Problem
Wix runs a data mesh: each product domain (Stores, Customer Care, Restaurants and others) owns its tables and documentation. Together they hold hundreds of tables with thousands of dimensions and metrics. Business users often do not know which domain holds the data, so they depend on domain experts, and metadata is fragmented. Even with a wizard-based exploration tool (Data Playground) on a semantic layer, users could not pick the right metrics.
Example from the article: a user asks "how many users registered recently?", and the assistant must first learn what "recently" means before it can pick a metric.
Solution in brief
Wix built Anna, a multi-agent assistant that turns a chat question into a structured Data Playground query. The core is retrieval: an LLM writes example questions for every dimension and metric, those questions are embedded in Vespa, and the user's question is matched question-to-question. Retrieval cuts thousands of options down to a few for the query-building agent.
Questions this case answers
- How to map a free-text business question onto one of thousands of governed metrics and dimensions when table and column metadata is sparse and inconsistent across domains
- How to stop an LLM assistant from answering with an unrelated metric when the requested data does not exist
- How to match a business user's question to the right metric and dimension among thousands in a semantic layer?
- How to let non-technical people query a data-mesh warehouse without knowing which domain owns the data?
Tools and methods
- Vespa — vector database for similarity search over embedded questions
- Apache Airflow — hourly DAG that re-embeds the semantic layer
- Trino — query engine that runs the generated SQL
- Cube — semantic engine that turns structured payloads into SQL
- Question-to-question retrieval — LLM-generated questions per metric or dimension as the search index
- Multi-agent pipeline — root, validation, divider, query-building and retry agents
Google or a one-time email code. No password.