January 2024 Summaries

This week's edition of This Week in Neo4j features a tutorial on creating knowledge graphs with Large Language Models (LLMs), an overview of the new features added to NeoDash 2.4, and discussions on using graphs for automation, including a proof-of-concept project that automated over 800,000 unique paths in a complex system. The newsletter also highlights various upcoming events, including NODES 2023 recordings and a meetup in Gothenburg or London, as well as resources for getting started with graphs, such as GraphAcademy tutorials and Neo4j AuraDB free trials. Additionally, the newsletter shares a tweet from S2 about their enthusiasm for using LLMs to traverse graphs and understand causality between entities, and introduces FlowMate, a BurpSuite extension that uses taint analysis to monitor parameter appearances in web responses.

This blog post explores how to scrape YouTube video transcripts into a knowledge graph for Retrieval Augmented Generation (RAG) applications. The project uses Google Cloud Platform, Neo4j, and LangChain to create a document from the transcript, store the resulting documents in a Neo4j graph database, and embed only the smaller child chunks of the text using SpaCy embeddings. The process involves setting up services such as Google Cloud Storage and Neo4j AuraDB instance, scraping transcripts from YouTube videos, chunking the transcripts into manageable pieces, loading the transcripts into the Neo4j graph database, and creating an index on the embedding property for vector search. The project demonstrates how to build a simple knowledge graph that can be used for RAG applications, with plans to explore building a basic RAG application in the next blog post.

NeoDash is a tool for building interactive graph dashboards with Neo4j, offering various visualizations such as charts, maps, and graphs. The latest release of NeoDash introduces new features including 3D graphs, interactivity through forms, and customizable interfaces using extensions. These features enhance the user experience, allowing users to explore complex graph data in a more intuitive way. With its modular design, NeoDash can be customized to fit individual needs, making it an essential tool for developers building Neo4j projects.

Neo4j has introduced two new DAG algorithms, Longest Path and Topological Sort, as part of its Graph Data Science library. Directed Acyclic Graphs (DAGs) are useful in various applications such as supply chain management, project scheduling, causal structures, citation graphs, microprocessors, and dependency management. A Gantt chart's critical path analysis can be performed using Neo4j's gds.dag.longestPath algorithm to identify the longest sequence of tasks that must be completed on time for a project to stay on schedule. The algorithm has been tested with a synthetic dataset containing 10,000 tasks and 5 million dependencies, showcasing its scalability. A benchmark comparison between the GDS algorithm and an equivalent Cypher query demonstrated the efficiency and robustness of Neo4j's specialized algorithms in handling complex graph operations.

This week's edition of This Week in Neo4j features a detailed look at deploying on Azure, an interview with Llamaindex co-founder Jerry Liu, and creating Knowledge Graphs from YouTube transcripts using the new Aura API. The newsletter also includes information on upcoming events such as NODES 2023, GraphSummit Series, and Data Day Texas, as well as resources for beginners to learn graph databases like Cypher Fundamentals, Relation Extraction, and deploying Neo4j on AKS with TLS. Additionally, there is a podcast episode featuring Jerry Liu discussing LlamaIndex and data technology, and a tutorial on integrating Neo4j Aura into GitHub Actions for automated testing.

A field engineer at Neo4j has created a step-by-step walkthrough of building a Retrieval Augmented Generation (RAG) application from PDF documents using GenAI-Stack and OpenAI. The project leverages Neo4j AuraDB for knowledge storage, LLM Sherpa for PDF document parsing, and OpenAI models for embedding and text generation. The walkthrough covers key components such as PDF document parsing and content extraction, Neo4j AuraDB setup, Python data ingestion, Neo4j vector index for semantic search, GenAI-Stack for fast prototyping, and OpenAI models for embedding and text generation. The project demonstrates an end-to-end pipeline from parsing and ingesting PDF documents to knowledge graph creation and retrieving a graph for given natural language questions.

This edition of This Week in Neo4j features a new course on building recommendations with Spring Boot, as well as articles and resources on LLMs, Vectors, Knowledge Graphs, and more. The newsletter also introduces the Neo4j Research panel, which allows users to share their experiences and influence the future of Neo4j products. Additionally, it highlights various upcoming events, including NODES 2023 recordings, GraphSummit Series, and a new course on building music recommendation backends with Spring Boot and Neo4j. The newsletter also features articles on building knowledge graphs, the challenges of virtual knowledge graphs, and a tweet from Tomaz Bratanic showcasing his work with LangChainAI and Neo4j.

The latest news from the world of graph databases includes the start of a new year, with topics covering Artificial Intelligence, LangChain, Knowledge Graphs, and RAG. The Neo4j Research panel invites users to share their experiences and influence the future of Neo4j products, offering compensation and direct connection with product development teams. Featured speakers include Alex Babeanu, who will discuss using graphs for real-time authorization decisions, and Ajit Jaokar, who explores the intersection of Knowledge Graphs and Large Language Models (LLMs). Additionally, articles cover LangChain's potential to turbocharge graph databases for intelligent conversations, simpler ways to query Neo4j Knowledge Graphs, and the construction of an efficient knowledge graph pipeline with LlamaIndex.

Vector search didn't work for RAG solutions because text embeddings often struggle with context sensitivity, contextual meaning, and evolving language use. The retrieved content based on embedding-based similarity search methods may impact the accuracy and correctness of generation in Large Language Models (LLMs). Challenges include context sensitivity, unrelated noise, reasoning of simple maths, information integration, negative rejection, conflicting knowledge detection, and counterfactual robustness. Some challenges can be tackled by finetuning a domain-specific embedding model or using advanced retrieval strategies to combine vector search with other search techniques. Specific test cases and evaluation metrics are needed for RAG solutions to address these limitations.