The limitation is often relational, not semantic
Conventional RAG retrieves chunks that are semantically similar to a query. That is useful for local questions but can struggle when the answer depends on relationships distributed across a corpus.
Microsoft Research’s GraphRAG work was motivated by this distinction. For global sensemaking questions over large corpora, it builds an entity graph and community summaries, and reported stronger comprehensiveness and diversity than a conventional RAG baseline for the evaluated class of questions (Edge et al., 2024).
Enterprise implementations need provenance and noise control
Adobe researchers described a production-oriented KG-RAG assistant that builds a low-noise graph with entity resolution, confidence scoring and links from facts back to source documents. Their evaluation reported fewer irrelevant responses and substantially more fully relevant answers than the prior production system (Mukherjee et al., 2025).
The important design principle is provenance. A graph should not become a second opaque knowledge store. Every relation used in reasoning should remain traceable to the document, record or authoritative source from which it was derived.
- 01Documents
- 02Entity / relation extraction
- 03Knowledge graph
- 04Graph-aware retrieval
- 05Grounded answer
Graph structure complements rather than replaces conventional retrieval.
Graph and vector retrieval solve different parts of the problem
KG²RAG uses semantic retrieval to find seed chunks and then expands and organises context through knowledge-graph relationships. This hybrid pattern is useful because vector search remains efficient for candidate discovery while graph structure improves coherence and multi-hop context (Zhu et al., 2025).
For enterprise systems, the architecture choice should follow the question type. Local fact retrieval may need only high-quality vector RAG. Questions about dependencies, ownership, causal chains, counterparties or cross-document consistency often benefit from explicit structure.
Arvec view: ontology gives the graph a business meaning
A knowledge graph becomes more useful when its entities and relations reflect the organisation’s domain semantics. An ontology can define what counts as a customer, obligation, shipment, facility, risk, control or document evidence—and how those concepts relate.
This creates a bridge from retrieval to decision support: documents provide evidence, the ontology provides meaning, the graph provides relationships and the LLM provides language-level reasoning.
Key implications
- 01
Choose retrieval architecture based on the question, not on a default RAG pattern.
- 02
Preserve provenance from graph facts back to source evidence.
- 03
Use ontologies where domain meaning and relationship consistency matter.
- 04
Combine vector retrieval and graph traversal rather than treating them as substitutes.
Related AIF7™ capabilities
AI Data Foundation
Creates semantic, provenance-rich enterprise knowledge.
AI Engineering & Delivery
Selects retrieval patterns appropriate to question type and scale.
AI Governance & Trust
Maintains source traceability for generated answers.
AI Strategy & Vision
Focuses knowledge architecture on decisions that create value.
Sources & further reading
01 · Edge et al. 2024
From Local to Global: A Graph RAG Approach to Query-Focused SummarizationMicrosoft Research / arXiv · 24 April 2024
02 · Mukherjee et al. 2025
From Documents to Dialogue: Building KG-RAG Enhanced AI AssistantsAdobe Research / arXiv · 21 February 2025
03 · Zhu et al. 2025
Knowledge Graph-Guided Retrieval Augmented GenerationarXiv · 8 February 2025