Cloudera uses deterministic code to ensure AI answer accuracy
Cloudera developed a natural language querying tool that uses AI only to vectorize questions, while on-premise agents execute deterministic SQL and Python. This approach ensures 100% accuracy and data privacy by avoiding direct generative model access to private data.
To address hallucination and privacy concerns, Cloudera developed a 'barely AI' natural language querying tool. Instead of letting generative models directly access private data, the system uses a public model only to vectorize user questions. These vectors are matched against heavily curated business and technical metadata. On-premise, air-gapped AI agents then execute pre-existing, deterministic SQL and Python code to generate answers, ensuring 100% accuracy and data privacy. The system returns the answer, the SQL code, and data lineage. If no match is found, the agent may generate new SQL but flags the answer as AI-generated rather than from the enterprise repository.
Christian noted that coding agents like GitHub Copilot, Cursor, and Claude Code are used internally with data masking techniques, achieving 75% or greater efficiency in coding tasks. The approach relies on extensive data governance, including renamed tables and synonym columns, to help AI semantically match user intent with technical schemas. Christian contends that relying on generative AI for direct data querying is unreliable and insecure; instead, leveraging existing deterministic codebases with AI as an interface layer ensures accuracy and privacy.