CONSIDER

CONSIDER is an AI-powered reflection platform for structured disagreement on polarised topics.

Instead of forcing consensus, CONSIDER helps users:

1. clarify their own position,
2. engage with principled counterarguments, and
3. reflect on where and why disagreement persists.

It is designed as a browser-based MVP with a modular LLM backend and post-conversation analysis outputs.

---

Features

• Three-stage flow: Clarification → Conversation → Contemplation
• Configurable disagreement intensity for discussion behaviour
• AI-generated counter-position profile tailored to oppose the user's stated stance
• Optional specific focus — ground the discussion in a concrete case rather than an abstract topic
• Structured post-conversation analysis:
  • key disagreements and agreements
  • potential common ground
  • nature and root of the disagreement
  • moral foundations scoring (Haidt)
  • epistemic humility rating
• Reviewable discussion transcript inline on the results page
• Multi-provider LLM support (GPT, Llama from TogetherAI & OpenRouter)
• Frontend + backend separation for easy iteration

---

Tech Stack

Layer	Technology
Frontend	React + Vite
Styling	Tailwind CSS + Framer Motion
Backend	Node.js + Express
Database	MongoDB
Chat model	OpenAI (gpt-4o-mini) or Llama (Llama-3.3-70B) - set to Llama in this repo
Analysis model	Claude (claude-sonnet-4-5) via OpenRouter + gpt-4o for extraction

---

Repository Structure

CONSIDER/
├── frontend/                        # React/Vite client
│   ├── src/
│   │   ├── App.jsx                  # Routes
│   │   ├── api.js                   # Axios API calls
│   │   ├── Home.jsx                 # Landing page
│   │   ├── Access.jsx               # Password gate
│   │   ├── DebateIntro.jsx          # Intro screen
│   │   ├── RVDSelect.jsx            # Topic selection
│   │   ├── toggle.jsx               # Disagreeability + focus config
│   │   ├── CurrentPosition.jsx      # Position clarification chat
│   │   ├── Chat.jsx                 # Discussion chat interface
│   │   ├── Results.jsx              # Post-conversation analysis
│   │   └── components/
│   │       └── chat/
│   │           ├── ChatInput.jsx
│   │           └── MessageBubble.jsx
│   └── package.json
│
└── backend/                         # Express API + model orchestration
    ├── server.js                    # Entry point, middleware, route mounting
    ├── connect.js                   # MongoDB connection
    ├── chatRoutes.js                # /chat and /generate-profile endpoints
    ├── prompts.js                   # Prompt builders
    ├── analysisRoutes.js            # /analyze-conversation endpoint
    ├── analysisPrompt.js            # Analysis prompt builders
    ├── config.env.example           # Environment variable template
    └── package.json

---

User Flow

/access           → password gate
/                 → landing page
/debate-intro     → explains the format
/select-rvd       → user picks a topic
/toggle           → sets disagreeability level + optional specific focus
/current-position → clarification chat (AI extracts user's position)
/play             → disagreement chat (AI argues the opposing side)
/results          → full post-conversation analysis

---

Getting Started

Prerequisites

• Node.js
• A MongoDB Atlas account (or local MongoDB instance)
• API keys for at least one LLM provider (see below)

---

1. Clone the repository

git clone https://github.com/BillyhfPsyc/CONSIDER.git
cd CONSIDER

---

2. Backend setup

cd backend
npm install

Copy the environment variable template:

cp config.env.example config.env

Then fill in config.env:

# MongoDB
ATLAS_URI=mongodb+srv://<user>:<password>@cluster.mongodb.net/Consider

# LLM provider for chat ('openai' or 'together')
LLM_PROVIDER=openai

# OpenAI (required if LLM_PROVIDER=openai)
OPENAI_API_KEY=sk-...

# Together AI (required if LLM_PROVIDER=together)
TOGETHER_API_KEY=...

# OpenRouter (required for analysis — always used regardless of LLM_PROVIDER)
OPENROUTER_API_KEY=...

# Server
PORT=3001

# CORS — set to your frontend's origin
FRONTEND_ORIGIN=http://localhost:5173

Start the backend:

npm start

The server will run on http://localhost:3001 by default.

---

3. Frontend setup

cd frontend
npm install

Create a .env.local file:

VITE_API_URL=http://localhost:3001
VITE_SITE_PASSWORD=your-chosen-password

Start the frontend:

npm run dev

The app will be available at http://localhost:5173.

---

API Reference

Chat

Method	Endpoint	Description
POST	/chat	Send a message in either clarification or disagreement context
POST	/generate-profile	Generate a fictional AI opponent profile based on the user's position

Analysis

Method	Endpoint	Description
POST	/analyze-conversation	Run full post-conversation analysis on a completed discussion

Other

Method	Endpoint	Description
GET	/health	Health check (used by Render keepalive)

---

LLM Providers

CONSIDER is designed to work with multiple LLM providers:

Chat (discussion + clarification):

• openai — uses gpt-4o-mini
• together — uses meta-llama/Llama-3.3-70B-Instruct-Turbo

Switch between them with the LLM_PROVIDER environment variable.

Analysis: Always routed via OpenRouter. Uses:

• anthropic/claude-sonnet-4-5 for philosophical analysis
• openai/gpt-4o for structured data extraction

---

Environment Variables — Full Reference

Backend (backend/config.env)

Variable	Required	Description
ATLAS_URI	Yes	MongoDB connection string
OPENAI_API_KEY	If LLM_PROVIDER=openai	OpenAI API key
TOGETHER_API_KEY	If LLM_PROVIDER=together	Together AI API key
OPENROUTER_API_KEY	Yes	OpenRouter key (for analysis)
LLM_PROVIDER	No	openai (default) or together
PORT	No	Server port (default: 3001)
FRONTEND_ORIGIN	Yes	Frontend URL for CORS

Frontend (frontend/.env.local)

Variable	Required	Description
VITE_API_URL	No	Backend URL (default: http://localhost:3001)
VITE_SITE_PASSWORD	No	Site access password (default: test-password)

---

Deployment

The project is structured for straightforward deployment:

• Frontend — deploy the frontend/ directory to Vercel. Set VITE_API_URL to your deployed backend URL in the Vercel environment settings.
• Backend — deploy the backend/ directory to Render as a Node.js web service. Add all config.env variables as environment variables in the Render dashboard. The /health endpoint is used for keepalive pings.

---

CONSIDER can be accessed at: https://consider-chi.vercel.app

The application is currently password protected.

If you would like access, please contact
david.lyreskog@psych.ox.ac.uk
with a request for the password and a brief description of your intended usage..

Licence

© 2026 Design Bioethics Lab, Department of Psychiatry, University of Oxford. Licensed under CC BY-NC 4.0 — free to use and adapt for non-commercial purposes with attribution.