tencentmusic/supersonic

What it solves

SuperSonic addresses the reliability issues of traditional "Chat BI" (Text2SQL) systems, which often struggle with hallucinations and complex SQL syntax when used in large-scale real-world applications. It bridges the gap between natural language querying and structured data by combining LLM-powered chat interfaces with a governed semantic layer.

How it works

The platform unifies two paradigms: Chat BI for natural language interaction and Headless BI for defining semantic data models (metrics, dimensions, and tags).

When a user asks a question, SuperSonic uses a pipeline to process the request:

1. Knowledge Base & Schema Mapper: Matches user query text against a dictionary of schema elements.
2. Semantic Parser: Uses a mix of rule-based and LLM-based logic to generate a semantic query statement.
3. Semantic Corrector: Validates and fixes the statement using rules or LLMs.
4. Semantic Translator: Converts the validated semantic statement into executable SQL, offloading complex join and formula logic from the LLM to the semantic layer to reduce errors.

Who it’s for

• Business Users: Who want to query data and generate visualizations using natural language without knowing SQL.
• Analytics Engineers: Who need to build and maintain governed semantic data models that ensure consistency across the organization.

Highlights

• Hybrid Parsing: Combines rule-based and LLM-based parsers for better efficiency and reliability.
• Reduced Hallucinations: Incorporates business terms and column values from the semantic model into prompts to guide the LLM.
• Granular Access Control: Supports data access restrictions at the dataset, column, and row levels.
• Extensible Architecture: Allows custom implementations via Java SPI, including chat plugins and memory for few-shot prompting.