Problem
Business users in a pharma company need answers from clinical study data spread over dozens of linked tables and many documents. Dashboards cover only anticipated questions and multiply into a maintenance burden. Each tile is a hand-written SQL query, so a metric copied across tiles drifts apart. A chatbot over raw tables guesses joins and definitions and gives confident, almost-right answers, which pharma cannot accept.
Example from the article: after filtering a dashboard to one study, two numbers appear side by side, 41 subjects treated and 39 subjects, because one is planned and the other actual.
Solution in brief
Put a thin semantic layer of Databricks Metric Views (joins, measures, dimensions, descriptions) on tidy tables. Build a dashboard on it covering 50-70% of expected questions, and confine a Genie chatbot to the same views for the rest. Add a document assistant so one question can use both tables and protocols.
Questions this case answers
- How to let non-technical users self-serve trustworthy numbers from many linked tables while a chatbot answers only from governed metric definitions
- How to answer questions that need both table data and text documents such as study protocols
- How to let non-technical business users self-serve trustworthy numbers from many linked tables
- How to make a text-to-SQL chatbot (Genie) return correct answers instead of almost-right ones
Tools and methods
- Metric Views (Databricks) — joins, measures, dimensions and descriptions shared by the dashboard and Genie.
- AI/BI Dashboard — covers common questions and shows what the data contains.
- Genie space — natural-language questions, answered with the SQL it ran.
- Knowledge Assistant — answers questions from an indexed document corpus, with sources.
- Supervisor agent — routes a question to Genie, the Knowledge Assistant or both, and merges the answers.
Google or a one-time email code. No password.
