SQLegit: A Testing Framework for Trustworthy NL2SQL

Determine the correctness of NL2SQL translations using testing techniques.

Overview

About SQLegit

TL;DR: SQLegit is an evaluation harness for NL2SQL systems. Given an NL2SQL translation, it runs a test suite with fast-running test cases to provide a correctness judgment.

Quick Start

Prerequisites

• Python 3.10 is recommended (the Docker image uses Python 3.10).
• A dataset JSON (see the run modes below) and SQLite DB folder.
• One LLM provider credential set:
  • Azure OpenAI via langchain-openai (e.g. AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT, OPENAI_API_VERSION), or
  • DeepSeek via langchain-deepseek (DEEPSEEK_API_KEY)

Docker

Build:

docker build -t sqlegit .

Run (mount your data; pass env via --env-file):

docker run --rm -it \
  --env-file .env \
  -v "/your-data-path:/SQLegit/data" \
  sqlegit /bin/bash

Try It

run_judgment.py is the main entrypoint. It reads a dataset JSON and writes a JSONL file under a sibling results/ directory.

python run_judgment.py --help

Requirements:

• dataset JSON contains at least db_id, question, and query or SQL
• for BIRD, evidence is used as a hint
• --predicted_sql_path points at a text file with one SQL per line (BIRD/Spider); the code reads the first tab-separated field if present
• --schema_file_path is a JSON list with db_id entries (Spider tables.json-style)
• --db_root_path contains per-db folders: <db_id>/<db_id>.sqlite

Example:

python run_judgment.py \
  --judge_name "sqlegit+sem+nos+crs+orc+slf+nlr(gpt-4o-mini-0708)" \
  --benchmark_name nl2sql-bugs \
  --db_root_path /path/to/databases \
  --data_file_path /path/to/NL2SQL-Bugs.json \
  --schema_file_path /path/to/tables.json

How --judge_name is interpreted:

• if it contains sqlegit (case-insensitive), the runner builds a SQLegitJudge
  • backbone model is parsed from the name if it contains gpt-... (defaults to gpt-4o-mini-0708)
  • any of these substrings enable the matching check: sem, nos, crs, orc, nlr, slf

Consensus (run_consensus.py)

If you run SQLegit checks separately (one JSONL per check), you can combine them with a short-circuit + weighted vote policy:

python run_consensus.py --help

To generate the per-check JSONLs, run SQLegit with only one check enabled each time, e.g.:

python run_judgment.py \
  --judge_name "sqlegit+sem(gpt-4o-mini-0708)" \
  --benchmark_name nl2sql-bugs \
  --db_root_path /path/to/databases \
  --data_file_path /path/to/NL2SQL-Bugs.json \
  --schema_file_path /path/to/tables.json

Example:

python run_consensus.py \
  --benchmark-name nl2sql-bugs \
  --db-root-path /path/to/databases \
  --data-file-path /path/to/NL2SQL-Bugs.json \
  --sem-jsonl /path/to/judgments_sem.jsonl \
  --nos-jsonl /path/to/judgments_nos.jsonl \
  --orc-jsonl /path/to/judgments_orc.jsonl \
  --crs-jsonl /path/to/judgments_crs.jsonl \
  --slf-jsonl /path/to/judgments_slf.jsonl \
  --nlr-jsonl /path/to/judgments_nlr.jsonl \
  --out-jsonl /path/to/final_consensus.jsonl

Output Files

run_judgment.py writes under:

• a sibling results/ directory next to your --data_file_path (i.e. <data_dir>/results/)
• file name pattern:
  • judgments,dataset=<benchmark>[+<nl2sql_model>],judge=<judge_name>[+3-shots][+cot].jsonl

Each line is a JSON dict containing:

• final_judgment: true / false / "UNDETERMINED"
• one key per enabled SQLegit check (e.g. semantic_check, oracle_result, ...), each with details like:
  • judgment, confidence, tokens_used, elapsed, per-test-case results, traces, etc.

Code Structure

|-- src/
|   |-- judges/         # LLMJudge + SQLegitJudge implementations
|   |-- testers/        # semantic/noise-row/cross-model/oracle/self-consistency/... checks
|   |-- db_utils/       # schema parsing + execution helpers
|   |-- eval/           # evaluation scripts (Spider/BIRD-style)
|   |-- evalution.py    # evaluation entrypoints (name kept for compatibility)
|-- templates/          # prompt templates for LLM-based components
|-- assets/             # design diagrams (PDF)
|-- run_judgment.py     # run a judge and write JSONL
|-- run_consensus.py    # combine per-check SQLegit outputs into a final decision

Contributing

Contributions and suggestions are welcome.

If you find bugs, encounter problems when running the code, or have suggestions for SQLegit, please submit an issue or reach out to me ([email protected]).