This podcast was simply amazing, I can't thank Karel enough for how much he taught me about DSPy, how to use it for Extreme Multi-Label Classification (XMC), and the applications of XMC in Biomedical NLP, Recommendation, Job Listings, and more. I am beyond grateful to have the opportunity to share this knowledge in the Weaviate podcast!

The podcast begins with an overview of Extreme Multi-Label Classification. How in the world do we prompt LLMs to categorize inputs into thousands of classes?!

To solve this, Karel has developed a novel Infer-Retrieve-Rank (IReRa) DSPy program. Infer first takes the input and outputs coarse labels for it. These coarse labels are then mapped to the thousands of classes (typically managed in ontologies) with the retrieval system and... you guessed it, Vector Embeddings! The Rank LLM component then takes the classes from the vector search and sorts them by relevance to the query.

Karel then took me through the details of the DSPy compiler! There is just so much opportunity with this from understanding how we tweak the descriptions of tasks we give to our language models, to populating the prompt with in-context learning examples. We discussed all sorts of things from model compression (e.g. can we prompt Mistral or Llama 7b to rival the performance of GPT-4 or Gemini Ultra at a particular task in an LLM pipeline, such as re-ranking or query writing?), diving into the latest on Teacher-Student optimization, input-dependent prompting, and so much more! We then concluded the podcast by discussing IReRa's applications for Recommendation Systems and what lead Karel to Biomedical NLP! Thanks again Karel, I learned so much from this one!

YouTube: https://www.youtube.com/watch?v=_ye26_8XPcs

Spotify: https://podcasters.spotify.com/pod/show/weaviate/episodes/XMC-dspy-with-Karel-DOosterlinck---Weaviate-Podcast-87-e2fehtk