Vector Space Workshop

Listening to the World with Multimodal RAG.

You'll build a multimodal search system that answers questions like "What's on the radio in Berlin right now?" and returns live audio clips, transcripts, and related news articles from stations from a single query.

​What you'll build:

​Live radio streams from ~20 Radio Garden stations are already being chunked, transcribed, and queued for you. You'll experiment with:

• ​Multimodal embeddings: one Gemini E2 call site that puts audio chunks and article text into the same vector space
• ​Upsert + live feedback: write points to your own Vector Search Engine (Qdrant Cloud)
• ​Query layer: cross- and multi-modal retrieval with filtering and hybrid search
• ​Search Engineering: Qdrant's APIs enabling recommendations, diversity sampling, time-decay boosting, and relevance retrieval loops.

​⚠️ PLEASE BRING: laptop, charger, headphones. We will be using audio data, so headphones are a must!

​Who should come:

​Search engineers. ML engineers. Backend developers building retrieval systems.
Anyone who's touched embeddings and wants to see what multimodal search in production is like and learn from the AskNews team how you can turn multimodal retrieval projects into a viable business.

​What you get:

• ​A working repo running against your own Qdrant Cloud collection
• ​Hands-on experience with production-grade retrieval APIs

​What you need:

​A laptop with a browser. Headphones.

​Built with:

​Qdrant · Gemini 2 Embeddings · AskNews · Radio Garden · HackerSquad

‍