Problem
At BlaBlaCar, software engineers know the production code but treat the data warehouse as a black box because of SQL dialects, dbt syntax and Airflow operators. Data analysts, buried in "quick questions", rebuild business logic as ad-hoc SQL that ends up scattered across connected sheets and notebooks. Data quality errors surface only after a feature ships. A generic LLM knows SQL syntax but not company definitions: "search intent" meant something different in 2024 than in 2022.
Example from the article: an engineer asks "Show me the driver churn rate", and the assistant must reuse the data team's churn logic instead of inventing a definition.
Solution in brief
Data Copilot is an assistant inside the engineer's IDE. A Python script exports table schemas, golden (curated) queries and table samples from BigQuery into plain repository files, and GitHub Copilot finds them with its built-in search. Each answer is code that an analyst reviews in a pull request.
Questions this case answers
- How to let non-analyst domain experts get trustworthy answers from a warehouse when business terms are defined only in analysts' production queries
- How to ground an LLM assistant in curated company knowledge without building retrieval infrastructure
- How to let software engineers answer their own data questions without queuing behind analysts?
- How to make an LLM write SQL that uses our business definitions instead of guessing them?
Tools and methods
- GitHub Copilot in VS Code — LLM and retrieval layer (
#codebasesemantic search,#textSearchliteral search) - BigQuery — the warehouse and the source of schemas and samples
- dbt models and verified reports — source of the golden queries
- Python export script — writes context as Markdown, SQL and JSON files
- Data Health Card — heuristic soft warnings next to every analysis
Google or a one-time email code. No password.
