Phase 3Module 20 of 22

Graph RAG

Some questions are never answered in a single similar paragraph. Graph RAG walks relationships the documents imply but never state together. Extraction and entity resolution cost more than naive RAG — use it when hops matter.

Vector RAG is 'find pages like this question.' Graph RAG is 'start at this company and follow acquired-by until you hit a country.'

Visual Workflows

Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.

Overview

100%

Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

100%

Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Microsoft-style GraphRAG idea

100%

Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Build communities, summarize, then retrieve a subgraph plus text.

Hybrid is common

100%

Loading diagram...

Scroll inside the frame to explore · use + / − to zoom up to 200%

Vectors find the paragraph. The graph finds the hop the paragraph never states.

Key Takeaways

1.Graph RAG retrieves a subgraph (and often text chunks) then generates.
2.It is for multi-hop and relational questions, not a default replacement for vector RAG.
3.Typical flow: extract a graph from the corpus, resolve entities, optionally cluster communities and summarize (Microsoft GraphRAG-style).
4.At query time, identify seed entities, pull a bounded subgraph plus related text, and prompt the LLM with both.
5.Hybrid is common: vectors find the seed passage, the graph expands.

Real Example

Scenario

'Which products of companies Acme acquired in 2024 mention GDPR?' needs acquisition edges and then document text — neither store alone is enough.

What you would do

In RAG Engineering, apply Graph RAG to this scenario: 'Which products of companies Acme acquired in 2024 mention GDPR?' needs acquisition edges and then document text — neither store alone is enough. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.

Practice Task

Open the Code Walkthrough below and run it locally. Change one parameter related to Graph RAG (e.g. model, temperature, top_k, or tool name), observe the difference in output, and write 2–3 sentences explaining what changed.

Code Walkthrough

Highlighted lines show where Graph RAG happens in the code.

Graph RAG

1# Graph RAG — minimal example2from openai import OpenAI3
4client = OpenAI()  # create API client5
6# Ask the model to explain this topic7response = client.chat.completions.create(  # core API call for Graph RAG8    model="gpt-4o-mini",9    messages=[10        {"role": "system", "content": "You explain graph rag clearly."},11        {"role": "user", "content": f"What is graph rag?"},12    ],13    temperature=0,14)15print(response.choices[0].message.content)  # show output for debugging

Commands

Commands to Remember

Graph RAG = retrieve a subgraph (and often chunks), then generate
Use it for multi-hop / relational questions, not every FAQ
Extraction + entity resolution quality bounds the whole system
Eval needs multi-hop questions, not only single-paragraph facts

Cheat Sheet

Quick recap

quick ref

•Retrieve subgraph + text, then generate
•For multi-hop / relational questions
•Extraction quality bounds the system
•Eval with hop questions, not only FAQs

Common Mistakes

✕Running Graph RAG on every FAQ — expensive and slower
✕Skipping entity resolution so hops never connect
✕Unbounded traversal that dumps the whole graph into the prompt