Beyond Vectors: Evolving GenAI through Transformative Tools and Methods
summary
A concrete pipeline: documents from docs, blogs, support and GitHub are split into 512-character chunks, embedded on GCP, stored in Neo4j and linked to URLs; KNN on those embeddings creates KNN on cosine embeddings edges, then FastRP and PageRank surface high-centrality chunks. EDA finds junk clusters (one community had median length 512 and JSON noise) and extreme redundancy (17 docs at 99% similarity), so the workflow collapses duplicate nodes and produces heatmaps of community usage to prioritize fixing blind spots.
AI-written summary · about
chapters (20)
- 0:00Introduction to talk
- 1:12Leveraging data for better RAG
- 2:38Building internal chatbot example
- 5:52Five principles of high-quality data
- 7:59Using EDA for data quality
- 9:21Graph data science for insights
- 13:15Cleaning and community analysis
- 15:00Handling text similarity issues
- 16:48Collapsing duplicate nodes
- 17:31Explainable AI with citations
- 18:14Tracking conversation flow
- 20:01Analyzing conversation patterns
- 21:48Dashboarding for insights
- 22:52Conversation length analysis
- 23:58Context document usage
- 25:41Most frequently used texts
- 27:29Community frequency insights
- 28:14Clustering responses
- 29:40Evaluating response quality
- 30:00Summary and closing remarks
related talks
01h 17mGraph Database and Retrieval Augmented Generation
30mFrom Unstructured Docs to Graph Intelligence: A Framework for Building Graph+Embedding Pipelines
01h 46mVector Databases (for GenAI & beyond)
01h 00mGraph Mastery to AI
01h 40mKnowledge Graphs & GraphRAG: Techniques for Building Effective GenAI Applications
41mGenAI, vector/semantic/hybrid search, RRF, NLP, (...) and other buzzwords
27mWhen Vectors Aren’t Enough: Unlocking AI’s Hidden Connections with Graphs
30mPerformant GraphRAG Improving Neo4j to Drive Autonomous AI ...