Design Brief: Conversational Analytics

Project Introduction

Build a conversational copilot inside the Client Portal that lets portal users ask plain-language questions and receive grounded answers, status lookups, and inline charts/tables. The experience includes a chat UI with inline renderers (tables, Chart.js visuals with copy/download), backed by an orchestrator (LLM + tools) that provides a safe SQL translator (parameterized, allow-listed views with row/time limits), tenant-scoped order/tracking lookups, and RAG over product documentation with citations. Guardrails include observability of prompts/responses, cost/latency, success/error tagging, and human-fallback paths. Out of scope: any state-changing actions, complex multi-hop analytics, external data joins, and proactive recommendations.

Users and Needs

• Portal users needing order/tracking answers → Quickly get trusted status without navigating multiple screens or opening tickets; identity-aware lookups scoped to their orders; results easy to copy/download. Constraints: tenant-scoped access; no state-changing actions.

• Non-technical portal users seeking analytics → Ask questions like “How many orders in the last 6 months?” and see inline charts/tables they can copy/download. Constraints: queries pass through a safe SQL translator with allow-listed views and row/time limits; no complex multi-hop analytics; no external data.

• Portal users with “how do I” questions → Self-serve answers via documentation with citations. Constraints: RAG is limited to product docs; human-fallback available when confidence is low or content is missing.

Tasks

1. Enter a plain-language question in the portal chat (e.g., counts, trends, groupings).

2. View a grounded response with inline visualization or table when applicable.

3. Retrieve order/tracking details via identity-aware lookup or by specific ID.

4. Ask support “how do I” questions and receive cited answers from product documentation.

5. Copy or download the returned table/chart from the inline renderer.

6. Use the human-fallback path when the copilot cannot answer or guardrails are triggered.

Timeline and Deliverables