The LLM never writes the query: declarative search layer over sensitive records (judes.club)

A new declarative search layer has been introduced for an internal assistant designed to retrieve sensitive personnel records, which allows staff to search using plain language requests without directly writing queries. This approach enhances security by preventing users from generating improvised queries that could inadvertently request unauthorized data. Instead, the system utilizes a fixed vocabulary, where all search criteria are predefined and validated against a registry. This means that the AI model can understand the request but cannot manipulate the structure of the searches themselves, ensuring that access to sensitive data remains controlled and compliant with privacy regulations. The significance of this development lies in its dual-phase searching mechanism, which prioritizes data integrity and security. During the first phase, the system narrows down possible matches based on indexed fields, optimizing for speed and resource management given the scale of records (around 10 million). The second phase, which only executes when necessary, retrieves full records for in-depth evaluation of un-indexed fields. This structured approach minimizes the risk of exposing sensitive information and allows for granular permission controls, only displaying data that each user is authorized to see, thereby ensuring compliance and safeguarding privacy within AI-powered tools in the workplace.