Capstone, your own GraphRAG
Build a GraphRAG knowledge graph about your CV, a book or a hobby, test it with ten questions, and fix wrong answers.
The goal
Build a graph about something you know well, then prove it can answer real questions. This is the piece you can show in a portfolio or a job interview.
Pick a topic
Choose something with clear things and clear links between them. Some ideas:
- Your CV: you, your jobs, skills, projects and the tools each project used.
- A book or series: characters, places, and who is related to whom.
- A hobby: recipes and ingredients, a football team and its players, plants and their care needs.
Build it in six steps
- Click + New graph and give it a name.
- Add 15 to 30 nodes. Give each a type, a one-line description and a few properties.
- Connect them with clear labels: Worked at, Used, Friend of, Needs.
- Click Save graph, then Generate embeddings.
- Write 10 questions: 5 that one node can answer, and 5 that need a relationship.
- Ask all 10, and note which were right, which were wrong, and why.
When an answer is wrong
| What you see | Likely cause | Fix |
|---|---|---|
| The wrong nodes are matched | Node descriptions are too short or vague | Write clearer descriptions, then regenerate embeddings |
| The right node is matched, but the answer misses a fact | The link is missing or badly labelled | Add or rename the relationship |
| The facts are there but the answer is clumsy | The model is too small | Try another free model, or a paid one (Chapter 5) |
You did it
In 5 days you have called an LLM from Python, turned text into numbers, searched by meaning, used a vector index, built RAG by hand, and run a working GraphRAG app on your own data. That is the core toolkit of Applied AI.
Keep going with GraphRAG Chats
The app is open source at github.com/dooa-ansari/GraphRag-Chats. Star it, read the code, and try one of these next challenges:
- Follow relationships two steps instead of one.
- Turn a pasted paragraph of text into nodes and relationships automatically, using an LLM.
- Add a setting to pick the chat model from the screen.
You are done when your graph answers at least 8 of your 10 questions correctly.
Done reading and tried the practice task? Mark the chapter complete to save your progress.