Hi Kyuubi community,
I'd like to propose a new engine type: Data Agent Engine, which enables
users to perform data analysis through natural language.
What it does: Users connect via JDBC/REST, ask questions in plain language,
and an AI agent (ReAct loop) autonomously explores schemas, generates SQL,
executes queries, verifies results, and self-corrects — without user
intervention.
Key architecture decision: The agent's SQL tool uses Kyuubi's JDBC driver
to connect back to Kyuubi Server, following the same pattern as the
existing JDBC Engine. This ensures all queries go through the Server
gateway, preserving AuthZ (Ranger), audit logging, multi-tenancy, and
resource isolation.
Client → Kyuubi Server → Data Agent Engine
│ (sql_query tool)
JDBC Driver (user creds)
│
Kyuubi Server → Spark/Trino/Hive
Implementation: Java + LangChain4j on the JVM, consistent with all
existing Kyuubi engines.
Full proposal: https://github.com/apache/kyuubi/discussions/7373
Looking forward to your feedback!
Best,
Zhigang
