From Unstructured Docs to Graph Intelligence: A Framework for Building Graph+Embedding Pipelines
summary
A practical pipeline called D2GEP converts PDFs and transcripts into subject‑predicate‑object triples plus passage and node embeddings, using LangChain and spaCy for parsing, LLMs/LLM graph transformer for triple extraction, Olama/Gemini or OpenAI for embeddings, and Neo4j with a vector DB (or Neo4j embeddings) for storage. Demo runs from a docker‑compose stack, chunks at ~300–600 tokens, stores 384‑dim vectors, retrieves drug/entity relations (Jeffopixent) and reports about a 10–20% accuracy or grounding improvement versus plain GraphRAG or vector‑only retrieval.
AI-written summary · about
chapters (13)
- 0:00Introduction and welcome
- 1:07Problem with RAG limitations
- 5:10Introducing D2G framework
- 7:21Pipeline components overview
- 10:59Document processing steps
- 13:07Entity and relationship extraction
- 15:21Generating and storing embeddings
- 17:07Hybrid retrieval with Graph RAG
- 19:37Demo setup and tools
- 21:46Neo4j data population
- 23:11Querying the knowledge graph
- 26:31RAG vs hybrid approach comparison
- 27:39Q&A and closing
related talks
36mGraphRAG Explained 🔥 Building Smarter LLMs
01h 17mGraph Database and Retrieval Augmented Generation
15mAgentic GraphRAG: Simplifying Retrieval Across Structured & Unstructured Data
01h 00mComplex Data Ingestion with Open Source AI
22mStructuring the Unstructured : Empowering the AI Workflows
30mFusing NLP and Graph: Building a Conversational Agent for Enriched Financial Data
20mPractical GraphRAG: Making LLMs smarter with Knowledge Graphs
31mBeyond Vectors: Evolving GenAI through Transformative Tools and Methods