Stop Cyberviolence

Data for Good - Saison 13

Last updated: 2026-04-08

This page contains an introduction and the technical documentation for the project Stop Cyberviolence as a part of season 13 hosted by Data for Good France.

What is Monstermessenger?

Monstermessenger is a multilingual AI-powered chatbot designed to help young people (10–13 years) and parents address cyberbullying and online safety concerns. It supports two user audiences and six languages, offering context-aware conversations backed by curated educational documents.

The system is built on LangGraph, FastAPI, and React, with a persistent session layer, an IDP-based authentication flow, and remote feature-flag management via OpenFeature.

Dual-Variant Architecture

A single codebase serves two audiences, selected at deployment time via CHATBOT_VARIANT:

Variant Target Address style
teenager Children/teens (10–13 years) Direct (“tu”, “you”)
parent Parents of children Formal (“votre enfant”, “your child”)

Each variant gets its own:

  • Set of context-collection questions and system prompts (via the i18n layer)
  • RAG knowledge-base collection (docs_youth / docs_adult)
  • Frontend deployment (separate GCS bucket)
  • Feature-flag context (some flags are variant-specific)

Key Features

🗣️ Multilingual Support (6 Languages)

French (FR, default), English (EN), Spanish (ES), Italian (IT), German (DE), Portuguese (PT).
All prompts, questions, and UI strings are in YAML files under api/i18n/, with a two-layer in-memory cache.

🤖 Intelligent Conversation Flow

  1. Context Collection — Structured questions gather the situation. An optional AI-driven dynamic follow-up mode (feature-flagged) generates additional questions based on the user’s answers.
  2. AI Advice — The give_advice node, driven by gemini-3.1-flash-lite-preview, produces empathetic, age-appropriate guidance.
  3. RAG — Educational documents are retrieved from PGVectorStore using multi-query parallelised search, filtered by an LLM relevance score, and injected into the advice prompt. A two-level semantic cache (embedding cache + cosine-similarity query cache) keeps latency low.
  4. Ongoing Support — Follow-up conversation supported by a rolling summary persisted via the MemoryLake async queue.

🔐 Identity & Session Management

  • IDP Authentication: Clients obtain a signed token from an Identity Provider and exchange it at POST /auth/login for an application session.
  • Persistent Profiles: User, AppSession, Conversation, and IDPLogin records are stored in PostgreSQL and eagerly loaded through a repository pattern (AppDataService).
  • Session Resumption: Returning users are offered the Simple Resume Prompt (floating one-click card) or the full Conversation History Modal — controlled by the enable_simple_resume_prompt feature flag.

🎛️ Feature Flags (OpenFeature)

Runtime toggle of UI and backend behaviour via GET /features. The ApiFeatureProvider on the frontend fetches flags on startup. See Backend API for the full flag catalogue.

📱 Modern Tech Stack

Layer Technologies
Backend FastAPI, LangGraph, SQLAlchemy async, Alembic, psycopg3
AI Google Gemini (via langchain_google_genai), LangGraph tool nodes
Data PostgreSQL (multi-schema: app, langgraph, rag, cache)
Frontend React 18, TypeScript, Vite, Tailwind CSS, OpenFeature
Deployment Google Cloud Run (backend) + Cloud Storage (frontend), GitHub Actions

System Architecture

┌───────────────────┐   ┌────────────────────┐   ┌──────────────────────┐
│   React Frontend  │   │  FastAPI Backend   │   │  LangGraph AI Agent  │
│                   │   │                    │   │                      │
│ • AuthContext     │   │ • /auth/login      │   │ • collect_context    │
│ • BrandScreen     │◄──┤ • /chat/init       ├──►│ • give_advice        │
│ • MonsterAvatar   │   │ • /chat            │   │ • research_tools     │
│ • HistoryModal    │   │ • /features        │   │ • ongoing_support    │
│ • OpenFeature     │   │ • /chat/history    │   │ • summarize          │
└───────────────────┘   └─────────┬──────────┘   └──────────────────────┘
                                  │
                    ┌─────────────▼─────────────┐
                    │        PostgreSQL           │
                    │  schema: app               │
                    │    Users, AppSessions,     │
                    │    Conversations,          │
                    │    FeatureFlags            │
                    │  schema: langgraph         │
                    │    Checkpoint store        │
                    │  schema: rag               │
                    │    PGVectorStore            │
                    │  schema: cache             │
                    │    Semantic query cache    │
                    └────────────────────────────┘

Project Structure (Top-Level)

stopcyberviolences-chatbot/
├── api/                    # Backend — FastAPI + LangGraph
│   ├── agents/service1/    # Conversation agent (nodes, memory, tools, manifests)
│   ├── i18n/               # YAML prompts, questions, UI strings
│   ├── database/           # Multi-schema models, engines, Alembic
│   ├── services/           # AppDataService, FeatureFlagService, RAG, Chat
│   ├── routers/            # auth, chat, features, contacts, files, feedback
│   ├── core/               # Logging, dependency singletons
│   └── main.py
├── frontend/               # React SPA
│   └── src/
│       ├── components/     # BrandScreen, MonsterAvatar, HistoryModal, …
│       ├── context/        # AuthContext
│       ├── providers/      # ApiFeatureProvider (OpenFeature)
│       └── hooks/          # useChat
├── rag/                    # Document indexing scripts + raw PDFs
├── deployment/             # deploy.py CLI, Dockerfiles, shell scripts
├── discord_bot/            # Discord bot client
└── doc/                    # This documentation (Quarto .qmd files)

Use the sidebar to navigate between documentation sections.