AI Video Course Generator

This project creates short AI-powered video courses from a topic prompt.

What this system does

• Generates a course layout (chapters + subtopics)
• Generates slide content and narration
• Generates audio (TTS) for each slide
• Stores courses/slides in Postgres with strict user ownership
• Supports multiple AI providers/models
• Includes two chatbot modes:
  • Course chat (ask questions about the current course)
  • Home topic chat (query previous records by topic, upload PDF if no records exist)
• Uses Pinecone RAG + local Ollama embeddings for retrieval
• Uses Clerk for authentication across the app and API routes

Tech stack

• Next.js (App Router)
• TypeScript
• Drizzle ORM + Neon/Postgres
• Ollama (local LLM + embeddings)
• Pinecone (RAG vector store)
• Optional providers for slide generation:
  • Kimi (Moonshot)
  • Gemini
  • OpenAI

Run locally

1. Install dependencies

npm install

2. Configure environment variables in .env (or .env.local)

Minimum local setup:

NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY=replace_me
CLERK_SECRET_KEY=replace_me
AUDIO_STORAGE=local

OLLAMA_BASE_URL=http://127.0.0.1:11434
OLLAMA_MODEL=mistral:latest
OLLAMA_EMBED_MODEL=nomic-embed-text
OLLAMA_REQUEST_TIMEOUT_MS=1800000

PINECONE_API_KEY=replace_me
PINECONE_INDEX_HOST=replace_me
PINECONE_NAMESPACE_PREFIX=course

Optional model keys:

MOONSHOT_API_KEY=
MOONSHOT_BASE_URL=https://api.moonshot.ai/v1
MOONSHOT_MODEL=kimi-k2.5

GEMINI_API_KEY=
GEMINI_API_KEYS=

OPENAI_API_KEY=
OPENAI_BASE_URL=https://api.openai.com/v1

3. Start Ollama and pull required models

ollama serve
ollama pull mistral:latest
ollama pull nomic-embed-text

4. Start app

npm run dev

5. Open app

http://localhost:3000

AI model behavior

• Course layout generation uses selected provider mode (local-ai / global-ai)
• Slide generation uses selected Slide Model from UI:
  • ollama:mistral:latest
  • ollama:llama3.1:8b
  • kimi:kimi-k2.5
  • gemini:gemini-2.5-flash
  • openai:gpt-4o-mini
• If selected slide model fails, system falls back to local Ollama

RAG + chatbot behavior

• Course and slide records are indexed by topic namespace in Pinecone
• Home chat queries previous topic records
• If topic records do not exist, home chat asks user to upload a PDF
• Uploaded PDF text is indexed into Pinecone for later reuse
• Chat answers are generated by local Ollama

Main API routes

• POST /api/generate-course-layout
• POST /api/generate-video-content
• GET /api/course?courseId=...
• POST /api/course-chat
• POST /api/home-chat
• POST /api/rag-upload-pdf

Notes

• Route files are canonical under app/api/* and all protected routes enforce Clerk auth plus user-scoped authorization.
• Keep secrets out of git and rotate any key that was ever exposed.