uygnoey Agentic analytics for data warehouses
Instructions
Mental model: analytics is a context + verification problem
Analytics questions often have a single correct answer tied to a specific, up-to-date entity in a data model, so correctness depends on mapping the question to the right entities, sources, and definitions.
Guardrails for the three dominant failure modes
- Concept–entity ambiguity: clarify what entity and grain the user means.
- Data staleness: confirm the “as-of” point and freshness lag conventions; avoid assuming “yesterday” is available.
- Retrieval failure: always consult the intended sources of truth (semantic layer, lineage graph, governed docs) before writing raw SQL.
Semantic-layer-first workflow (required)
Follow the warehouse workflow template in:
templates/warehouse_skill_skeleton.md
Use the domain reference-doc format in:
references/reference_doc_skeleton.md
Validation and reporting
- Separate observations (what data shows) from interpretations (what it may imply).
- Never invent columns, tables, or results.
- End every final answer with a provenance footer (see the template).
Examples
- Use
templates/warehouse_skill_skeleton.mdas a starting point for a production warehouse skill. - Use
references/reference_doc_skeleton.mdto create governed domain reference docs (grain, key tables, gotchas, query patterns, cross-references).