Due to their impressive natural language capabilities, recent large language models (LLMs) are paving the way for fast prototyping of NLP applications in any business domain. Most practical use-cases however will benefit from a structured, pipeline approach in which LLMs can be complemented with supervised models or even rule-based approaches. In this talk, I showcase how to build such a structured pipeline with the open-source NLP toolbox spaCy, and its recent extension ‘spacy-llm’. I discuss how to design a production-ready NLP pipeline, while managing different (and occassionally conflicting) performance features such as accuracy, speed, memory usage, reliability, maintainability and customizability.

→  Venue: Belgian NLP meetup

→  Slides: Speakerdeck

→  Repo: Github