retrieval-search-orchestration_skill

Overview

This skill helps design retrieval strategies for querying knowledge graphs in Retrieval-Augmented Generation (RAG) systems. It prescribes pattern selection, query decomposition, ranking and constraint configuration, and provenance design to make answers traceable and accurate. Use it to produce a concise retrieval strategy specification tailored to your query mix and KG topology.

How this skill works

It inspects the query distribution, KG type, and operational constraints to recommend a primary retrieval pattern (global-first, local-first, U-shaped, etc.). It provides decomposition patterns for multi-hop or composite queries, ranking and filtering rules, fallback strategies, and a provenance annotation scheme so every returned fact links to source evidence. The output is a structured retrieval strategy spec you can apply to orchestration code or evaluation pipelines.

When to use it

• Designing or auditing retrieval strategies for GraphRAG systems
• Handling multi-hop or composite queries that need query decomposition
• Choosing between global-first, local-first, or hybrid search patterns
• Implementing provenance, citation, and evidence-tracing for generated answers
• Defining fallback and graceful degradation policies for sparse results

Best practices

• Start by profiling the query distribution and dominant query types before picking a pattern
• Prefer local-first for entity-centric lookups and global-first for thematic summarization
• Use U-shaped hybrid when queries mix breadth and depth, accept higher latency for coverage
• Apply LLM-as-controller or self-ask chains for multi-hop tasks and limit sub-query depth to control cost
• Annotate retrieved evidence with source IDs, timestamps, confidence and support chains for reproducible citations

Example use cases

• Biomedical RAG: find drugs linked to gene mutations using query decomposition and provenance
• Enterprise search: combine global topic discovery with local entity neighborhoods for customer insights
• Legal research: temporal and constraint-guided retrieval for case law filtered by jurisdiction and date
• Knowledge graph QA: ReAct-style interleaving to verify multi-hop paths with supporting documents
• Hybrid pipelines: SPARQL/Cypher tool-augmented queries for structured traversal and LLM synthesis