Multi-Agent SQL Correction Pipeline

This repository contains experiments and pipelines for Text-to-SQL, covering the full workflow from knowledge graph construction to SQL query generation, semantic refinement, correction, and evaluation.

The repository supports the paper "Unified Knowledge Graphs for Adaptive Semantic Refinement in Text-to-SQL", submitted to the ISWC 2026 Industry Track.

The work is developed in collaboration with the IBM T.J. Watson Research Center.

Index

• autoKG/ - framework for building knowledge graphs and schemas from textual data, with examples, scripts, and documentation.
• Structural_KG/ - construction of RDF/Turtle graphs from SQLite database schemas.
• graph_decoration_description/ - pipeline for enriching graphs with semantic descriptions and supporting information.
• multi_agent_text_to_sql/semantic_agent_description/ - multi-agent pipeline that uses semantic graphs and concept descriptions to enrich the dataset and generate SQL.
• multi_agent_text_to_sql/type_agent/ - pipeline analogous to the semantic-agent pipeline, but based on column data types.
• compute_pipeline_EX/ - analysis of the base, type, and description extraction results to measure pipeline correction behavior.
• test with bloked gold/ - extraction and reuse of gold results to speed up query evaluation.
• base_model_test/ - SQL query generation with a base vLLM model and prediction saving.

Summary

The general repository workflow is:

1. build or enrich the database graph;
2. enrich the dataset with semantic or structural information;
3. generate SQL queries with dedicated models or agents;
4. evaluate and compare predictions against gold outputs.

Component Documentation

Each component includes a dedicated README with local usage instructions:

• autoKG/README.md
• autoKG/script/README.md
• Structural_KG/README.md
• graph_decoration_description/README.md
• multi_agent_text_to_sql/semantic_agent_description/README.md
• multi_agent_text_to_sql/type_agent/README.md
• compute_pipeline_EX/README.md
• test with bloked gold/README.md
• base_model_test/README.md

Requirements

Requirements vary by component. The main workflows use Python, RDF tooling such as rdflib, LLM inference tooling such as vllm, transformers, and torch, SQLite benchmark databases, and SLURM/GPU infrastructure for large model runs.

See the component-specific README files for exact inputs, paths, and execution commands.