business-question-answering

name	business-question-answering
description	Translate business questions into metrics, SQL evidence, and actionable recommendations
tags	["business","metrics","recommendation","analysis"]
version	1.0.0

Business Question Answering

Use this workflow for questions like "where is revenue strongest?", "what are the opportunities?", "which segment should we prioritize?", or "what changed?".

Workflow

1. Clarify silently when possible

   • Infer likely metric, dimension, grain, and time window from available schema and context.
   • Ask the user only when multiple materially different business definitions would change the answer.

2. Get evidence

   • If the data lives in a connected database, load database-analyst and use list_tables/describe_table/execute_sql.
   • Before writing ad-hoc SQL, always call list_sql_pairs to check if an authoritative SQL pair already exists for the question.
   • IMPORTANT: list_sql_pairs returns a file-level preview only. You MUST call read_sql_pair for each potentially relevant file to inspect ALL individual pairs before concluding there is no match.
   • When a SQL pair matches the user's question (by intent or semantic similarity), you MUST call execute_sql_pair to run the pair's SQL as-is. Do NOT rewrite or substitute columns from the pair.
   • Only fall back to execute_sql when no matching SQL pair exists.
   • Do not answer quantitative questions without at least one SQL query when SQL tools are available.

3. Validate

   • Run a total/check query when recommendations depend on shares or rankings.
   • Compare at least two cuts when the user asks for opportunities, e.g. segment and region, or current and prior period.

4. Recommend

   • Separate facts from interpretation.
   • Tie every recommendation to a number from the query result.
   • Include caveats about sample size, filters, and grain.

Output shape

Use this compact structure unless the user asks otherwise:

1. Answer — one sentence with the main finding.
2. Evidence — small table or bullets with actual numbers.
3. Recommendation — 1–3 actions.
4. Caveat / next check — one line.