An agentic system that orchestrates Power BI and Snowflake through governed, multi-stage pipelines — from natural-language question to validated insight.
Each layer has a single responsibility and communicates only with its immediate neighbors.
VS Code · Web · Teams
Intent → Route → Build → Validate
List · Metadata · Generate · Validate · Execute · Export
Semantic Models · DAX
SQL API · Governed Execution
Markdown · Tables · Charts · Reports · Exports
Every user question flows through a governed, multi-stage pipeline.
Parse the user's natural-language question to determine intent type — KPI lookup, trend analysis, data exploration, or report generation.
Decide whether to query Power BI semantic models or Snowflake based on the intent and available metadata.
Construct a source-specific prompt with metadata, few-shot examples, and business glossary. Validate all generated queries before execution.
Run the validated query against the data source, format results with interpretation, and return actionable output.
Beyond the core pipeline — the capabilities that make this production-grade.
Short-lived user tokens exchanged for service credentials. No raw secrets in the extension.
Every generated query is validated against a policy engine before execution.
Full logging of every request — who asked, what ran, what was returned.
Metadata and frequent queries cached to reduce latency and API calls.
Embedding-based routing maps questions to the right data source automatically.
Domain terms resolved to exact column references — "PMPM" becomes the right measure.
Curated question→query pairs injected into prompts for higher accuracy.
Dangerous or ambiguous operations require explicit user approval before execution.
Plain English, query, data table, chart suggestion, and export — all in one response.
Conversation context maintained so follow-up questions resolve correctly.
Explore each layer in detail.