GraphRAG is easy to sell with one question:

What are the hidden relationships across this entire corpus?

Plain vector RAG is easy to embarrass with the same question. It retrieves nearby chunks, misses distant connections, and struggles to summarize themes that no single passage states.

Then a user asks:

What is the cancellation window in policy version 12?

The expensive graph may now be slower, staler, and less direct than a simple filtered search.

GraphRAG is not better RAG. It is a different retrieval investment whose return depends on the shape of your questions.

Classify the Question Before Choosing the Architecture

Use four classes.

Point lookup

Examples:

• exact policy limit;
• account field;
• product specification;
• known document clause.

Plain keyword, metadata-filtered, or vector retrieval often wins. The answer should come from a small number of authoritative passages.

Entity neighborhood

Examples:

• which projects involve this vendor?
• what incidents, owners, and systems relate to service X?
• what claims concern this drug?

A graph can collect connected entities and relationships while preserving source text.

Multi-hop relationship

Examples:

• which supplier change contributed to failures in two regions?
• how are these people, decisions, and projects connected?
• which dependencies connect a vulnerability to affected products?

Graph traversal or hybrid retrieval can expose paths that chunk similarity misses.

Corpus-wide synthesis

Examples:

• what are the dominant themes across all reports?
• which communities and recurring tensions appear?
• how has strategy changed across years?

Microsoft GraphRAG's Global Search targets this class by using generated community reports in a map-reduce process. Its own query documentation calls the method resource-intensive.

The first break-even input is your real query distribution, not a demo question selected to favor graphs.

Understand What You Are Buying

GraphRAG indexing is not “add Neo4j.”

Microsoft's documented pipeline:

load documents
→ chunk documents
→ extract entities, relationships, and claims
→ detect communities
→ generate community reports
→ embed chunks, entities, and reports

Every generated layer has cost and error.

Entity extraction can split one entity into aliases or merge different entities. Relationship extraction can invent or omit edges. Community detection depends on the graph it receives. Generated reports can compress uncertainty into authoritative prose.

The graph is a derived index. It is not automatically the truth.

Keep source references on nodes, edges, claims, and reports. Evaluate graph construction separately from answer generation.

Measure Total Indexing Cost

The project's repository explicitly warns that GraphRAG indexing can be expensive.

Count:

• model calls for extraction and summarization;
• embeddings;
• compute and storage;
• retries and cache;
• entity resolution;
• human review for critical domains;
• reindexing after schema or prompt changes;
• migration between versions;
• permission propagation;
• monitoring and operations.

Then divide by successful graph-advantaged answers, not total questions.

If 95% of traffic is point lookup and only 5% benefits from graph structure, routing those 5% to a graph may be economical. Sending every query through global graph search probably is not.

Freshness Can Destroy the Value

Graph indexing introduces update lag.

A new document may require:

• entity extraction;
• relationship updates;
• community recalculation;
• report regeneration;
• embedding updates;
• permission updates.

For frequently changing operational data, direct tools or conventional search may be more current. Use graphs for relatively stable relationships and a live API for current state.

Track two clocks:

• source freshness;
• derived-graph freshness.

An answer must disclose which snapshot it used. “The graph says” is not acceptable when the graph trails the source by days.

Permissions Are Harder Than Retrieval

Suppose Alice may read documents A and B, while Bob may read B and C. A community report generated from A, B, and C can leak information even if the final query filters source chunks.

Permission-aware GraphRAG must control:

• which documents enter the same derived artifact;
• provenance of every entity, edge, and summary;
• authorization at query time;
• deletion and revocation propagation;
• caches;
• cross-tenant entity resolution.

One approach is permission-homogeneous indexes. Another is fine-grained lineage and policy filtering. Both increase complexity.

If your security model cannot explain who may see a generated community report, the graph is not ready for sensitive data.

Benchmark Retrieval Modes Side by Side

Microsoft GraphRAG exposes useful modes:

• Basic Search: vector RAG baseline;
• Local Search: graph entities/relationships plus text chunks;
• Global Search: community reports for corpus-wide questions;
• DRIFT Search: local search expanded with community context and follow-up questions.

Build a labeled evaluation set by question class. For each mode measure:

• answer correctness;
• citation and evidence precision;
• multi-hop path validity;
• unsupported-claim rate;
• retrieval latency;
• model tokens;
• index amortization;
• freshness;
• permission correctness;
• operator effort;
• cost per successful answer.

Include “no answer” cases. A system that confidently synthesizes a theme from insufficient evidence is not better because it is comprehensive.

Use a Break-Even Scorecard

Graph investment becomes more attractive when:

• relationship-heavy questions are frequent and valuable;
• the corpus is large enough that local chunks miss global structure;
• relationships are stable enough to amortize indexing;
• entity resolution quality is measurable;
• users can tolerate query latency;
• permissions can propagate through derived artifacts;
• the team can operate the pipeline;
• graph-backed answers materially improve outcomes.

Plain or hybrid RAG is favored when:

• questions are mostly point lookup;
• metadata filters and lexical search are strong;
• documents change rapidly;
• strict freshness dominates;
• indexing budget is limited;
• permissions are highly dynamic;
• graph extraction errors are costly;
• citations must map directly to small passages.

Do not turn this into a fixed weighted score copied across businesses. Use it to make assumptions explicit.

Hybrid Usually Means Routing, Not Blending Everything

A practical system can route:

exact identifier or policy lookup → lexical/metadata search
semantic factual question → vector RAG
entity relationship question → local graph search
corpus-wide theme → global graph search
uncertain complex query → bounded DRIFT or agentic retrieval
real-time account state → direct tool

The router can begin with deterministic features and a small classifier. Record the chosen path and evaluate routing errors.

Hybrid does not require concatenating results from every retriever on every query. That increases tokens and makes evidence conflicts harder to reason about.

Evaluate the Graph Itself

Sample nodes, edges, claims, and community reports:

• entity precision and alias resolution;
• relation accuracy;
• source coverage;
• temporal validity;
• contradiction representation;
• unsupported summary claims;
• deletion propagation;
• permission leakage.

Use domain experts for high-risk corpora. A graph that improves benchmark answers while inventing a critical relationship is not production-ready.

Version extraction prompts and models. A change can alter the entire derived topology.

Start With the Cheapest Falsifiable Experiment

Before indexing the company:

1. collect representative questions;
2. label their shape;
3. build a strong plain-RAG baseline;
4. graph a bounded, valuable corpus;
5. compare modes on paired questions;
6. include indexing and operations in cost;
7. inspect failure cases;
8. decide which query classes earned graph routing.

GraphRAG is compelling when the structure is the answer. It is overkill when the user needs one fresh paragraph.

The break-even point is not a model score. It is the moment when better answers to graph-shaped questions repay the cost and risk of maintaining another derived representation of reality.