The answer is in your documents. Just not in any one of them.
Finding information isn't the same as connecting it.
Most AI search (RAG) finds passages that look like your question. That breaks when the answer is spread across files never written to be read together.
Which vendors are affected by the new data-residency rule, and when can we renegotiate their contracts?
From January, customer data must be stored in the EU. Vendors processing personal data must comply or be replaced.
Northwind Analytics processes customer records. Hosting: US-East. Owner: Data team.
This agreement renews on 1 March unless either party gives 60 days' notice.
The challenge isn't finding relevant passages. It's connecting facts that were never written together.
From isolated passages to connected knowledge.
A knowledge graph links entities (companies, people, policies) through relationships (how they connect).
- EU Data-Residency Rule
- requires → EU Storage
- applies to → Customer Data Processing
- Northwind Analytics
- performs → Customer Data Processing
- hosts data in → US-East
- governed by → Vendor Agreement
- Vendor Agreement
- renews on → 1 March
- requires notice → 60 Days
The contract never mentions the regulation. The graph still finds it, through Northwind Analytics.
Northwind Analytics is affected. It processes customer data hosted in US-East, and the new rule requires EU storage from January. Its contract renews on 1 March with 60 days' notice. That notice deadline is the key date for renegotiation.
- Policy_update_Q3.pdf
- Vendor_register.csv
- MSA_NW_Analytics_2023.pdf
RAG searches by similarity. GraphRAG also searches by entities and relationships.
Same question, same three files. What each approach finds, and how it finds it:
Conventional RAG
- Similarity
- Entities
- Relationships
- Policy_update_Q3.pdf by similarity
- Vendor_register.csv by similarity
- MSA_NW_Analytics_2023.pdf missed
The contract talks about renewal, not data residency. It doesn't look like the question.
GraphRAG
- Similarity
- Entities
- Relationships
- Policy_update_Q3.pdf by similarity
- Vendor_register.csv by entity: Northwind Analytics
- MSA_NW_Analytics_2023.pdf by relationship: vendor → agreement
It follows rule → vendor → contract. This is a multi-hop question: several facts, linked.
| Conventional RAG | GraphRAG | |
|---|---|---|
| Searches by | Semantic and keyword relevance | Relationships between entities, often alongside semantic search |
| Connects facts | Through information retrieved together | Through explicitly represented relationships |
| Cross-document questions | Depends on retrieving all necessary passages | Can follow connections to discover additional evidence |
| Best suited for | Direct questions, summarisation and document lookup | Multi-hop reasoning, dependencies and relationship-heavy questions |
| Trade-off | Simpler to build and operate | Requires accurate graph construction and maintenance |
Similarity finds passages related to the question. A graph finds information related to each other.
Some questions are really about connections.
Legal and contracts
Which contracts allow termination following a change of control?
CompanyAgreementTermination clause
Compliance and risk
Which suppliers are exposed to the new regulation?
RegulationActivitySupplierContract
Due diligence
How is this director connected to the companies we're reviewing?
DirectorShareholdingCompany
Research
Which studies support this finding, and which contradict it?
StudyShared claimStudy
Engineering
Which systems will be affected if we retire this component?
ComponentServiceDownstream system
Institutional knowledge
Why did we choose this vendor in the first place?
VendorEvaluationDecision
The answer isn't just a fact to look up. It's a set of relationships to uncover.
However, not every question needs a knowledge graph.
Conventional RAG is enough
The answer sits in one place.
What is this product's warranty period?
GraphRAG pays off when
- Facts are scattered across several documents
- Relationships matter: who owns, depends on or governs what
- Reasoning takes several steps
- The same entity appears under different names
- Evidence is indirect, linked through other entities