The challenge
The purchasing department of a major grocery chain was spending excessive time on manual data preparation for assortment decisions. Buyers could not run flexible, ad-hoc queries on assortment KPIs across product groups without technical support, and there was no standardised, scalable way to query third-party market-research data and internal transactional data together.
At a glance
LanguageHungarian natural language → SQL
DeploymentOn-premise LLM
FrontendResponsive web chat UI
What we did
SWICON unified market-research and internal assortment data into one queryable database spanning 200+ product groups, built a KPI calculation layer and documented the schema as machine-readable context for the model. On top sits an AI agent with a text-to-SQL engine that translates Hungarian natural-language questions into multi-join SQL, with schema-constrained generation to prevent invented fields, fuzzy semantic matching, result commentary and on-request charts. Because the dataset contains licensed data, the LLM runs on-premise: no cloud model, no third-party data sharing.
The result
Buyers query the unified database in plain Hungarian through a responsive chat interface. The pilot delivered natural-language querying on 4–6 product groups and 5–6 standardised KPIs, with hallucination prevention, complex multi-join query generation and an optional Power BI dashboard layer on the same database.
