Everything is a graph, including LLM Applications (and that’s handy)
summary
Claims that thinking of NLP systems as Haystack pipelines as directed graphs reduces complexity, shows actual components used (embedders, rankers, generators), and demonstrates two working builds: a Notion question-answering app and a Hacker News summarizer wired into Streamlit and hosted on Hugging Face. Examples show validating node type connections, adding custom components like a Hacker News fetcher, looping a pydantic validator loop to retry until schema-conformant output, and that these pipelines run and produce useful results.
AI-written summary · about
chapters (13)
- 0:00Graph-based AI applications
- 1:51LLM interactions and capabilities
- 3:16Retrieval augmented generation
- 5:20Embedding text into vectors
- 7:25Hastack’s graph pipelines
- 9:11Building Hastack pipelines
- 11:40Agentic pipelines and decisions
- 13:45Live coding web QA demo
- 18:01Hacker News summaries pipeline
- 20:52Notion QA with custom components
- 23:50Looping pipelines for extraction
- 25:58AutoQuizzer multi-graph demo
- 28:48Q&A on training and design
related talks
29mComposing AI Applications as a Graph with Haystack
30mGraphs as the Backbone of Scalable Agentic Apps
22mBuilding Cypher Queries With Open-Source LLMs
30mBridging Classic ML Pipelines with the World of LLMs
35mBuilding LLM-based Agents
38mImprove LLM-based Applications with Fallback Mechanisms
41mEnhancing LLMs with Graph Technology
01m#NODES2025 countdown