What Now? (Co teď?)

AI-powered web platform for people caring for an elderly or ill loved one.

The project combines:

• guided intake form,
• legal-information retrieval (RAG),
• AI-generated personalized recommendations,
• interactive map of nearby social services.

Why This Project Exists

When caregiving starts suddenly, people usually do not know:

• what support they are entitled to,
• what practical steps to take first,
• which services are available nearby.

What Now helps users move from uncertainty to an actionable plan in a few minutes.

Product Workflow (End-to-End)

1. User opens frontend and fills a 4-step caregiving questionnaire.
2. Frontend sends form data to backend endpoint POST /api/v1/answer.
3. Backend converts answers into a compact retrieval query.
4. Backend performs semantic search over indexed legal chunks in PostgreSQL + pgvector.
5. Backend asks AI model for a grounded response based on retrieved sources.
6. Backend finds relevant social services (semantic + optional geo-radius filter).
7. Frontend renders the response (Markdown) and service map (Leaflet).
8. User can continue in chat (POST /api/v1/chat/message or WebSocket /api/v1/chat/ws).

High-Level Architecture

Frontend (Next.js)

• Multi-step form (/pomoc)
• Result view with AI response and map modal
• Dedicated map page (/mapa)
• Responsive UI for desktop and mobile

Backend (FastAPI)

• Laws ingestion and chunk indexing (JSON/PDF)
• Semantic search over legal text
• First-answer generation from form data
• Chat endpoint (HTTP + WebSocket)
• Social-service filtering and semantic matching

Data Layer

• PostgreSQL 16
• pgvector extension for embedding similarity
• Alembic migrations
• Main entities: law chunks, social services, service locations, target groups

Technology Stack

Frontend

• Next.js 16.1.6
• React 19.2.3
• TypeScript 5
• Tailwind CSS 4
• @tailwindcss/typography
• Leaflet 1.9.4
• React Leaflet 5.0.0
• react-markdown + remark-gfm
• Lucide React
• ESLint 9

Backend

• Python 3.13+
• FastAPI (async API)
• Uvicorn
• SQLAlchemy 2 (async)
• asyncpg
• pgvector
• Alembic
• Pydantic + pydantic-settings
• httpx
• sentence-transformers
• pypdf

Infrastructure / Tooling

• Docker + Docker Compose
• uv (Python package/dependency manager)
• PostgreSQL image: pgvector/pgvector:0.6.0-pg16

Optional/Experimental Parts in Repository

• tasks/ service (Celery, Playwright, RabbitMQ) prepared for async parsing workflows
• Services are currently commented out in docker-compose.yml and are not part of the default runtime
• _gemma4/ scripts in backend are experimental CLI utilities

Repository Structure

what-now/
├── docker-compose.yml
├── frontend/                  # Next.js application
├── backend/                   # FastAPI application
├── tasks/                     # Optional task workers (currently not active in compose)
└── test_docker.sh             # Convenience script for compose validation/start

API Overview

Base prefix: /api/v1

• GET /laws/health - laws slice health check
• POST /laws/ - upload/index law documents
• GET /laws/ - semantic law search
• POST /answer - first personalized answer from intake form
• POST /chat/message - one-shot chat response
• WS /chat/ws - realtime chat channel
• GET /map/services - paginated map services with filters
• GET /map/services/{source_service_id} - service detail
• POST /map/services/search - semantic + geo service search

Quick Start (Docker)

1. Prerequisites

• Docker + Docker Compose
• API credentials and model configuration for backend AI endpoint

2. Configure environment

Create backend environment file:

cp backend/.env.example backend/.env

If .env.example is not present, create backend/.env manually and provide at least:

# API
API_HOST=0.0.0.0
API_PORT=8001

# Database
POSTGRES_USER=backend
POSTGRES_PASSWORD=secret
POSTGRES_HOST=backend-db
POSTGRES_PORT=5432
POSTGRES_DB=whatnow

# AI
AI_MODEL_NAME=your-model-name
AI_SERVER_URL=https://your-ai-endpoint
AI_REQUEST_TIMEOUT=60

3. Run containers

docker compose --env-file backend/.env up -d --build

4. Open services

• Frontend: http://localhost:3000
• Backend API docs: http://localhost:8001/docs

5. Stop environment

docker compose --env-file backend/.env down

Local Development (Without Docker)

Backend

cd backend
uv sync
source .venv/bin/activate
alembic upgrade head
python main.py

Frontend

cd frontend
npm install
npm run dev

Data & Search Notes

• Legal text is chunked and stored with embeddings.
• Query and chunk vectors are compared using cosine distance in pgvector.
• Social-service matching combines semantic similarity with optional location radius.
• If user location is missing, semantic matching still works without geo filtering.

Current Scope and Limitations

• Primary product language is Czech.
• AI quality depends on configured external AI server/model.
• No active scheduler runtime flow is currently wired into backend runtime path.

Additional Documentation

• Frontend details: frontend/README.md
• Backend details: backend/README.md

License

This project is licensed under the MIT License. See the LICENSE file for details.