---
title: P3 Myeloma Follow-up Chatbot
emoji: 🩺
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
---

# P3 Myeloma – AI Follow-up Chatbot

An adaptive follow-up chatbot for the AI-Enhanced Personalized Patient Profile (P3)
Myeloma intake. The participant answers the 16 Shared Decision Making questions, and
the app uses Google Gemini to ask tailored follow-up questions, gently probe apparent
contradictions in their preferences, and end with a priorities summary they can bring
to their care team.

It does **not** give medical advice, diagnoses, or treatment recommendations. It only
helps the patient articulate their own values and concerns.

## Setup on Hugging Face Spaces

1. Create a new Space, choose **Gradio** as the SDK.
2. Upload `app.py`, `knowledge_base.py`, `supabase_store.py`, `Archive_2.zip`,
   `requirements.txt`, `packages.txt`, and this `README.md`.
3. Add your Gemini key as a secret:
   - Go to **Settings → Variables and secrets → New secret**
   - Name: `GEMINI_API_KEY`
   - Value: your key from https://aistudio.google.com/apikey
4. Add the Supabase secrets described below.
5. The Space will build and launch automatically.

## Saving and resuming sessions (Supabase)

Participants sign in with a **User ID** and an **email address**. Everything they
do is saved against that email, so closing the page and returning later — even on
another device — picks up exactly where they left off. The User ID is checked on
every return visit, so entering someone else's email does not open their session.

### One-time setup

1. **Create the tables.** In Supabase, open **SQL Editor → New query**, paste the
   whole of `supabase_schema.sql`, and run it. It is idempotent, so re-running it
   after an update is safe.
2. **Copy your credentials.** In Supabase, go to
   **Project Settings → API** and copy the **Project URL** and the
   **`service_role`** key.
3. **Add them as Space secrets** (Settings → Variables and secrets):

   | Name | Value |
   |---|---|
   | `SUPABASE_URL` | `https://YOUR-PROJECT-REF.supabase.co` |
   | `SUPABASE_KEY` | your **service_role** key |

   Use the `service_role` key, not the `anon` key. Row Level Security is enabled
   with no policies, so the `anon` key can read and write nothing — which is the
   point. Gradio runs your Python on the server, so the key is never sent to a
   browser.
4. **Check it.** Locally, or in the Space terminal:

   ```bash
   SUPABASE_URL="https://YOUR-PROJECT-REF.supabase.co" SUPABASE_KEY="YOUR-SERVICE-ROLE-KEY" python3 verify_supabase.py
   ```

   It writes a throwaway row, reads it back, and deletes it. It never touches
   real participant data.

If the Supabase secrets are missing or the service is unreachable, the chatbot
still runs — it just tells the participant that this visit will not be saved,
and it will not overwrite anything already stored for them.

### What gets stored

`p3_sessions` — one resumable row per email: the participant's User ID, the 16
answers, the conversation they see, the full model-facing conversation, and the
follow-up counter.

`p3_messages` — an append-only log of every turn. Session rows are replaced when
a participant starts over; this log never is, so no transcript is lost.

For a quick look at what has been collected, open the `p3_session_overview` view
in the Table Editor.

### What participants see

- They are asked for their **User ID first, then their email address**, before
  anything else. Both are stored in Supabase.
- Their User ID is shown in the **top right of the page** for as long as they are
  signed in, so they can note it for next time.
- If a returning participant enters the wrong User ID for an email, they are
  refused. The **second** wrong attempt tells them the User ID is displayed at
  the top right of the page.
- The conversation **opens with all 16 questions and their own answers**, so the
  profile the follow-up is built on is visible without leaving the chat.
- The questionnaire auto-saves as it is filled in, so a half-finished intake
  survives a closed tab.
- **Start Over** first shows a warning that it erases the current chat history and
  starts a new session, and only erases anything once confirmed. The transcript
  log in `p3_messages` is kept either way. **Use a different email** just signs
  out; nothing is deleted.
- If the same session is open in two places, the older one is told to reload
  rather than being allowed to overwrite the newer conversation.

### Changing a participant's User ID

```sql
update public.p3_sessions set user_id = 'P3-NEW' where email = 'someone@example.com';
```

A row created before User IDs existed has none stored; the first sign-in after
that adopts whatever User ID is entered, and it is enforced from then on.

### Deleting a participant's data

```sql
delete from public.p3_sessions where email = 'someone@example.com';
delete from public.p3_messages where email = 'someone@example.com';
```

## Model

Uses `gemini-2.5-flash`. To change it, edit the `MODEL` variable in `app.py`.

## How it works

- The 16 yes/no answers are turned into a structured profile.
- A rule-based `detect_tensions()` step flags likely contradictions (for example,
  wanting an aggressive approach while prioritizing quality of life, or living alone
  with no caregiver but preferring home medications).
- Those flags are passed to Gemini, which asks one follow-up question per turn.
- After enough exchanges (capped at 8), the model returns a `<<SUMMARY>>` of the
  patient's priorities, which the app formats and displays.

## Disclaimer

This tool is for informational and preparation purposes only and is not a substitute
for professional medical advice. Always consult a qualified healthcare provider.