Extending 3DCityDB with Oracle AI Database Support and LLM-Based Natural Language Interfaces

The open-source 3D City Database (3DCityDB) is a widely used solution for storing and managing semantic 3D city models based on the CityGML standard. Its latest version 5 introduces a redesigned and highly generic schema along with a JSON-based schema mapping mechanism, which was initially supported only by PostgreSQL with the PostGIS extension. This paper presents the extension of 3DCityDB to the Oracle AI Database, which offers a database-native natural language interface for interacting with the database. Following a model-driven approach, we ported the complete schema from PostgreSQL to Oracle within a unified relational modelling environment. We implemented a lightweight Oracle adapter that shares a common interface with the PostgreSQL version in the citydb-tool to keep import, export, and query workflows consistent across both database platforms. One of the main research challenges arises when integrating LLM-based natural-language-to-SQL (NL2SQL) through the Oracle SELECT AI package. It typically conveys database semantics via static annotations on fixed physical tables and columns, whereas 3DCityDB v5 adopts a generic Entity-Attribute-Value (EAV) schema whose semantics are encapsulated as JSON in metadata tables. As a result, annotations alone cannot express the structure required for correct query generation. To bridge this gap, we propose a schema-driven approach that automatically derives a dedicated database view from the JSON schema mappings and renders it into a compact, LLM-readable context, which is then injected into the SELECT AI prompt. This enables general-purpose LLMs to resolve hierarchical relationships and generate recursive SQL.