GraphRAG & KG+LLM Systems
KG + LLM Architectures
11 min+105 XPStep 1/11
Three depths of coupling
- Graph as retrieval index — the LLM never sees the schema. Your pipeline does entity linking + traversal + chunk fetch, then hands the model curated context. Lowest risk; bounded by what your pipeline anticipated.
- Text-to-query — the model writes Cypher/SPARQL from natural language. Arbitrary questions become possible; so do arbitrary bad queries. Needs schema-in-prompt, validation, and a repair loop.
- Graph tools — the model calls
find_entity(name),neighbors(id, rel_type),shortest_path(a, b)and composes results. Middle path: flexible like queries, bounded like tools — every call legible and loggable.