Assist Chatbot (Single Next.js App)

Unified Next.js + TypeScript + Prisma + PostgreSQL chatbot platform with multi-tenant bot management and embeddable widget.

Architecture

• Next.js App Router (UI + API Route Handlers)
• Prisma ORM + PostgreSQL
• JWT auth
• AES-256-GCM encryption for per-bot AI keys
• In-memory rate limit
• Vanilla JS widget (public/widget.js) with Shadow DOM

Gemini Defaults

• Default model: gemini-2.5-flash
• Default provider base URL: https://generativelanguage.googleapis.com/v1beta/models
• API key header to provider: x-goog-api-key

Pages

• /login
• /dashboard
• /bots
• /bots/[id]/edit
• /bots/[id]/style
• /bots/[id]/knowledge

API Routes

Auth

• POST /api/auth/register
• POST /api/auth/login

Bots (JWT required)

• GET /api/bots
• POST /api/bots
• GET /api/bots/[id]
• PATCH /api/bots/[id]
• GET /api/bots/[id]/models (checks available Gemini models using this bot's API key)

Knowledge (JWT required)

• GET /api/bots/[id]/knowledge
• POST /api/bots/[id]/knowledge
• POST /api/bots/[id]/knowledge/import (upload PDF/Excel and convert into knowledge entries)
• PATCH /api/bots/[id]/knowledge/[knowledgeId]
• DELETE /api/bots/[id]/knowledge/[knowledgeId]

Widget / Chat (public)

• GET /api/widget/config/[id]
• POST /api/chat
• OPTIONS /api/chat (CORS preflight)
• OPTIONS /api/widget/config/[id] (CORS preflight)

Setup

1. Install dependencies:

npm install

2. Create env file:

cp .env.example .env

3. Start PostgreSQL:

docker compose up -d

4. Run migrations:

npx prisma migrate dev --name init

5. Run app:

npm run dev

Widget Embed

<script src="https://your-domain.com/widget.js" data-bot-id="YOUR_BOT_ID" data-api-base-url="https://your-domain.com"></script>

Security

• Encrypted API keys at rest
• Domain whitelist check for widget config and chat
• Per-bot rate limiting in memory
• Prompt scope control with role + topics + knowledge

Knowledge File Import

• Upload PDF or Excel (.xlsx, .xls, .csv) in /bots/[id]/knowledge
• Excel is generally better for structured store data (products, prices, inventory)
• PDF is useful for long-form descriptions, policies, and personal site context
• Imported content is chunked into multiple knowledge rows automatically