Packt GraphRAG Workshop Impressions
Gifted review ticket from Packt Publishing — honest, unpaid opinion below.
On September 19 I took part in Packt's hands-on virtual workshop "How Modern AI Systems Really Find Answers: Build GraphRAG Applications", led by Dr. Alessandro Negro, Chief Scientist at GraphAware. As promised, here are my impressions.
What I want to stress first: this was a real coding workshop. Alessandro did not stay at the level of architecture slides — he truly worked through the code, notebook by notebook, almost cell by cell. Seven modules, seven notebooks, all backed by a genuine Python package (ingestion, agent loop, enrichment, extraction, similarity, text-to-Cypher, evaluation), scripts and tests. The repository is shared with participants, so the whole build can be revisited afterwards.
The incremental build was the best part of the design. It starts with a one-shot RAG baseline over 10-K filings (Docling parsing, local embeddings, a Company → Document → Chunk graph in Neo4j). Then each module adds one layer: a LangGraph agent with semantic, full-text and navigation tools and grade-and-retry steps; enrichment with Wikidata, Sharadar and New York Times news; LLM-based entity and relationship extraction; similarity linking and entity resolution; a text-to-Cypher tool; and LLM-as-judge evaluation. Every layer was motivated by a limitation the previous one had just visibly hit.
The graph was not just a slide: several times Alessandro opened Neo4j and showed the actual graphs, such as 10-K chunks and resolved entity groups.
What I particularly appreciated was the honesty. Text-to-Cypher was presented as an escape hatch, not a default — with a measured limit: single-hop generated queries are reliable, multi-hop joins are not yet. When two sources disagree about a company's sector, the conflict is kept as data, not silently resolved. And the evaluation uses an anchored rubric with the rationale written before the score, so the number follows the reasoning. For a quant, that attitude matters more than any single technique.
Back to my teaser's question: where does graph-grounded reasoning earn its place? The workshop reinforced my view. Questions that defeat plain vector search — how a company's priorities shifted across consecutive annual reports, which executives moved between rival firms — are relational and qualitative, and that is where the graph pays off. It complements my quantitative models rather than replacing them: numbers stay with the math, while the graph surfaces network structure and the risks and opportunities buried in filings and news, with a trail back to the source.
The Q&A was lively and technical: chunking, entity resolution, Louvain vs Leiden, evaluation, database alternatives.
It exceeded my expectations, in both content and delivery. Thank you, Alessandro and Packt Publishing!
