parameter-free algorithm used to combine multiple ranked result lists into a single, optimized ranking, ensuring the most relevant documents across all sub-queries are prioritized.

Does this skill add significant latency?

The skill includes a fast-path heuristic check (<1ms) to skip decomposition for simple queries and uses parallel execution for retrieval to keep total latency minimal.

Query Decomposition

Name: Query Decomposition
Author: yonatangross

byyonatangross

•

データサイエンスとML

Breaks down complex search queries into independent sub-concepts to improve retrieval accuracy and coverage in RAG systems.

The Query Decomposition skill provides a robust framework for handling multi-concept and multi-hop queries within RAG (Retrieval-Augmented Generation) pipelines. It employs a two-tier strategy—using fast heuristic detection for simple cases and LLM-driven analysis for complex ones—to split compound questions into searchable fragments. By executing parallel retrievals and merging results through Reciprocal Rank Fusion (RRF), this skill ensures that Claude can find relevant documentation even when information is scattered across disparate sources, significantly improving the recall and precision of AI-driven search applications.

主な機能

01Heuristic-based fast-path detection for multi-concept queries

02Seamless integration with HyDE (Hypothetical Document Embeddings)

0369 GitHub stars

04LLM-driven concept extraction for complex multi-hop reasoning

05Parallel retrieval execution for optimized latency

06Reciprocal Rank Fusion (RRF) for robust result blending

ユースケース

01Enhancing RAG pipelines for complex, multi-topic user requests

02Improving recall for multi-hop research questions in large datasets

03Resolving 'X vs Y' comparison queries in technical documentation

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit query-decomposition

For use in Claude.ai and ChatGPT

Download Skill