🦙 Sandbox-NotebookLM

Enterprise-ready alternative to Google NotebookLM with multi-user support, powered by LlamaIndex and multiple AI models.

Live Demo: https://ai-sandbox.oliver.solutions/notebookllama/ Repository: https://bitbucket.org/zlalani/sandbox-notebookllamalm-nextjs

---

🚀 Quick Start (Docker)

Prerequisites

• Docker and Docker Compose
• Git

1. Clone

git clone git@bitbucket.org:zlalani/sandbox-notebookllamalm-nextjs.git
cd sandbox-notebookllamalm-nextjs

2. Configure environment

Create backend/.env:

# Core (required)
OPENAI_API_KEY=sk-...
LLAMACLOUD_API_KEY=llx-...

# Model-specific (optional)
ANTHROPIC_API_KEY=sk-ant-...
GOOGLE_API_KEY=AI...
GROQ_API_KEY=gsk_...
DEEPSEEK_API_KEY=sk-...

# Podcast generation (optional)
ELEVENLABS_API_KEY=...

# Database
pgql_user=postgres
pgql_psw=admin
pgql_db=postgres_nextjs

# Microsoft SSO (optional)
AZURE_CLIENT_ID=your-client-id
AZURE_AUTHORITY=https://login.microsoftonline.com/your-tenant-id
AZURE_REDIRECT_URI=https://your-domain.com/notebookllama/

Create frontend/.env.production (baked into build):

NEXT_PUBLIC_API_URL=https://your-domain.com/notebookllama-back
NEXT_PUBLIC_WS_URL=wss://your-domain.com/notebookllama-back
NEXT_PUBLIC_AZURE_CLIENT_ID=your-client-id
NEXT_PUBLIC_AZURE_AUTHORITY=https://login.microsoftonline.com/your-tenant-id
NEXT_PUBLIC_AZURE_REDIRECT_URI=https://your-domain.com/notebookllama/

For local dev, use frontend/.env.local:

NEXT_PUBLIC_API_URL=http://localhost:9000
NEXT_PUBLIC_WS_URL=ws://localhost:9000

3. Build and start

docker compose up -d --build

This starts: backend (port 9000), frontend (port 4000), PostgreSQL (port 5433), Redis (port 6380).

4. Initialize database

On first run:

docker compose exec backend /app/.venv/bin/python -c \
  "import sys; sys.path.insert(0, '/app/src/notebookllama'); from database import init_db; init_db(); print('Done')"

---

🌐 Access Points

Service	URL
Frontend	http://localhost:4000
Backend API	http://localhost:9000
API Docs (Swagger)	http://localhost:9000/docs
Health Check	http://localhost:9000/api/health

---

📚 Features

Core

• Multi-Notebook Management — organize documents into collections
• 6 AI Models — GPT-5, Claude 4.5, Gemini 2.5 Pro, GPT-4o, Gemini Flash, GPT-4
• 40+ File Formats — PDF, DOCX, PPTX, XLSX, CSV, images (OCR), audio (transcription), video (multimodal)
• Background Processing — non-blocking uploads with real-time status
• Document Summaries — AI-generated summaries, highlights, Q&A pairs

Analysis

• Cross-Document Synthesis — themes, insights, comparative findings across all docs (persists to DB)
• Studio — 7 additional output types generated from your documents:
  • Flashcards — 15-20 study cards with 3D flip animation
  • Quiz — 10-12 multiple choice questions with scoring
  • Mind Map — SVG radial tree visualization
  • Slide Deck — 8-12 slide presentation with PPTX download; slide diagrams (flowcharts, bar charts) rendered in preview and exported; supports custom .pptx template upload (Claude analyzes and rebuilds it) and per-slide AI editing without regenerating the full deck
  • Report — executive summary + sections + conclusions with PDF download
  • Infographic — visual blocks with stats and emojis
  • Data Table — structured comparison table
  • All Studio modules accept a custom prompt to guide content generation and style

Chat

• Real-time WebSocket Chat — ask questions across all documents
• Multiple Sessions — organize conversations, rename, share, delete
• Source Citations — see which documents were used
• Markdown Rendering — formatted AI responses

Podcast

• AI-Generated Podcast — two-voice audio discussion from your documents
• Customizable — length (5-30 min), voices, theme, instructions
• Built-in Player — listen or download directly

Collaboration

• Notebook Sharing — share by email with Read / Write / Share permissions
• Shared Sessions — make chat sessions visible to collaborators

Admin

• System stats, cost estimation, user management, task monitoring

Auth

• Local email/password + Microsoft SSO (MSAL PKCE)
• Role-based access (admin / regular user)

---

🎯 AI Models

Model	Provider	Input	Output
GPT-5	OpenAI	$1.25/1M	$10/1M
Claude 4.5	Anthropic	$3/1M	$15/1M
Gemini 2.5 Pro	Google	$1.25/1M	$5/1M
GPT-4o	OpenAI	$5/1M	$15/1M
Gemini Flash	Google	$0.075/1M	$0.30/1M
GPT-4	OpenAI	$30/1M	$60/1M

---

🏗️ Architecture

Tech Stack

Frontend: Next.js 15 (App Router), React 19, TypeScript, Tailwind CSS 4, React Query, Zustand, MSAL, WebSocket, CSS custom property theme system (light/dark)

Backend: FastAPI, SQLAlchemy, Python 3.13, uv package manager

AI/ML: LlamaCloud (indexing), LlamaIndex (RAG), LlamaParse (parsing), LlamaExtract (structured extraction), Gemini 2.5 Pro (video multimodal), ElevenLabs (voice), python-pptx, weasyprint

Infrastructure: Docker Compose, PostgreSQL, Redis

Database Schema (10 Tables)

Table	Key columns
users	email, username, is_admin, auth_provider
notebooks	name, model_type, synthesis_data, studio_data, podcast_path
documents	filename, llamacloud_file_id
notebook_documents	notebook_id, document_id
document_summaries	summary, highlights, questions, answers
chat_sessions	title, is_shared, notebook_id
chat_messages	role, content, sources
document_shares	permission_level (READ/WRITE/SHARE/ADMIN)
background_tasks	status, task_type

---

🔌 API Endpoints

Full docs: http://localhost:9000/docs

Notebooks:

• GET/POST /api/notebooks/ — list / create
• GET/PUT/DELETE /api/notebooks/{id} — get / update / delete
• POST /api/notebooks/{id}/synthesis — generate cross-doc analysis
• GET /api/notebooks/{id}/synthesis — get saved synthesis
• POST /api/notebooks/{id}/podcast — start podcast generation
• POST /api/notebooks/{id}/share — share with user
• GET /api/notebooks/{id}/studio — get all saved Studio outputs
• POST /api/notebooks/{id}/studio/{type} — generate (flashcards / quiz / mindmap / slides / report / infographic / datatable)
• GET /api/notebooks/{id}/studio/slides/download — PPTX file
• GET /api/notebooks/{id}/studio/report/download — PDF file
• GET /api/notebooks/{id}/studio/mindmap/download — SVG file
• POST /api/notebooks/{id}/studio/slides/from-template — generate PPTX from uploaded template (multipart)
• POST /api/notebooks/{id}/studio/slides/edit/{index} — regenerate single slide via prompt

Documents:

• POST /api/documents/upload/{notebookId} — upload file
• GET /api/documents/task/{taskId} — task status
• GET /api/documents/{documentId}/summary — get analysis

Chat:

• WS /api/chat/ws/{notebookId}?session_id={id} — real-time chat
• GET /api/chat/{notebookId}/sessions — list sessions
• POST /api/chat/{notebookId}/sessions — create session

Auth:

• POST /api/auth/signup — register
• POST /api/auth/login — login
• POST /api/auth/microsoft — SSO login

Note: All download endpoints require the JWT token via Authorization: Bearer header. The frontend uses fetch() with the auth header for all binary downloads.

---

🔐 Microsoft SSO Setup

1. Register Azure AD app (SPA with PKCE, User.Read permission)
2. Add redirect URI matching NEXT_PUBLIC_AZURE_REDIRECT_URI
3. Set AZURE_CLIENT_ID, AZURE_AUTHORITY, AZURE_REDIRECT_URI in backend/.env
4. Set matching NEXT_PUBLIC_* vars in frontend env

SSO users are auto-created on first login. Local accounts are merged if the same email logs in via SSO.

---

📖 Usage Guide

Create a notebook:

1. My Notebooks → New Notebook → choose name + AI model

Upload documents:

1. Open notebook → Select Files → wait for processing (~1 min/doc, ~1 min/10min of video)

Studio:

1. Open notebook with processed documents
2. Scroll to Studio section → click any card (Flashcards, Quiz, etc.)
3. Results load or generate fresh; PPTX/PDF download available for Slides/Report

Cross-doc analysis:

1. Click "Cross-Doc Analysis" → wait 30-60s → results persist

Podcast:

1. Click "Podcast" → choose length + voices → "Generate" → wait 3-5 min

Chat:

1. Click "Chat" → create/select session → ask questions

---

🐛 Troubleshooting

Backend 500 on all routes after deploy:

docker compose build backend && docker compose up -d backend

column notebooks.studio_data does not exist:

docker compose exec backend /app/.venv/bin/python -c \
  "import sys; sys.path.insert(0, '/app/src/notebookllama'); from database import run_studio_migration; run_studio_migration(); print('Done')"
docker compose restart backend

Frontend shows old UI:

docker compose build frontend && docker compose up -d frontend

Database errors:

docker compose logs postgres --tail=30
docker compose restart postgres

Podcast stuck: Connect to DB and reset stuck tasks:

docker compose exec postgres psql -U postgres -d postgres_nextjs -c \
  "UPDATE background_tasks SET status='FAILED', error_message='Timeout', completed_at=NOW() WHERE task_type='podcast_generation' AND status='IN_PROGRESS';"

---

🚢 Production Deployment

See scripts/ for automated migration scripts:

• scripts/1_backup.sh — backup DB + files before migration
• scripts/2_deploy.sh — pull, build, switch from systemd to Docker
• scripts/3_cleanup.sh — remove build artifacts after verification

For developer notes, see CLAUDE.md.

---

📁 Project Structure

sandbox-notebookllamalm-nextjs/
├── backend/
│   ├── src/
│   │   ├── api/
│   │   │   ├── main.py
│   │   │   └── routes/           auth, notebooks, documents, chat, admin
│   │   └── notebookllama/
│   │       ├── database.py       SQLAlchemy models
│   │       ├── studio_generators.py  7 Studio LLM generators
│   │       ├── audio.py          podcast generation
│   │       ├── background_tasks.py
│   │       └── llm_factory.py
│   ├── Dockerfile
│   └── .env
├── frontend/
│   ├── src/
│   │   ├── app/
│   │   │   ├── notebooks/[id]/page.tsx   main notebook page
│   │   │   └── ...
│   │   ├── lib/api.ts
│   │   └── types/index.ts
│   ├── Dockerfile
│   └── .env.production
├── scripts/
│   ├── 1_backup.sh
│   ├── 2_deploy.sh
│   └── 3_cleanup.sh
├── docker-compose.yml
├── CLAUDE.md
└── README.md

---

---

🎨 Theme System

All pages support light and dark mode. The toggle is in the top navigation bar (☀ / ☾).

The theme is built on CSS custom properties (--bg, --fg, --primary, --border, etc.) defined in globals.css. The selected theme is persisted to localStorage and applied before hydration to prevent flash.

---

Version: 3.1.0 | Updated: March 15, 2026 | Status: Production