knowledge-graphlisted

**CONSUMES**: User knowledge-graph / GraphRAG task + corpus description + optional existing graph schema, extraction configs, or DB choice **PRODUCES**: Applied graph judgment rules + GraphRAG architecture recommendation + extraction prompting plan + entity-resolution pipeline + graph-DB selection + Text2Cypher/SPARQL-Star query patterns + cost guardrails # Knowledge Graph & GraphRAG Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 **License**: Apache 2.0 --- ## What This Pack Does AI agents reach for GraphRAG the moment "knowledge graph" appears, then default to full Microsoft GraphRAG — paying complete LLM pre-summarization cost when LazyGraphRAG would match global-search quality at a >700x lower query cost (indexing = 0.1% of full GraphRAG). They ask "which graph database do we store it in?" — not knowing that default Microsoft GraphRAG writes **Parquet + an embedded LanceDB store and needs no graph DB at all**, so the question is a category error. They pick Global search for every query, ignoring that it runs a parallel Map-Reduce over entire community-report levels. They extract entities with a bare zero-shot prompt and skip entity resolution, splitting "Olympic Winter Games" and "winter Olympic games" into two nodes that fragment every traversal. They copy a "magic" 0.9 merge threshold no record-linkage tool actually publishes. And when a graph DB IS warranted, they pick Neo4j by reflex without checking