Structural Abstention Boosts Reliability in AI-Database Systems.

Large Language Models (LLMs) have made natural language interfaces to databases (NLIDB) a credible reality. However, a significant deployment challenge arises from LLM text-to-SQL systems' tendency to produce fluent but incorrect answers, such as hallucinated columns or mis-aggregated totals. When users or other AI agents cannot inspect the generated query, these errors are indistinguishable from correct outputs, posing a critical reliability problem. To tackle this, a novel architectural pattern called "structural abstention" is proposed. This pattern involves a trusted, deterministic kernel paired with a generative shell. The core invariant is that the generative component can only influence *which* question the system answers, never *what value* it returns. The generative shell interprets ambiguous input and phrases responses, while the deterministic kernel matches fully specified questions against a bounded set of answerable shapes, compiling them into queries through deterministic execution. The system requires user confirmation before any value is computed, and requests that the kernel cannot precisely express are declined rather than approximated. This differs from statistical abstention (like selective prediction) because refusal is based on unrepresentability within the trusted kernel, not a confidence estimate. The paper specifies this pattern, provides a five-decision recipe, and includes a two-year production case study, demonstrating its effectiveness in enterprise AI deployments and operational dashboards.