Step 1: Define Core Questions + Build Deterministic MVP

• Identify ~30–50 core business questions (e.g. yield %, batch recovery £ vs standard).
• Build predefined SQL queries (Query Toolbox) with fixed parameters.
• Create interface to ask questions and return tabular or chart-based answers.
• Responses are traceable, deterministic, and use structured SQL only.
• No tagging or vector retrieval at this stage.

Step 2: Develop Tagging & Metadata Layer

• Build tagging schema: context, metric, product, scheme, date fields, etc.
• Implement admin UI for defining & reviewing tags (manual + auto-suggest).
• Store tag metadata in a central index (Postgres or JSONB).
• Link tags to SQL query templates to improve interpretability.
• Introduce semantic roles (measure, attribute, key) to guide field usage.

Step 3: Admin UI for Prompt Feedback + Training

• Store prompt → response pairs with accuracy feedback.
• Admin panel to:
  • Review prompt history and tools used.
  • Refine tagging, template logic, or create new queries.
  • Capture user corrections to improve future responses.