From Unstructured Docs to Graph Intelligence: A Framework for Building Graph+Embedding Pipelines

Session Track: Knowledge Graphs

Session Time: 08:30 am - 09:00 am November 6

Session description

Enterprise data lives in documents (PDFs, contracts, emails, product manuals) but extracting actionable insights from them remains a major challenge. In this session, Satej Sahu will introduce a practical, modular framework that transforms unstructured and semi-structured documents into a hybrid graph and embedding representation, enabling LLM-based reasoning and GraphRAG applications. The framework—D2GEP (Document-to-Graph-and-Embeddings Pipeline)—parses raw text into knowledge graph triples, embeds both textual content and graph structure, and stores them for hybrid retrieval in systems like Neo4j and vector databases. Using open-source tools and sample data (e.g. legal texts, scientific publications, or customer-service transcripts), I will demonstrate how to: - Parse and chunk documents using LangChain or spaCy - Extract entities and relationships into Cypher-friendly graph triples using LLMs - Generate node- and passage-level embeddings with models like OpenAI or SentenceTransformers - Store structured data in Neo4j and unstructured embeddings in Pinecone or FAISS - Enable natural language querying via GraphRAG — combining vector similarity and Cypher queries This session will walk through an end-to-end, reproducible pipeline, with reusable code, a template schema, and prompt engineering examples for extracting domain-specific knowledge.