EvoOntology vs neocarta
Renmin University's self-evolving ontology layer for data agents: builds a workload-grounded ontology over tables, files and databases, serves it via MCP, and evolves it from agent trajectories. — versus — Neo4j Labs' semantic layer for data agents: ingest warehouse schema, business glossary, metrics and query history into one graph, then serve it over MCP so agents route queries and write grounded SQL.
Both give data agents a semantic map of the warehouse over MCP; neocarta's graph is ingested from schema, glossary and query history, EvoOntology's is built and evolved from the workload.
| EvoOntology | neocarta | |
|---|---|---|
| Stars | 493 | 140 |
| Forks | 41 | 26 |
| Language | Python | Python |
| License | MIT | Apache-2.0 |
| Last activity | today | 7 days ago |
| Topics | data, knowledge-graphs | knowledge-graphs, data |
| Curated connections | 4 | 8 |
EvoOntology — the curator's take
The research answer to stale semantic layers: instead of hand-maintained definitions, EvoOntology builds an ontology from your data and real workload, lets agents fetch only the semantics a step needs through MCP, and proposes targeted updates from execution trajectories, publishing a new version only when paired evaluation beats its parent. Claude Code and Codex plugins make it easy to try. It is two-week-old research code from a university lab: treat the evolution loop as an experiment, keep versions reviewable, and give it read-only credentials. If you want definitions humans own and govern, a curated semantic layer (ktx, neocarta) is the conservative choice.
neocarta — the curator's take
The honest fix for Text2SQL: the model is not bad at SQL, it is blind to your data landscape. Neocarta pulls schema metadata, foreign keys, sample values, glossary terms, governed metric definitions and real query history into a Neo4j graph — only metadata crosses over, data stays in the source — then serves it to agents over MCP with full-text, vector and hybrid search that returns columns, types, example values and the FK references needed to build a join. Embeddings are optional; catalog search works from schema alone. It ships a runnable LangGraph + BigQuery agent so you can see the routing loop end to end. Read the label though: Neo4j Labs, explicitly experimental, not a supported product, 98 stars, and it assumes you are willing to stand up and maintain a Neo4j instance next to your warehouse.