LLM-powered BI that checks the semantic layer before guessing SQL

What it does

SuperSonic is a Java-based analytics platform that lets business users ask questions in plain English and get charted answers. Under the hood, analytics engineers define metrics, dimensions, and relationships in a Headless BI semantic layer. The system then uses that curated model to interpret the natural language query and generate the SQL, rather than asking the LLM to wing it from raw schema names.

The interesting bit

The project treats the LLM as a translator working with a glossary, not a database expert working from memory. It feeds business terms and column values from the semantic layer into the prompt to reduce hallucination, and offloads joins and formulas to the semantic layer so the LLM only handles the easy parts.

Key highlights

• Natural-language Chat BI for business users, paired with a Headless BI semantic-modeling interface for analytics engineers.
• Hybrid semantic parser: rule-based engines handle deterministic scenarios, LLM-based engines handle the fuzzy ones.
• Schema mapper and knowledge base keep the LLM grounded in actual business terms rather than raw database metadata.
• Built-in multi-turn chat, auto-completion, and post-query recommendations.
• Dataset-, column-, and row-level access controls.

Caveats

• The public demo resets every weekend, so do not treat it as a persistent sandbox.
• Build-and-development documentation links are Chinese-language, which may slow down non-Chinese contributors.

Verdict

Worth a look if you need governed natural-language analytics and have the engineering bandwidth to maintain a semantic layer. Skip it if you just want a lightweight LLM-to-SQL proxy without the BI infrastructure.