Chat-to-SQL: Ask your database in plain language (technical guide)

What is Chat-to-SQL?

Chat-to-SQL is an architecture that combines Retrieval-Augmented Generation (RAG) with language models to translate questions in plain language into executable SQL queries.

It’s not magic: it’s applied engineering with a proven stack.

System architecture

User → Question in plain language
    ↓
Catalog embedding (pgvector)
    ↓
Relevant context + Schema
    ↓
LLM generates SQL
    ↓
Validation + Execution
    ↓
Response in natural language

Technical stack

• PostgreSQL 16 with pgvector extension for embeddings
• FastAPI as inference API
• LangChain for RAG pipeline orchestration
• Fine-tuning of the model with the client-specific catalog

Accuracy and limitations

In our deployments, the system achieves:

• 94% accuracy on inventory queries
• 87% accuracy on complex financial queries
• < 2 seconds average latency

Queries that fail are usually ambiguous or refer to data that doesn’t exist in the schema. The system learns with each correction.

Typical implementation: 10 days to pilot

1. Days 1-5: ERP connection, schema extraction, embedding generation
2. Days 6-10: Validation with real data, adjustments, and scoped pilot deployment

View interactive demo →