Navi TV web app

Streamlit research testbed for an LLM-driven conversational content recommendation system.

Note: This is an unfinished research project. It is full of bugs and holes!

Conversational

• Memory
• Relevance, recency, importance (RRI) ML model

Content

• History
• User engagement
• RRI vector
• Device context (audio / bluetooth, device in motion, …)
• Mood - detected with pre-trained (but user fine-tuned?) ML model
• Channel selector ML model
• Prompt generation NLP models - one per channel

Architecture sketch: content recommendations

• Learn (and update) the RRI vector from conversations with the user
• Use mood, device context and user content history & engagement to select channel
• Generate prompt for channel and return list of content titles
• Use the RRI vector to filter content titles & descriptions

Development

Requires docker and a text editor.

1. Copy .streamlint/secrets-template.toml to .streamlint/secrets.toml and fill in the secrets.
2. Copy .streamlit/youtube-api-example-config.json to .streamlit/youtube-api.json and fill in the YouTube API config.
3. Run OPENAI_APIKEY=<your-api-key> docker-compose up -d.

The Streamlit app should now be available at http://localhost:8501.

Happy hacking!