SCHEMA wins in Neo4j SimpleKGPipeline test

The test deliberately held the LLM and chunking constant, changing only how the schema was supplied. To split by recipe rather than a fixed size, the author wrote an H1MarkdownTextSplitter that cuts the Markdown at each H1 heading, so the curry rice and nikujaga sections become one chunk each. The same two recipes include shared ingredients — beef, onion, carrot, potato — which gave the comparison a way to see whether those shared links and the cooking sequence were represented.

The results followed a clear pattern. FREE, with no schema, produced near-duplicate node types for the same object and multiple relationship types meaning similar things, leaving text-to-Cypher queries uncertain about which type to pick. EXTRACTED made type names more consistent because the LLM inferred one schema from the whole document, but it did not model which step uses which ingredient or the order between steps, so process-oriented questions stayed unanswerable. SCHEMA, with human-defined node types, relationship types, and patterns such as Recipe-REQUIRES_INGREDIENT->Ingredient and Step-NEXT_STEP->Step, let the author traverse shared ingredients and step order. Even so, across repeated runs with the same schema and input, node names varied, suggesting a residual constraint common to all three modes.