Athena: a wise database fine tuning mechanism

Abstract

The translation of natural language into SQL queries (Text-to-SQL) has advanced significantly with the advent of Large Language Models (LLMs). However, conventional approaches often fail in two critical aspects: susceptibility to schema hallucinations (Schema Linking) and the generation of inefficient queries that disregard database performance. This work presents Athena, a self-adaptive SQL agent designed to act not merely as a translator, but as an autonomous Database Administrator (DBA). The proposed architecture implements a sequential refinement pipeline composed of three modules: Grounding, Cognitive Processing, and Physical Execution. To mitigate hallucinations, a Retrieval-Augmented Generation (RAG) strategy is employed, injecting semantic few-shot examples into the context. The cognitive core, driven by the DeepSeek Coder model, operates in a dual-stage process: first translating the user’s intent and subsequently refining the code via a heuristic knowledge base grounded in Relational Algebra and academic literature (Are-SQL). Finally, the system introduces a "Test-Drive" mechanism for physical optimization, where candidate indexes are created provisionally and validated through a 10% improvement rule, ensuring that only empirically effective optimizations are retained. Experiments conducted on the TPC-H benchmark over PostgreSQL demonstrate that Athena outperforms GPT-4-based approaches in adherence to rewriting rules and achieves drastic reductions in complex query latency (e.g., from 281s to 3.9s), validating the efficacy of the proposed hybrid architecture.