In brief

What this examines

Graph retrieval-augmented generation addresses a specific failure of ordinary vector search: questions whose answers are distributed across entities, events, and documents rather than located in one passage. The decision to use it should follow the shape of the questions, not the novelty of the technology.

Why it matters

Vector retrieval remains the right default for passage-shaped questions. A graph earns its cost when the system must traverse relationships, compare state over time, resolve identity across records, or summarize patterns across an entire corpus.

Key ideas

  • Use vector retrieval when the answer lives in a small number of relevant passages.
  • Use graph retrieval for multi-hop, corpus-wide, temporal, and conflict-aware questions.
  • Entity resolution is the production gate: a bad identity graph poisons every downstream answer.
  • Provenance and authorization must survive indexing and retrieval.
  • Pilot against real questions and a representative corpus before committing to the architecture.

The question vector search cannot answer

Nearest-neighbour retrieval returns passages that resemble the query. It does not inherently preserve the chain connecting a facility to a component, a component to a part number, and that part number to a supplier. Multi-hop and corpus-wide questions need structure that survives retrieval.

How graph retrieval works

A graph pipeline extracts entities and relationships, resolves identities, attaches provenance, and builds communities or neighbourhoods that can be searched locally or summarized globally. The graph is compiled evidence, not a decorative database added beside the vector index.

Where it breaks

The most important failure modes are identity errors, missing provenance, authorization leakage, and stale temporal state. Graph quality is not merely an extraction score; it is whether the system can return the right connected evidence to the right identity at the right time.

Choose by evidence

A useful pilot compares vector, hybrid, and graph approaches on the same corpus and decision questions. Define pass conditions for answer quality, traceability, latency, operating cost, and update behaviour before selecting the production pattern.