AnythingLLM Setup: RetailFlow Virtual Staff

How to turn the text personas in this folder into embedded chatbots for the AI in Delivery course. Do this once per persona. Budget ~10–15 min per bot the first time.

What you need per bot

Each bot = one AnythingLLM workspace with three knowledge documents + a system prompt + an embed.

Documents to upload to the workspace (_backstories/): 1. retailflow_company_overview.md, shared company facts (every bot) 2. delivery_scenario_context.md, shared delivery framing (every bot, for this course) 3. The persona’s own backstory, e.g. priya_sharma_data_analytics.md

System prompt: paste the contents of that bot’s bots/<name>/prompt.txt, then append the line below so the bot is in delivery mode:

DELIVERY MODE: The board has already funded the AI initiatives. The learner is a delivery lead figuring out how to ship one of them. Answer their scoping, data, stakeholder, human-in-the-loop, roadmap and risk questions from your role. Disagree with other executives where your priorities differ. If you don’t know a precise figure or policy, say so. Never invent one.

Steps (per bot)

Create workspace named after the person (e.g. “Priya Sharma, Data”).
Upload the three documents above; embed/save so they’re in the workspace’s vector store.
Paste the system prompt (prompt.txt contents + the DELIVERY MODE line) into the workspace’s chat/system prompt setting.
Set chat mode to “query”/“chat” as preferred; keep temperature moderate so it stays in character.
Create an embed (Embeddable Chat Widgets → New embed → select this workspace). Copy the <script> snippet AnythingLLM generates.
Paste the snippet into bots/<name>/index.qmd, replacing this placeholder block:

Chatbot loading...
Re-render the Quarto site and check the bot page loads the widget.

Priority order for Friday

Stand these up first (the core afternoon stakeholders): Priya, Marcus, Emma, Tom Walsh, David. Sarah and Lisa are optional/nice-to-have.

If you only have time for the morning demo, one bot is enough (Tom Walsh or Priya) to do the “watch it confidently hallucinate a policy” moment. The afternoon can fall back to the text personas as printed briefing cards (see the course’s facilitator quick reference).

Guardrail reminder

The delivery_scenario_context.md already tells every bot to stay in character, admit when it doesn’t know, and avoid inventing precise figures or quoting model versions/stats. If a bot drifts, tighten the system prompt rather than the documents.