Knowledge Graphs and LLMs in Action (Alessandro Negro, Vlastimil Kus, Giuseppe Futia etc.)（Z-Library）

# Knowledge Graphs and LLMs in Action ## 【One-Line Pitch】 A practical guide for ML/AI engineers, data scientists, and data engineers who want to combine knowledge graphs with large language models to build reliable, explainable AI systems that reduce hallucinations and ground LLM outputs in structured domain knowledge. ## 【Book Arc】 - **Opening (~0%–8%)**: Establishes the "killer combination" thesis—knowledge graphs provide structured relationships and explainability while LLMs offer natural language understanding and generation. Introduces the hybrid intelligent system concept and the four pillars of knowledge graphs, with early use cases in drug discovery and customer support. - **Early (~17%–33%)**: Covers foundations of intelligent systems—what intelligence means, knowledge acquisition and representation, reasoning engines (deductive, inductive, and LLM-assisted), and the design philosophy behind hybrid approaches. Part 2 begins with hands-on KG construction from ontologies and structured data using RDF/LPG choices and tools like neosemantics. - **Middle (~42%–58%)**: Dives into real-world KG building from structured sources (biomedical, pharmaceutical, clinical domains) and then shifts to extracting knowledge from unstructured text using LLMs. Covers named entity disambiguation (NED) in depth, including a full SoHO knowledge graph case study and an open-LLM approach with Ollama and Llama 3.1. - **Late (~67%–75%)**: Moves into machine learning on graphs—node classification, link prediction, clustering, feature engineering (manual and semiautomated), graph embeddings, and graph neural networks (GNNs). Includes practical applications like anti-money laundering and movie recommendation systems. - **Ending (~83%–92%)**: Focuses on information retrieval—knowledge graph–powered retrieval-augmented generation (Graph RAG), natural language querying of KGs, schema-based approaches, intent detection, and building a full QA agent with LangGraph and Streamlit. Concludes with a policing-domain investigation use case and future directions. ## 【Key Takeaways】 - **Knowledge graphs and LLMs are complementary, not competing** (Early): KGs provide structured, explainable relationships while LLMs offer flexible natural language understanding. Together they mitigate each other's weaknesses—KGs reduce hallucinations and ground outputs, LLMs make KGs accessible via natural language. - **Start KG design from business needs, not technology** (Early): The book advocates an iterative top-down approach beginning with business and domain understanding before data modeling. This prevents building technically correct but practically useless graphs. - **RDF vs. LPG is a goal-driven decision** (Middle): Choose between Resource Description Framework and Labeled Property Graphs based on your query patterns, edge property needs, and tooling ecosystem. The book walks through this trade-off with concrete examples. - **LLMs can build KGs from unstructured text, but require careful prompt engineering** (Middle): Named entity recognition and relation extraction from documents become feasible with LLMs, but success depends on well-designed prompts, normalization, cleansing, and entity resolution strategies. - **Named entity disambiguation is the hard part** (Middle): Moving from recognition to disambiguation requires domain ontologies, candidate selection, and context-aware matching. The book demonstrates both traditional and open-LLM approaches (Llama 3.1) for medical domain NED. - **Graph machine learning follows an encoder–decoder pattern** (Late): From shallow embeddings like Node2Vec to message-passing GNNs, the encoder–decoder framework unifies graph representation learning. Understanding this pattern helps you apply node classification and link prediction across domains. - **Graph RAG outperforms vector-based RAG for structured queries** (Late): Vector-based retrieval struggles with multi-hop reasoning and precise relationship queries. Knowledge graph–powered RAG provides complete context and enables reasoning agents that can answer complex questions reliably. - **Schema-aware prompting is key for natural language KG querying** (Late): Converting user questions to Cypher queries works best when you provide the graph schema, descriptive annotations, and structured output formats to the LLM. Intent detection and response summarization complete the pipeline. ## 【Reading Tips】 - **Skim Part 1 (Chapters 1–2)** if you already understand KG basics; it's conceptual groundwork. Focus instead on the "deciding whether to use a KG" checklist and the hybrid reasoning engine discussion. - **Deep-read Chapters 3–4** for hands-on KG construction from structured data—the RDF vs. LPG decision framework and the biomedical/pharmaceutical case studies are directly reusable. - **Pay special attention to Chapter 7 (NED)** —it's the most detailed real-world example (SoHO knowledge graph) and covers the full pipeline from schema definition to disambiguation to KG-based use cases. - **Chapter 14 is the heart of the LLM integration** —the schema-based querying approach, intent detection, and "reasoning first vs. answer first" discussion are immediately actionable for building production QA systems. - **Skim Chapter 15** if you're not building agents with LangGraph specifically, but do read the architecture overview and the policing investigation case study for a complete end-to-end example. ## 【Coverage Limits】 This guide covers the book's progression from KG foundations through construction, machine learning, and LLM integration. The excerpts do not cover the appendices (graph introduction, Neo4j setup, structured source building) in detail, nor the final chapter's future directions beyond what's listed in the table of contents. ##