AI SQL Agent for Large Databases

/

AI SQL Agent for Large Databases

About Project

AI SQL Agent is an enterprise-grade natural language to SQL solution that enables non-technical users to interact with complex databases using plain English. The system converts business questions into optimized T-SQL queries, executes them securely against SQL Server, and presents both the resulting data and a human-readable summary. To overcome the challenges of large database schemas, the solution uses intelligent two-stage vector-based schema pruning to identify the most relevant tables and columns before generating SQL. This significantly reduces LLM context requirements while improving query accuracy, performance, and cost efficiency. Built with LangGraph, FastAPI, FAISS, and LLM technologies, the platform incorporates read-only guardrails to prevent unauthorized database modifications. An automated self-correction mechanism analyzes SQL execution errors and retries failed queries, improving reliability without manual intervention. The solution demonstrates our expertise in AI application development, intelligent automation, enterprise data solutions, natural language processing, and custom software development.

image-2

image-3

Core Features

Natural Language Queries
SQL Generation
Schema Pruning
Table Selection
Column Selection
Vector Search
Semantic Search
SQL Guardrails
Read-Only Access
Query Execution
Error Correction
Automatic Retries
Result Summarization
Database Analytics
FAISS Indexing
Schema Embeddings
REST API
Enterprise Scalability
Context Optimization
Cost Optimization

Technology Stack:
AI & Orchestration:

Technology	Purpose
LangGraph	StateGraph orchestration of the agent workflow - nodes, edges, self-correction loop
LangChain / LangChain Community	LLM chains, prompt templates, output parsing, FAISS vector store wrapper
Groq (qwen/qwen3-32b)	High-speed inference for SQL generation, manifest generation, and summarization
HuggingFace sentence-transformers (all-mpnet-base-v2)	Local CPU embedding model for table and column vectors
Azure OpenAI / Ollama	Optional alternative embedding and LLM providers
LangSmith	Optional tracing and observability of agent runs

Vector Search & Data Processing:

Technology	Purpose
FAISS	Approximate nearest-neighbour L2 similarity search over table and column vectors
NumPy	Storage and manipulation of pre-computed float32 embedding arrays (.npy)
scikit-learn	Similarity computation and hybrid re-ranking support
pandas	Tabular result handling and transformation

API & Runtime:

Technology	Purpose
FastAPI	REST API exposing /query, /health, and /refresh-schema
Uvicorn	ASGI server running the application
Pydantic	Request and response schema validation
Python 3.12+ / uv	Implementation language and dependency, environment, and script management
PyYAML / python-dotenv	Configuration files and environment credential management

Client: South Africa based client
Region: Africa
Duration: 6 months
Team Size: 1