DIVER: A Robust Text-to-SQL System Unveiled at SIGMOD 2026, Powering ChatBI

Research Background

With large language models (LLMs) reshaping many industries, natural‑language database interfaces (Text‑to‑SQL) are becoming increasingly important for allowing non‑technical users to query massive data using everyday language. Existing models achieve high scores on benchmarks such as BIRD, but their performance collapses—often by more than 10%—in real‑world scenarios where expert‑written evidence is unavailable.

Key Insight

Fine‑grained analysis across models of different scales reveals that value linking —the component of expert evidence that maps concrete values in a query to database entries—is the most critical yet weakest link. Current value‑linking methods rely on a single, static semantic‑similarity recall, which fails when users ask ambiguous questions or when databases contain many values without clear semantics (e.g., “K‑12”). This deficiency, hidden by expert evidence in benchmark settings, is the root cause of the robustness problem.

Core Innovations

1. Automated Expert Agent – DIVER builds a multi‑agent collaboration framework consisting of a “Decomposition Assistant”, an “Exploration Assistant”, and an “Evidence Assistant”. The system first breaks a complex natural‑language query into single‑entity sub‑clauses, then conducts multi‑turn dialogues with the database for each sub‑clause, and finally synthesizes the discovered facts into clear reasoning evidence, achieving full‑process automation that mimics human expert workflow.

2. Dynamic Interactive Value Linking – To overcome the limitations of static semantic matching, DIVER introduces a dynamic, interactive value‑linking method that iteratively moves from “guess” to “verify” within a structured reasoning loop. The system employs eight meta‑tools that allow flexible database interaction, and all exploration follows a strict JSON‑constrained workflow called Chain of Thoughts and Facts (CoTF): hypothesis → observation → verification.

3. Adaptive Evidence Generation – The “Evidence Assistant” intelligently aggregates the structured information verified by the Exploration Assistant into natural‑language evidence tailored to the downstream Text‑to‑SQL model. For large‑parameter prompting models such as GPT‑4, it produces detailed, long‑form evidence; for fine‑tuned models like CodeS, it generates concise, core‑instruction evidence, thereby maximizing downstream accuracy and robustness.

Experimental Validation

Comprehensive experiments demonstrate that DIVER substantially improves the robustness and accuracy of mainstream Text‑to‑SQL models in expert‑free real‑world scenarios.

Performance boost: On the BIRD benchmark, DIVER raises the execution accuracy (EX) of models of various scales, with the largest gain of +10.82% for GPT‑4o.

Robustness: On challenging datasets that simulate dynamic schema changes and ambiguous questions (e.g., DR.Spider), existing models suffer severe degradation, whereas DIVER improves performance by over 10% and achieves multiple new state‑of‑the‑art results, confirming the effectiveness of dynamic interactive value linking.

Application to ChatBI

ChatBI incorporates DIVER’s core capabilities—automated expert agent and dynamic interactive value linking—into a next‑generation intelligent analysis platform. The system no longer depends on simple semantic matching; instead, it initiates a hypothesis‑observation‑verification reasoning loop to handle vague business queries while maintaining high execution accuracy and robustness without human intervention.

Additional features built on DIVER include a business semantic layer that aligns natural‑language terms with database fields, multi‑turn context memory for follow‑up questions, and AIOps‑driven causal insight, forecasting, and full‑domain visibility that extend analysis from “what happened” to “why it happened” and “what may happen next”.