Natural Language to T-Sql issue: sqlcoder-7b-2 fails on complex T-SQL joins & date logic (offline, 40GB GPU)

Models
1

Project Overview

I am building a Natural Language → T-SQL system for Microsoft SQL Server (T-SQL).

Expected behavior:
If a user asks a natural-language question (e.g.,
“How many users are using smartphones last month?”),
the system should generate a valid and logically correct T-SQL query.

Constraints

  • Maximum GPU memory: 40 GB
  • Deployment: Local GPU inference only
  • No internet access after training (fully offline deployment)
  • This restricts model size and external API usage

Current Architecture

  • LLM: defog/sqlcoder-7b-2
  • Fine-tuning: ~2,500 complex SQL queries
    • Multi-table JOINs
    • Aggregations
    • Date logic
  • Schema Handling (RAG):
    • Tables and column descriptions stored separately
    • Embedded using MiniLM
    • Retrieved via cosine similarity
  • Generation Flow:
    1. User NL query
    2. Retrieve relevant schema context
    3. Inject schema into prompt
    4. Generate T-SQL

What Works

  • Simple queries
  • Single-table queries
  • WHERE / GROUP BY / HAVING
  • Basic aggregations

Issue

For complex queries involving:

  • Multiple JOINs
  • SQL Server date functions (DATEADD, DATEDIFF, CONVERT)
  • Cross-table business logic

the model often:

  • Chooses incorrect JOIN paths
  • Misses required tables
  • Hallucinates columns
  • Produces SQL Server–invalid date syntax
  • Generates logically incorrect queries

This happens despite fine-tuning and schema grounding.

Questions

  1. Is this mainly a 7B model limitation for complex for this project?
  2. Would explicitly injecting foreign-key relationships / join graphs into the prompt help?
  3. Is a query-planning stage (join planning → filters → final SQL) recommended?
  4. Any best practices for T-SQL–specific correctness?
  5. Given offline + 40 GB GPU constraints, would:
    • Larger quantized models
    • Multi-stage planners
    • Rule-based join resolution + LLM
      be more reliable?
  6. Are there any open-source or production-grade Natural Language to Sql architectures that handle complex joins reliably under similar constraints?

Goal

To generate correct, production-ready T-SQL for complex NL queries under offline and 40 GB GPU constraints.

Thanks in advance for any guidance or references!

1 Like
2

While the small size of the model seems to be one factor, it appears that even large models may struggle to solve the problem unless the task is broken down.

1 Like