November 2024 Summaries

This is a summary of an article about using Graph Retrieval-Augmented Generation (GraphRAG) to streamline the process of ingesting commercial contract data and building a Q&A agent. The approach diverges from traditional RAG by emphasizing efficiency in data extraction, rather than breaking down and vectorizing entire documents indiscriminately. It uses a four-stage approach: targeted information extraction using LLMs and prompts, storing information extracted into a knowledge graph with Neo4j, developing simple knowledge graph data retrieval functions, and building a Q&A agent using Microsoft Semantic Kernel. The article provides an example of how to implement this approach, including creating a knowledge graph, defining data retrieval functions, and building a chatbot agent that can answer questions about contracts. The GraphRAG approach minimizes inefficiencies found in traditional vector search-based RAG by focusing on extracting only relevant information, reducing the need for unnecessary vector embeddings, and simplifying the overall process.

The Neo4j Developer Survey 2024 has been launched, aiming to create better tools for the community. The survey's first 500 unique responses will receive a $15 Amazon voucher. A project called SupplyChainInsights uses GraphRAG and fine-tuned LLMs to enhance global health supply chain analytics, providing detailed insights and efficient data retrieval. Meanwhile, a custom knowledge system was developed to improve organisational learning and decision-making by managing information across various platforms and fostering collaboration. The concept of Agentic RAG is explored, enabling powerful systems capable of handling complex queries and delivering highly relevant results. Additionally, the community can look forward to upcoming events such as Neo4j Live: Cognitive Sciences and Dynamic GraphRAG on November 28, and a meetup in London, UK on December 12. A featured community member, Astrid Paramita Mochtarram, will be presenting at NODES 2024 with her session on creating personalised travel agents using AI Agentic design patterns and Neo4j.

The Neo4j Query API offers a more streamlined experience with versioning, implicit and explicit transactions, and a single Cypher statement per request. The URL paths have changed to include the API name and support versioning, allowing for gradual migration without interruption. The Query API also includes response notifications with categorization, JSON with type information, and changes in the response format. When planning your migration, it's essential to be curious about your application's usage, remember that a slow and steady approach is better than a hasty swap-over, take advantage of running both APIs for back-to-back testing, and read the Query API documentation thoroughly.

We are thrilled to have hosted a successful conference, NODES 2024, which brought together thousands of attendees for a day of learning and coding, featuring over 140 unique technical talks on graph-related topics. Our keynote speaker, Ben Lorica, provided valuable insights from his conversations with AI leaders, while the Trends in Generative AI panel offered a pragmatic approach to understanding GenAI's real-world applications. The Neo4j team shared updates on their platform, including features aimed at simplifying database development and enabling easier scaling for AI workflows. Attendees were impressed by the top-rated talks, which covered topics such as Nodestream, climate modeling, and graph-based authorization. To make it easy to access all the content, a comprehensive playlist of slides and code repositories is available on YouTube.

In a world driven by data, companies are using database management systems (DBMS) to securely store and organize their data. A DBMS is software that stores data in a structured way, enabling storage and organization of data using defined formats and database structures. The market value of the DBMS industry tops $91 billion, with relational databases being the most widely used type. However, other types of DBMS models such as hierarchical, object-oriented, NoSQL, and graph databases are also gaining popularity. Companies use DBMSes for various purposes including data analysis, security, scalability, and compliance. While DBMSes offer numerous benefits, they can also be complex, brittle, and expensive to maintain. Examples of popular DBMS brands include Amazon RDS, Apache Cassandra, IBM Db2, Microsoft Access, MongoDB, MySQL, Neo4j, Oracle Database, PostgreSQL, SQL Server, and others. Graph databases are becoming increasingly popular due to their flexibility, intuitive organization, simplicity, speed, and scalability, making them suitable for handling challenging workloads and meeting analytical needs.

GraphRAG, a combination of knowledge graphs and retrieval-augmented generation (RAG), has evolved into various techniques with a growing body of research papers and software integrations. The second GenAI Graph Gathering brought together experts to discuss the progress and challenges in using knowledge graphs for retrieval, observing that applications often start with unstructured or structured data but typically stall in the pilot phase. A pattern catalog was curated to distill information from research papers, and proven approaches were implemented in tools and libraries to help address the "cold start" problem and provide guidance on domain-specific GraphRAG approaches. The discussion also covered knowledge graph construction, visualization, and ontology development, highlighting the importance of schemas for interoperability, explainability, and grounding. Advanced graph retrieval techniques, such as contextual retrieval, query-focused summarization, and GNNs, were explored, emphasizing the need for case-dependent approaches. Ultimately, GraphRAG continues to evolve with a focus on cross-organ collaboration, successful individual outcomes, and benefiting from the collective knowledge and technologies.

The GenAI Graph Gathering is an event that brought together experts in the field of graph-based artificial intelligence to discuss and share knowledge on GraphRAG, a technique that combines knowledge graphs with retrieval-augmented generation (RAG). The gathering aimed to catch up on recent developments and compare notes on various aspects of GraphRAG. Researchers have found that projects using GraphRAG often start with either unstructured or structured data, but those using mixed data tend to succeed more consistently. The group discussed ways to improve developer experience, including the use of templates and guidance for domain-specific approaches. They also explored knowledge graph construction, GraphRAG techniques, and real-world experience. Key concepts include knowledge graphs as an information architecture that can be simple or comprehensive, and ontologies as a set of concepts and categories that show properties and relations between them. The gathering highlighted the importance of interoperability, explainability, grounding, and balancing ease of use with rigor in ontology development. Advanced graph retrieval techniques continue to explore various methods, including contextual retrieval, query-focused summarization, and GNNs. Ultimately, the goal is for each guest to be successful on their path and for everyone to benefit from GraphRAG, which continues to evolve as a broad spectrum of approaches and technologies.

Welcome to This Week in Neo4j, your fix for news from the world of graph databases! We are back this week after a longer break due to NODES 2024. The team is celebrating the launch of a new GraphAcademy Course, which teaches how to create and query knowledge graphs using large language models. Additionally, GraphRAG Python Package has been released, allowing developers to build AI agents with Neo4j and LangChain tools. A livestream event, Neo4j Live: HybridAGI, is scheduled for November 19, while a conference will be held at AWS re:Invent in the US on December 02-06. Furthermore, Satej Sahu, Principal Data Engineer at Zalando SE, has shared his insights on using graph data to enhance Retrieval-Augmented Generation models through multi-agent integration.

The authors of a recent blog post have released a fine-tuned Neo4j Text2Cypher (2024) model, which demonstrates the potential benefits of fine-tuning foundational models on the Neo4j Text2Cypher (2024) Dataset. The dataset is used to translate natural language questions into Cypher queries, and the authors found that fine-tuning techniques can significantly improve performance over baseline models. The best-performing fine-tuned model achieved improvements in both translation-based Google BLEU score and execution-based ExactMatch score, outperforming its baselines by a considerable margin. However, the authors also caution against potential risks and pitfalls associated with fine-tuning, including changes in data distribution and access to training and test sets. Overall, the release of this fine-tuned model highlights the potential for Neo4j Text2Cypher (2024) tasks to be enhanced through fine-tuning techniques.

The guide focuses on setting up a Neovim environment for Cypher language support, starting from scratch. To achieve this, the author assumes no existing configuration and provides step-by-step instructions. The setup includes installing the necessary packages, including Cypher Language Server, LSPs, and plugins like `mason` and `nvim-lspconfig`. The author also discusses options for customizing the color scheme to improve aesthetics.

The GenAI Stack is a prebuilt development environment for building support agents that utilize Generative AI (GenAI). It uses Docker, LangChain, Ollama, and Neo4j to create an integrated system. The stack includes tools like GraphRAG, which combines knowledge graphs with retrieval-augmented generation (RAG) to improve accuracy, relevance, and provenance in support agents. RAG is a technique that combines retrieval mechanisms with generative AI models to produce more accurate and contextually relevant responses. The GenAI Stack supports using open-source LLMs like Llama and Mistral instead of relying on third-party providers. An example support agent app demonstrates how the stack can be used to build a chat interface that queries an internal knowledge base, retrieves relevant information, and generates new questions in the style of highly ranked existing ones. The GenAI Stack allows developers to customize the code and configuration to suit their needs, making it easier to build and deploy GenAI applications.

We explored how various fine-tuned and foundational LLM-based models perform in translating natural language questions to Cypher queries using the newly released Neo4j Text2Cypher (2024) Dataset. The results showed that closed-foundational models, such as OpenAI's GPT and Google's Gemini, demonstrated strong performance with user-friendly APIs and reliable output, though they can be costly. Previously fine-tuned models haven't quite matched these giants, but they demonstrate real potential for improvement through techniques like fine-tuning. We benchmarked four fine-tuned models and 10 foundational models to assess their performance side by side, using two evaluation procedures: translation-based evaluation and execution-based evaluation. The closed-foundational models delivered the best overall performance, with a match ratio of about 30 percent in the execution-based evaluation and outperforming previously fine-tuned models in the translation-based evaluation.

Are you a Java developer looking to integrate graphs and Neo4j into your applications? There's a new course on GraphAcademy: Building Neo4j Applications with Spring Data, which teaches how to use the Spring Data Neo4j library to interact with Neo4j. The course covers mapping graph data models to Java application domain models, converting between Neo4j and Java data types, reading and writing data, and using custom Cypher statements. You'll also explore SDN architecture, driver lifecycle, and transaction management. In this hands-on course, you'll develop an application that reads, writes, and runs queries against a Neo4j graph database of movies and actors. The course assumes you have knowledge of creating Java programs, Spring data, and completing previous courses on Neo4j Fundamentals and Cypher Fundamentals.

The Neo4j Text2Cypher (2024) Dataset is a machine learning dataset designed to help train and benchmark Text2Cypher models with ease. It was created by combining publicly available datasets, cleaning and organizing them for smoother use. The dataset consists of 44,387 instances, with 39,554 in the training split and 4,833 in the test split. The data preparation involved identifying and gathering datasets, combining and cleaning the data, creating training and test splits, and splitting remaining datasets. The resulting dataset is designed to support machine learning models that translate natural language into programming or domain-specific languages, such as turning plain text into Cypher query language.

The Text2CypherRetriever is a significant step forward for integrating Neo4j with Generative AI models, simplifying the retrieval process by using natural language processing to generate Cypher queries. This tool makes it easier for developers of all skill levels to easily retrieve data without a deep understanding of Cypher, providing flexibility and accessibility in retrieving information from Neo4j databases. The Text2CypherRetriever uses a large language model to translate user inputs directly into Cypher queries, eliminating the need for complex setups and making data retrieval simpler. It is particularly useful when precise, context-specific information is required without managing vector embeddings, offering simplicity, flexibility, and accessibility in its approach.

Support for MySQL and SQL Server has been added to Neo4j Aura, allowing users to import data from these databases into Free, Professional, and Business Critical Instances. The new import capability enables users to connect to existing data sources, define a graph data model, and initiate an import job without requiring code to move data into Neo4j. Initially, the feature is launched with connectivity to PostgreSQL, with plans for additional RDBMS and cloud data warehouse integrations in the future. Users can take advantage of this new capability by signing into the latest console preview and following along with a hands-on example using the Chinook dataset from PostgreSQL. The process involves setting up the data source, generating a graph model, importing data, and exploring the resulting graph with the Query co-pilot feature.