graphrag-system-design_skill

Overview

This skill designs complete GraphRAG systems that combine graph retrieval with LLM reasoning. It advises on pattern selection, technology stack, integration pipelines, and domain customizations. Use it to create production-ready architectures that preserve relational context, enable multi-hop reasoning, and provide provenance for generated answers.

How this skill works

Given a domain and retrieval requirements, the skill analyzes query complexity, data characteristics, and compliance constraints to recommend a retrieval pattern (Hybrid Symbol-Vector, Subgraph-on-Demand, or Community-Based Global Summarization). It maps those patterns to a concrete tech stack for graph, vector, orchestration, and LLM layers, then designs an ingestion-to-generation pipeline with synchronization, routing, and provenance. Finally, it produces a detailed specification including deployment, monitoring, and performance targets.

When to use it

• You need multi-hop or relationship-aware retrieval beyond flat vector RAG.
• You must combine precise structured queries with fuzzy semantic search (entity disambiguation).
• You require explainability and provenance for LLM outputs in regulated domains.
• You plan to deploy a production system with scaling, monitoring, and update requirements.
• You need domain-specific ontology integration (healthcare, finance, legal).

Best practices

• Start by characterizing query patterns, latency SLAs, and compliance constraints before choosing a pattern.
• Prefer a proof-of-concept that validates retrieval quality and latency before full-scale build.
• Design clear synchronization between graph and vector indices and choose an update strategy (batch vs streaming).
• Use a query router to combine patterns when requirements vary across query types.
• Include provenance tracking in the pipeline so every generated claim links to graph paths or source docs.

Example use cases

• Enterprise QA that needs entity-aware answers and source chains for auditability.
• Clinical decision support using a layered patient graph plus semantic retrieval of literature.
• Financial analytics with temporal graphs and hybrid retrieval for regulatory reporting.
• Customer support knowledge base that assembles subgraphs on demand for complex tickets.