Data Catalog Mediation for Secure LLM Analytical Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems using large language models (LLMs) for data interaction face challenges in maintaining security, access control, and accuracy, particularly in preventing unauthorized access to sensitive data and reducing hallucinations, while lacking structured mechanisms for user authorization and data handling.
Innovation Solution
A system that orchestrates an LLM-directed workflow to generate structured queries for a data catalog, enabling selective access based on user credentials, using metadata for grounding responses, and allowing dynamic user intervention to ensure secure and accurate data interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If LLMs interact directly with data catalogs to enable natural language queries, then user interaction intuitiveness is improved, but security risks and unauthorized access increase
Solution Approach 1:
The patent introduces a data catalog as an intermediary layer between users and the underlying data storage systems. The data catalog mediates all data access requests, translating natural language queries into structured queries while enforcing access control policies. This mediator prevents direct access to sensitive data, thereby reducing security risks while maintaining intuitive user interaction through natural language interfaces.
2Speed
If LLMs are used to generate data queries and responses, then query processing speed is improved, but hallucination and inaccuracy increase
Solution Approach 1:
The patent implements feedback mechanisms where the data catalog validates and verifies data before returning results to the LLM. The system receives feedback about data availability, accessibility, and schema information from the data catalog, allowing the LLM to adjust its responses accordingly. This feedback loop reduces hallucinations while maintaining fast query processing speeds through efficient data catalog indexing and query optimization.
3Object-affected harmful factors
If access control mechanisms are added to restrict LLM access to data, then data security is improved, but system complexity increases
Solution Approach 1:
The patent applies preliminary action by pre-configuring access control policies and data catalog schemas before data access is needed. Users and systems register their access requirements and data needs in advance, allowing the data catalog to automatically enforce appropriate access controls without adding complex real-time decision-making logic. This preliminary configuration approach enhances data security while minimizing system complexity during data access operations.
4Object-affected harmful factors
If structured query mechanisms are implemented for data access, then data access control is improved, but query processing time increases
Solution Approach 1:
The patent implements preliminary action by pre-compiling and indexing data catalog schemas, access control policies, and query execution plans before actual queries are processed. When a natural language query arrives, the system can quickly translate it into the appropriate structured query using pre-established mappings and access control rules, significantly reducing query processing time while maintaining robust data access control mechanisms.
Data Source
AI summary
A computer-implemented system, computer-implemented method, and computer-program product includes receiving a natural language query from a user for executing an analytical task; generating an analytical large language model (LLM) prompt based on the natural language query and, in response to generating the analytical LLM prompt, orchestrating an LLM-directed workflow for handling the natural language query by: automatically prompting, using the analytical LLM prompt, an analytical task-oriented LLM to generate a structured query for querying a data catalog application; querying the data catalog application using the structured query generated by the analytical task-oriented LLM; obtaining query results from the data catalog application, where the query results include metadata associated with at least one element accessible to the data catalog application; prompting the analytical task-oriented LLM to identify a given analytical task associated with a given analytical agent; and automatically executing, by the given analytical agent, the analytical task.