Ask Your DB | Moraldeep Sachdeo

Ask Your DB — Natural Language to SQL Agent

An interactive text-to-SQL system that translates plain-English analytics questions into executable SQL, validates safety, and measures quality with a rigorous evaluation harness.

Natural Language Analytics Agent

Ask a data question in English, get executable SQL and grounded answers in seconds

The system introspects live database schema, generates SQL with open-source models on Fireworks AI, executes the query, and retries once with error feedback when execution fails. It also includes SQL guardrails and benchmark scripts to track correctness, latency, and cost.

Overview

Ask Your DB is built around a practical product question: can a text-to-SQL agent be accurate enough for real analytics workflows while staying fast and cost-efficient? Instead of optimizing for demo output, the project emphasizes executability, safety, and measurable quality across both known and unseen schemas.

What The Product Does

Runs an interactive CLI where users ask database questions in plain English and see generated SQL plus returned rows
Grounds SQL generation in live schema metadata (tables, columns, and sample rows) to reduce hallucinated references
Executes generated SQL and performs an automatic repair pass when runtime errors occur
Evaluates output using strict, column-aware, and relaxed result matching to avoid misleading single-metric claims
Supports adversarial checks that block injection-like prompts and reject DDL or DML operations

Implementation Details

Built in Python with a modular agent and evaluation harness under src/
Uses SQLite as execution backend with Chinook plus a synthetic GitLab-style analytics schema for generalization testing
Uses open-source LLM endpoints served on Fireworks AI for text-to-SQL generation
Implements execute-then-retry logic to recover from dialect and naming mismatches automatically
Includes performance instrumentation for latency and token-cost estimation under production-like query loads

Database Description

The primary database is Chinook, a digital music store schema in SQLite with 11 related tables including Artist, Album, Track, Customer, Invoice, and InvoiceLine. This gives the agent realistic join paths for customer behavior, revenue, and catalog analytics. To test generalization beyond a familiar benchmark, the project also generates a second synthetic GitLab analytics schema with 7 tables and a larger templated question set.

In practice, the agent starts by reading schema metadata and sample rows from the active database. That context is injected into the prompt so SQL generation is grounded in actual table and column names rather than guessed structure.

Example Query Walkthrough

One representative flow from the CLI looks like this:

Question: Which 5 customers spent the most in total?

Generated SQL:

SELECT
  c.CustomerId,
  c.FirstName || ' ' || c.LastName AS CustomerName,
  ROUND(SUM(i.Total), 2) AS TotalSpent
FROM Customer c
JOIN Invoice i ON i.CustomerId = c.CustomerId
GROUP BY c.CustomerId, c.FirstName, c.LastName
ORDER BY TotalSpent DESC
LIMIT 5;

The agent executes this SQL directly and returns the top rows in the terminal, for example customer name and lifetime spend. If execution fails due to a dialect mismatch or naming issue, the error is fed back once and the query is auto-repaired before returning a final answer.

Why It Matters

This project focuses on the gap between demo-grade and production-minded AI analytics. The value is not only that it produces SQL, but that it reports whether results are truly correct, stays within practical latency and cost ranges, and fails safely when a query is unsafe.

Key Outcomes

100% executable SQL on evaluated dev and held-out suites with the hardened setup
~90% relaxed result-match on the core dev benchmark
p50 latency near interactive usage levels and significantly lower projected cost versus a GPT-class pricing baseline
100% rejection on tested injection and DDL/DML attack patterns

Role and Focus

Role: Solo builder focused on agent architecture, evaluation design, reliability hardening, and cost-performance tradeoffs.

Tech Stack: Python, SQLite, Fireworks AI, OpenAI-compatible SDK, CLI tooling, JSON-based eval pipelines.

Category: Text-to-SQL, agentic analytics, LLM evaluation, trustworthy AI systems.

Positioning: Think "terminal-native BI copilot," designed for measurability and safe execution rather than prompt demos.

Project Repository

View source on GitHub