AI / Agentic Systems

SQL Assistant MCP

An experimental agentic SQL system that combines MCP orchestration, multi-agent SQL workflows, compliance checking, result interpretation, and structured evaluation.

View Source Code
Agentic SQL Evaluation

SQL Generation

Generates SQL from natural-language questions.

Compliance Checker

Reviews generated SQL before execution.

Result Interpreter

Transforms query results into meaningful answers.

Project Overview

From natural language to evaluated SQL results

SQL Assistant MCP is a simplified and refactored implementation inspired by AINativeBench. The project is designed so each part of an agentic SQL workflow can be inspected and evaluated independently.

Rather than treating the system as a single black box, the workflow separates SQL generation, compliance checking, and result interpretation into distinct components.

MCP Orchestration

Coordinates the different components of the SQL workflow.

Multi-Agent Workflow

Independent agents handle generation, validation, and interpretation.

Structured Evaluation

Generated results are compared against ground-truth answers.

Step-by-Step Debugging

Individual notebooks make intermediate outputs inspectable.

Agentic Workflow

A modular SQL reasoning pipeline

The system breaks the SQL task into several inspectable stages instead of relying on a single generation step.

01

User Question

Receive a natural-language question that needs to be answered from a database.

02

SQL Generation

An agent generates the SQL query required to answer the question.

03

Compliance Check

A separate agent checks whether the generated SQL satisfies the expected constraints.

04

Result Interpretation

Query results are interpreted and transformed into the final answer.

Evaluation-driven development

The evaluation pipeline uses predefined test cases and ground-truth answers to measure the quality of generated SQL and final results.

Test Cases
SQLite Database
Ground Truth
Generated SQL
Compliance Result
Final Interpretation

Technology

Built for experimentation and evaluation

PythonuvMCPMulti-Agent SystemsSQLSQLiteJupyterLLMLangfuseGROQAINativeBench

Explore the implementation

Inspect the MCP server, individual agent crews, evaluation notebooks, test cases, and supporting utilities in the repository.

Open GitHub Repository