My ArcadeDB in Python paper is in the SciPy Proceedings
One embedded database for documents, a graph and vectors
My paper, ArcadeDB in Python: An In-Process Multi-Model Database, was accepted at SciPy 2026 and is now in the SciPy Proceedings. I wrote it alone. The research page on humem.ai has the method and the benchmark tables.
What it is
Some Python programs need transactions over documents, graph traversals, and vector search over the same data. The usual answer is three systems, each with its own copy of the data, and application code that keeps the copies in sync.
ArcadeDB is a multi-model database written in Java. The package runs it inside the Python process, with a Java runtime bundled in the wheel, so one database does all three and you do not need Java installed. You query documents with SQL, the graph with openCypher, and vectors through an HNSW-style index.
The paper shows a workflow that chains a vector search, a SQL filter, and a graph traversal over one dataset. It also benchmarks the package against SQLite, DuckDB, LadybugDB, and Chroma on a Q&A corpus. Each specialist wins on its own model. ArcadeDB is competitive on all three, and what it costs is memory and loading time.
The ArcadeDB project wrote the engine. I wrote the Python bindings and the packaging. The
package is named arcadedb from the 26.11.1 release on.