Confidence Scoring for LLM-Generated SQL in Supply Chain Data Extraction

2025-06-20

AI Summary

Large language models (LLMs) generate SQL queries for supply-chain data extraction without quantifying uncertainty, hindering reliability assessment. Method: We propose a confidence estimation method based on cross-modal embedding similarity, integrating dual-modality embeddings (e.g., text-embedding-3-large), cross-lingual back-translation verification, and meta-prompted confidence generation. Contribution/Results: We systematically compare three strategies—translation consistency, embedding semantic similarity, and LLM-reported confidence—and find that SQL–natural-language cross-modal embedding similarity significantly outperforms overconfident self-reported scores in accuracy discrimination. Our method achieves an AUC of 0.89 for SQL correctness identification—32% higher than self-reported confidence—enabling effective human review prioritization and automated query rejection in high-risk scenarios.

Technology Category

• Natural Language Processing: Information Extraction
• Machine Learning: Large Multimodal Models (LMMs)
• Data Mining & Knowledge Management: Intelligent Query Processing

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large Language Models (LLMs) have recently enabled natural language interfaces that translate user queries into executable SQL, offering a powerful solution for non-technical stakeholders to access structured data. However, one of the limitation that LLMs do not natively express uncertainty makes it difficult to assess the reliability of their generated queries. This paper presents a case study that evaluates multiple approaches to estimate confidence scores for LLM-generated SQL in supply chain data retrieval. We investigated three strategies: (1) translation-based consistency checks; (2) embedding-based semantic similarity between user questions and generated SQL; and (3) self-reported confidence scores directly produced by the LLM. Our findings reveal that LLMs are often overconfident in their own outputs, which limits the effectiveness of self-reported confidence. In contrast, embedding-based similarity methods demonstrate strong discriminative power in identifying inaccurate SQL.

Problem

• Assessing reliability of LLM-generated SQL queries
• Evaluating confidence scores for supply chain data extraction
• Addressing LLM overconfidence in self-reported SQL accuracy

Innovation

• Translation-based consistency checks for SQL
• Embedding-based semantic similarity assessment
• LLM self-reported confidence scoring