Automating Text-to-SQL in WMS: Choosing a Local LLM Without GPU
DeepSeek
Google/DeepMind
Alibaba/Qwen
A developer at the logistics company Aerosib-C automated the generation of SQL queries for invoicing using LLM models. Due to security requirements, local models were used, tested on a Linux server with 30 GB of RAM and no GPU. The best result was achieved by gemma4:26b, although response times reached tens of minutes. Coder models were outperformed by reasoning models. Plans include building an expert system for service classification.
A developer at Aerosib-C has implemented a task to automate invoicing for logistics services based on data from the WMS (warehouse management system) and handheld terminals. Local LLM models were chosen for generating SQL queries from text descriptions to avoid transferring sensitive business data to third parties. Testing was conducted on a Linux Debian server with 30 GB of RAM and no video accelerators. Out of seven tested models (deepseek-coder-v2:16b, gemma4:26b, qwen3-coder:30b, qwen3:14b, qwen3:30b, qwen3.5:9b), only three produced correct SQL code: gemma4:26b, qwen3:14b (suboptimal), and qwen3:30b. The winner was gemma4:26b, despite a very long response time (over 21 minutes). It was noted that specialized coder models performed worse than reasoning models. The author plans to create an expert system classifying services into groups to simplify selecting text descriptions for queries. The article concludes by emphasizing that AI is a tool that needs to be mastered and applied creatively, without generating incomprehensible code.
Source: Habr — хаб ИИ —
original