Sieves - Plug-and-Play NLP Pipelines With Zero-Shot Models
summary
A modular toolkit claims to speed prototyping of document processing by composing small tasks around zero-shot-first models and enforcing structured outputs (integrating Outlines, DSPy, LangChain, Ollama, vLLM and Hugging Face models). It builds a pipeline abstraction with ingestion (dockling), chunking, and task engines for classification, NER, IE and distillation, runs a demo that extracts EU country stances in about 15 lines of code, and returns plausible structured results ready for downstream training.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:09Speaker and project motivation
- 3:37Ideal NLP pipeline tool
- 6:32Wish list challenges
- 10:10Sieves and core features
- 12:56How Sieves works
- 17:29Live demo: information extraction
- 21:02Quality of life features
- 23:07When to use Sieves
- 24:30Q&A
related talks
43mLarge Language Models for Devs: from zero to your first LLM application
27mVibe NLP for Applied NLP
12mCodeless Generative AI Pipelines: Prompt Automation
22mStructuring the Unstructured : Empowering the AI Workflows
01h 34mA Practical Starter's Guide to building LLM based projects
26mDSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines
45mArchitecture of an NLP Deployment
18mLLMs and GenAI: A Guide to Building Your Own RAG Model for Document Interaction