Everything so far was practice. This is the project you can show someone.
The brief
Build a question-answering app over a corpus you care about, where plain vector RAG visibly falls short. Good corpora are ones with entities that recur across many documents: a codebase's docs and issues, a podcast's transcripts, a set of research papers, your own notes, public filings, a game wiki.
Requirements
- Corpus: at least 50 documents or pages, publicly usable or your own.
- Graph: built with a schema you designed, with entity resolution and chunk provenance.
- Two or more retrieval patterns, plus a router or agent that picks between them.
- Baseline: a vector-only retriever on the same corpus.
- Evaluation: a test set of at least 30 questions, scored on both systems.
- Write-up: a README with an architecture diagram, results table, token costs and an honest "where it fails" section.
Suggested path
| Step | Output |
|---|---|
| Choose corpus and write 30 questions | questions.md |
| Design schema from the questions | schema diagram |
| Build graph, inspect 20 facts | graph plus error notes |
| Baseline retriever | baseline scores |
| GraphRAG retrievers and router | improved scores |
| Evaluation and cost analysis | results table |
| Write-up and demo | public repo |
Stretch goals
- Add a global-search mode for theme questions.
- Show the graph path that led to each answer in the UI.
- Add incremental updates so new documents merge into the existing graph.
Share it
Publish your repo and write up what surprised you. A candid "GraphRAG did not help on these question types" result is as valuable as a win, because it shows you can judge when the technique earns its cost.
Where to go next
Articlegraphrag.com
Research papers on GraphRAG (opens in a new tab)
Continue into the research literature.
CodeGitHub
microsoft/graphrag (opens in a new tab)
Read the source of a production-grade indexing pipeline.
DocsNeo4j
Neo4j GraphRAG for Python (opens in a new tab)
Explore the retrievers and pipeline components you did not use.