Composing AI Applications as a Graph with Haystack
summary
Haystack 2.0 models AI apps as a directed cyclic graph so pipelines can loop and become agentic; components are Python classes with run methods that pass typed documents, embeddings, prompts and metadata. Demos show an indexing pipeline (in‑memory docstore → OpenAI embeddings for Mistral/Cohere pages), an auto‑quizzer that builds JSON quizzes, and a loop that uses self-reflection loops plus Pydantic as JSON schema validation to retry until structured output meets the schema; Llama 3 (via Grok) and Gradio are used for UI and testing.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:03Haystack design rationale
- 2:50Breaking down RAG
- 4:35Embedding into vector DB
- 6:21From pipelines to agents
- 8:27Haystack 2.0 features
- 9:29Custom components in Haystack
- 10:54Document embedder demo
- 11:35Indexing pipeline with URLs
- 13:00RAG pipeline with Jinja
- 14:24Self-reflection loops
- 17:14Bringing components to Haystack
- 17:35Auto quizzer demo
- 19:42Quiz generation pipeline
- 21:25Serializing pipelines
- 22:49Validating outputs
- 24:15Looping pipelines Q&A
- 26:15Haystack vs frameworks
- 27:58When to use custom components
- 28:19Closing demo & Q&A
related talks
34mEverything is a graph, including LLM Applications (and that’s handy)
37mBuilding Generative AI Applications
48mFinding the needle: a deep dive into the rewriting of Haystack
43mLarge Language Models for Devs: from zero to your first LLM application
31mRAGtime with Haystack! How to build a RAG with Haystack 2.0?
41mFrom RAGs to riches: Build an AI document interrogation app in 30 mins
36mApplied AI with Eclipse Graphene
48mBuilding AI Powered Applications with Java