TAHOE: Text-to-SQL with Automated Hint Optimization from Experience

Researchers introduce Tahoe, a system that optimizes LLM-based Text-to-SQL conversion through dynamic prompt engineering rather than model retraining. By consolidating debugging traces into reusable hints and modeling conflicting user intents as strategies, Tahoe increases query pass rates from 62% to 79% on Spider 2.0-Snow benchmarks while maintaining compatibility across weaker model backbones.

Tahoe addresses a critical gap in deploying language models for database access at scale. While LLMs have made SQL generation accessible, production systems demand handling of strict dialect requirements, massive schemas, and user preference shifts—challenges that fine-tuning cannot solve elegantly. The research treats prompt optimization as a data management problem, distinguishing between Syntax Hints derived from compiler feedback and Semantic Hints from execution results, enabling systematic knowledge accumulation without retraining.

The system's innovation lies in its Strategy Layer, which resolves conflicting user intents by modeling them as competing approaches under shared triggers, scored by recency and empirical success metrics. This design acknowledges that real-world deployments face contradictory requirements; Tahoe's attribution framework surfaces which strategies work best in which contexts. The reported improvements are substantial: raising pass rates 17.5 percentage points on GPT-5.5 while reducing compiler feedback cycles from 2.79 to 0.12 per candidate demonstrates meaningful efficiency gains.

For the AI infrastructure market, Tahoe's approach signals a broader trend toward post-training optimization layers that operate outside model weights. This has implications for cost efficiency—avoiding expensive fine-tuning cycles—and modularity, as demonstrated by the Hint Bank's transfer to weaker backbones like Doubao-2.0-lite with 19.7 percentage-point gains. Organizations building database applications gain a pragmatic alternative to continuous model upgrades. However, the paper leaves deployment-phase human feedback updates unexplored, suggesting real-world effectiveness remains to be validated when exposed to genuine production errors and user corrections.

• →Tahoe improves Text-to-SQL pass rates from 62% to 79% through hint-based prompt optimization without model retraining.
• →The system distinguishes Syntax Hints (compiler feedback) from Semantic Hints (execution and user feedback), enabling structured knowledge capture.
• →A Strategy Layer models competing user intents under shared triggers, with post-learning attribution revealing which approaches work best.
• →Hint Banks transfer across weaker models, delivering 19.7 percentage-point improvements on Doubao-2.0-lite and reducing compiler feedback cycles by 95%.
• →The approach addresses production deployment challenges—strict SQL dialects, massive schemas, evolving preferences—without costly supervised fine-tuning.