September 2025 Summaries

Retail inefficiencies, despite the industry's vast growth, often lead to lost sales and customer dissatisfaction due to the inability of sales associates to efficiently meet customer needs from extensive product catalogs. This issue is addressed through the innovative use of multi-agent store advisers powered by generative AI (GenAI) on Amazon Bedrock, Amazon SageMaker, and Neo4j knowledge graphs, orchestrated by Oneture Technologies. These digital advisers facilitate a personalized shopping experience by utilizing specialized agents that focus on product discovery, inventory management, and customer preferences, transforming what was once a frustrating shopping experience into an informed and guided journey. Neo4j's graph database technology enhances this system by enabling real-time, context-aware recommendations, which improve with continuous learning from customer interactions. This approach not only boosts conversion rates and operational efficiency but also enhances staff productivity and customer loyalty by allowing associates to focus more on building relationships rather than memorizing catalogs.

This week's Neo4j update highlights the launch of the NODES 2025 agenda, a global conference focused on graph database applications and Neo4j innovations, set for November 6. Key developments include the creation of MCP-powered GraphRAG retrievers and the use of Neo4j for agentic memory in a Multi-User Dungeon (MUD) game. The update also discusses fine-grained, graph-based access control for multi-tenant SaaS applications, enabling dynamic enforcement of permissions. Featured community member Annabelle Blangero, a neuroscientist and responsible AI expert, will speak at NODES 2025 on multi-hop reasoning using a Neo4j knowledge graph. Upcoming events include live streams, conferences, and meetups, offering opportunities for learning and engagement with graph technology.

In the realm of cybersecurity, visualizing attack strategies through graph networks, or attack graphs, is crucial for both attackers and defenders. Attackers often view systems as graphs to identify the shortest path to critical assets, which are valuable targets within a network. Defenders can adopt similar strategies using graph database technology, like Neo4j, and visualization tools such as G.V() to preemptively identify and fortify against potential attack paths. By conceptualizing systems as graphs, defenders can easily identify vulnerable nodes and paths, allowing them to implement targeted security measures. This method not only provides an efficient way to analyze complex networks but also helps cybersecurity teams to predict and mitigate potential threats effectively. The integration of tools like Neo4j and G.V() enables intuitive exploration of these attack paths without extensive coding, offering a robust defense strategy against cyber threats.

Alex Gilmore's article explores the strategies and methods for generating knowledge graphs, highlighting their role in enhancing context management within the GenAI ecosystem through Neo4j's capabilities. The piece delves into the dual components of knowledge graphs—construction and retrieval—emphasizing the advantages of using graphs to connect structured and unstructured data, which facilitates complex filtering and traversals. It discusses the differences between traditional vector stores and knowledge graphs, particularly in the context of applications like medical Q&A systems, where graph-based retrieval (GraphRAG) can offer more nuanced insights compared to similarity search methods. The article also outlines the architecture of a knowledge graph generation pipeline, from data ingestion to the post-processing and validation of entities and relationships, underscoring the importance of linking unstructured documents with structured data for enriched context. Additionally, it provides a detailed examination of the lexical and domain components of knowledge graphs, the processes of entity extraction and context management, and the potential for evolving these methodologies to improve the reliability and accuracy of AI-driven insights.

The Neo4j GraphRAG Python library has introduced a new feature called the ToolsRetriever class and the Retriever.convert_to_tool() method, enhancing the library's flexibility by allowing multiple retrievers to function as tools within a single query. This feature empowers a language model (LLM) to intelligently select the most suitable retriever(s) for a given task, improving the adaptability and efficiency of retrieval-augmented generation (RAG) applications. The ToolsRetriever enables seamless integration with the GraphRAG framework by adhering to the Retriever interface, facilitating dynamic tool selection without the need for manual configuration. Additionally, the Retriever.convert_to_tool() method allows any retriever to be converted into a tool, supporting the combination of diverse retrieval strategies and ensuring consistent result formatting. This development provides a robust and extensible solution for applications requiring diverse data sources or fallback mechanisms and is expected to enhance user experience by abstracting tool selection, allowing developers to focus on application-building rather than retriever management.

Neo4j's Startup Program is designed to support AI-focused startups by providing them with essential tools and resources to build scalable, explainable, and production-ready AI applications. The program offers credits for Neo4j Aura, a graph-native database, to help manage costs while prototyping and developing AI solutions. Startups benefit from expertise in graph analytics and GraphRAG, receiving personalized consultations from Neo4j engineers to optimize their AI architectures. The program also includes co-marketing opportunities and partnerships to boost visibility and product adoption among a community of over 300,000 graph developers and data scientists. By joining the program, startups can leverage Neo4j's infrastructure and community to accelerate their AI innovation and market presence.

The Neo4j Startup Program is designed to support emerging AI startups by providing them with the necessary tools and resources to build scalable, production-ready AI applications. This program offers startups up to $16,000 in credits for Neo4j Aura, a fully managed cloud graph database, allowing for rapid development and cost-effective scaling of AI models. It also provides technical expertise from Neo4j graph engineers, who assist in optimizing AI architectures and developing enterprise-grade applications using GraphRAG and graph analytics. Additionally, the program enhances product visibility through co-marketing opportunities and partnerships, leveraging Neo4j's extensive community of over 300,000 developers and data scientists for increased engagement and brand recognition. With these resources, startups can efficiently develop, validate, and launch innovative AI solutions, positioning themselves to compete effectively in the rapidly evolving AI landscape.

Neo4j's latest update highlights the introduction of Infinigraph, a new graph architecture designed to unify transactional and analytical workloads, allowing seamless horizontal scaling beyond 100TB without application rewrites. Additionally, the release of a new GraphRAG course on DataCamp, which teaches the integration of Neo4j with LangChain for structured data extraction and enhanced chatbot functionality, is announced. The update also discusses a graph-first approach to integration strategies, underlining how AI and knowledge graphs can enhance flexibility and efficiency while maintaining human oversight, and explores using graph theory to bolster cloud security by identifying hidden attack paths. The community is invited to engage in various events and educational opportunities, such as live AMA sessions, conferences, and workshops, to further explore the capabilities of graph technologies.

Maxime Guery's article on optimizing Weakly Connected Components (WCC) projections in the Neo4j Graph Data Science library offers strategies to efficiently manage graph projections to reduce time and memory consumption. By utilizing the Neo4j Cypher query language, users can transform graph topologies to project only essential nodes and relationships, thereby minimizing resource usage without altering the algorithm's outcomes. The article explores three methods: projecting the graph as stored, removing unnecessary nodes and bidirectional relationships, and an optimized version that retains only the essential nodes and relationships. Each method maintains the integrity of the WCC results, which consistently yield 4,967 communities, but differ in the resources required for execution. These optimizations are particularly beneficial for users of Neo4j Aura Graph Analytics, where reduced memory usage can offer significant cost savings.

The article explores the evolution and development of text-based adventure games, particularly focusing on Multi-User Dungeons (MUDs) and the novel concept of Agentic Dungeons. It discusses using GenAI to enhance text-based games by solving challenges related to command parsing and dynamic world-building, which traditional models struggle with due to limited context windows. The proposed solution involves utilizing agentic memory with Neo4j graph databases to store game environments and NPC interactions, ensuring consistency when players revisit areas. The article provides a detailed overview of implementing a game set in a haunted castle using Neo4j and Java, highlighting the use of Cypher for game mechanics and the Embabel Java/Kotlin framework for agent management. Additionally, it offers insights into enhancing gameplay with AI-driven dynamic storytelling, multi-player capabilities, and potential future improvements like multi-floor castles and interactive NPCs.

An AI agent is an application utilizing Generative AI (GenAI) models, such as large language models (LLMs), to autonomously achieve goals through reasoning, planning, and tool usage in an execution loop. These agents are capable of handling complex tasks by breaking them down into smaller steps, using tools to gather necessary information, and learning from interactions to improve outcomes. They extend beyond traditional question-and-answer functions, integrating structured memory, context, and orchestration to dynamically adapt and plan tasks. The development of AI agents involves careful context management and the use of knowledge graphs to provide structured reasoning and decision-making, with real-world applications seen in industries like healthcare, finance, and customer service. However, challenges such as infinite loops, context hallucination, and data privacy concerns need to be addressed. The future of AI agents points towards more autonomous and contextually aware systems, potentially transforming traditional applications into agent-driven workspaces.

AI agents are sophisticated applications that leverage generative AI models, primarily large language models (LLMs), to autonomously achieve specific goals through reasoning, planning, and iterative decision-making. Unlike simple chatbots, AI agents can use tools, access structured memory systems like knowledge graphs, and operate within a reasoning and execution loop to complete complex tasks. These agents are increasingly used across various industries, handling repetitive tasks to allow human workers to focus on more strategic responsibilities. The development of AI agents has been propelled by advancements in AI technology, such as improved multi-step reasoning, enhanced context windows, and standardized protocols for tool integration. While offering significant operational benefits, AI agents also pose challenges related to context management, security, and human trust, necessitating robust frameworks and best practices for their safe and effective deployment. As AI agents evolve, they are expected to transition from purely text-based systems to dynamic, context-aware systems capable of reasoning with structured knowledge, thereby transforming traditional business applications into agent-driven workspaces.

Maxime Guery's article discusses a solution for preventing Cypher query crashes by utilizing the Weakly Connected Components (WCC) algorithm from Neo4j's Graph Data Science plugin. This approach addresses common issues with concurrent transactions, such as deadlocks and missing nodes or relationships, which typically result in transaction rollbacks. By implementing the WCC algorithm, users can execute Cypher queries with concurrency without encountering these errors, optimizing complex refactoring and writing operations in large graphs. The article provides an example where the WCC algorithm significantly improves performance by allowing queries to be processed concurrently, demonstrating efficiency gains over traditional Cypher-only queries, particularly in large graphs. It highlights that while the solution is effective, it may not be applicable if the write operation cannot be split by communities or if there are no disjoint graphs, and transaction times may vary depending on the size of the communities involved.

Neo4j has introduced Infinigraph, a groundbreaking distributed graph architecture that permits organizations to manage over 100TB of both operational and analytical graph workloads within a single system, avoiding fragmentation, infrastructure duplication, and performance compromises. This innovation opens up new possibilities for use cases such as global fraud detection, compliance analysis, and AI-ready systems by supporting real-time data traversal and maintaining full ACID compliance. Infinigraph eliminates the need for separate transactional and analytical systems, allowing for seamless integration and operation at unprecedented scales, which enhances decision-making and reduces costs. Neo4j is also enhancing its AuraDB platform with features like property-based access control, AI-powered dashboards, and multi-database support, which aim to improve security, accessibility, and cost-effectiveness for enterprise-scale graphs. These developments continue Neo4j’s tradition of innovation in graph technology, offering scalable solutions that assist companies in building intelligent AI systems capable of reasoning and decision-making.

Neo4j introduces property sharding as a solution for scaling rich graph databases without compromising performance or ease of use. As graph datasets grow to include billions of nodes and relationships, the properties associated with these elements become a scalability bottleneck. Property sharding addresses this by distributing properties across multiple database shards while maintaining the integrity of the graph structure. This method allows for efficient management of nodes and relationships with extensive property payloads, ensuring that operations remain ACID compliant and transparent to users through standard Cypher queries. The system operates within an autonomous cluster using a Raft group to manage graph shards and property shards independently, enhancing high availability and operational flexibility. However, the initial version lacks automatic rebalancing, which will be introduced later, and the solution is most effective for property-heavy graphs. Property sharding is currently available in Neo4j's Early Access Program for users dealing with significant property workloads.

The blog post explores the implementation of Neo4j GraphRAG Retrievers as a Model Context Protocol (MCP) server, highlighting the integration of vector search with Cypher to enhance semantic search capabilities in Neo4j databases. It explains how the MCP Neo4j Cypher server allows large language models (LLMs) to query Neo4j databases using natural language without directly handling Cypher syntax, thanks to the Text2Cypher approach. The post introduces the GraphRAG VectorCypherRetriever, which manages the workflow of embedding generation and vector search, abstracting complexities from the LLM. By doing so, the LLM can utilize a single tool interface for efficient data retrieval. The article discusses setting up this retrieval system, including defining database connections, vector indices, and retrieval queries, and emphasizes the importance of embedding models in the process. The post underscores the modularity and efficiency of MCP servers in enabling LLMs to focus on reasoning and orchestration, rather than the technical details of data retrieval, and suggests that similar principles can be applied to other GraphRAG retrievers to enhance data access logic.

The blog post discusses a framework for evaluating retrieval quality in Model Context Protocol (MCP) agentic systems, particularly focusing on how these systems interact with graph databases like Neo4j. It highlights the need to move beyond traditional single-step Cypher query evaluations to a more dynamic, multi-step reasoning approach, which better reflects real-world agent interactions that involve iterative processing and exploration of data. The article introduces a new benchmark designed to assess the quality of final answers produced by agents using an agentic approach, incorporating real-world complexities such as typographical errors and informal language. This evaluation benchmark, developed using Claude 4.0 and hosted via LangChain, signifies a shift towards measuring the semantic quality of results rather than just technical query accuracy, emphasizing the importance of concise, accurate answers over sheer retrieval capability. The results of the evaluation indicate that while agents can effectively handle complex queries using the MCP-Neo4j-Cypher interface, performance is impacted by factors such as input noise and question complexity, with potential improvements suggested through enhancing schema access and refining retrieval strategies.