March 2025 Summaries

Neo4j has recently been used by Klarna to power their in-house AI-driven solutions, replacing several SaaS tools including Salesforce. Neo4j is also being utilized for various other applications such as modeling agentic memory, analyzing banking customer behavior, and integrating with the Quarkus framework. The company is launching a new course on GraphAcademy, which teaches users how to integrate Neo4j into Python projects. Additionally, Neo4j's integration with Google's Gen AI Toolbox has expanded its functionality for database management and Gen AI applications. Furthermore, various webinars and meetups are being organized by Neo4j, including ones focused on customer conversation analysis and using Neo4j in banking.

GraphRAG is an emerging approach that leverages the graph-based representation of data to provide richer context, dynamic queries, and more reliable answers. It outperforms traditional RAG by offering deep relationship awareness, contextual retrieval, rich multi-hop queries, and explainable reasoning. Agentic architecture uses function calling and tool usage to delegate work to specialized components, enabling adaptive problem-solving, seamless information flow, smart orchestration, human-like reasoning, action-taking, and effortless scalability. By combining GraphRAG with agentic architecture in NeoConverse, we unlock AI workflows that are powerful, adaptable, and highly intelligent. This approach sets the stage for richer, more context-aware applications, delivering seamless and efficient experiences by leveraging the best of both graph and agentic thinking.

At Neo4j Aura, chaos testing is a crucial aspect of ensuring the stability and performance of their cloud-based software. To tackle this challenge, they've developed an automated workflow using Argo Workflows alongside Backstage, their internal developer portal. Engineers can trigger chaos tests via a simple API request, which sends a POST request to a microservice that forwards it to the correct Kubernetes cluster. The Argo Workflow Controller then executes the test, overloading the database with specific Cypher queries and verifying its health after the test. Logs and results are collected and returned to Backstage, allowing engineers to monitor execution details without needing direct Kubernetes access. This approach enables chaos testing at scale, making it simple and accessible for engineers across different teams.

We're adding support for cloud data warehouses, including Snowflake, Databricks, and BigQuery, to the Import Service in Neo4j AuraDB, enabling users to easily import their data into Neo4j. A new generative AI feature will inspect the warehouse schema, identify likely primary and foreign keys, and generate a candidate graph data model. This feature will also work with CSV files that lack such key information. The new capability simplifies the task of managing, analyzing, and gaining insight from complex connected data within an enterprise ecosystem. Users can quickly design a graph data model and import data from existing sources using the Import Service in the Aura Console.

Graphiti, an open-source framework developed by Zep AI, is designed to provide a real-time dynamic memory layer for autonomous agents in post-RAG (retrieval-augmented generation) systems. Unlike traditional RAG approaches that struggle with frequent updates, Graphiti's temporally aware knowledge graph engine incrementally processes incoming data, instantly updating entities, relationships, and communities without batch recomputation. This allows it to provide a unified, evolving view of the agent's world, handling chat histories, structured JSON data, and unstructured text from multiple sources. Graphiti achieves low-latency retrieval through its hybrid search approach, combining semantic embeddings, keyword search, and direct graph traversal, making it suitable for real-time interactions. The framework also offers custom entity types using Pydantic models to enhance context extraction and improve the quality of agent interactions.

Graph databases are designed to store and manage data as a network of connected entities, unlike traditional relational databases that use tables and joins. This approach provides several key benefits, including performance improvements, flexibility, and insights into complex relationships and patterns. Graph databases excel in industries such as finance, retail, healthcare, manufacturing, technology, government, transportation, and telecommunications, where complex connections between data elements are critical. They offer solutions to challenges like complex queries, real-time relationship analysis, flexible data models, performance issues with deeply connected queries, pattern detection across large datasets, and fraud detection. To get started with graph databases, organizations should identify their use case, model their data, start small, and scale with confidence. With the right approach, graph databases can help transform how businesses work with connected data, providing a competitive edge in today's fast-paced digital landscape.

This article discusses how graphs can be used to model different types of agent memory. The authors, inspired by Harrison Chase's talk at the DeepLearning.AI Dev Day conference, explore various aspects of memory management in agentic systems and provide a framework for implementing these concepts in a graph database such as Neo4j. They identify four main types of memory: short-term, long-term, procedural, and temporal, each with its own characteristics and challenges. The authors propose data models for each type of memory, including semantic memory, which stores facts about the world, episodic memory, which stores remembered experiences, procedural memory, which stores how to do something, and temporal memory, which stores how data changes over time. They also discuss the importance of managing memory in agentic systems and provide examples of how these concepts can be implemented in practice. The article aims to provide a solid foundation for solving the problem of agent memory management and offers a starting point for further exploration and development.

The traditional SaaS model is facing a fundamental shift with the rise of GenAI (Generative Artificial Intelligence). Companies like Klarna have transformed their data architecture to support GenAI initiatives, eliminating over 1,200 SaaS applications in the process. This fragmentation challenge coincides with Microsoft's prediction that traditional business applications will transform in the AI era, shifting logic to an "AI tier" that operates across multiple data sources. To succeed with GenAI, organizations need a unified knowledge layer, which is where knowledge graphs come in. A knowledge graph enables dynamic flexibility by design, allowing AI to efficiently reason over complex, multi-domain relationships. This flexibility makes them ideal for AI-driven systems, supporting dynamic retrieval, contextual understanding, and deeper insights across disparate data sources. Companies can start unifying their SaaS data for GenAI by building a knowledge graph as an abstraction layer on top of their existing SaaS data, creating agentic workflows with GraphRAG, and expanding incrementally to consolidate applications.

The Neo4j LLM Knowledge Graph Builder provides a user-friendly interface for uploading PDF files and other web sources, generating graphs using LLMs, and interacting with the generated knowledge graph through a chat experience. The tool's front-end architecture is built using React for application logic and components, Axios for handling network calls and responses, and long polling or server-sent events (SSEs) for real-time data updates. The UI guides users through each step of the knowledge graph-building process and chat experience, making it accessible and easy for users with varying levels of technical expertise. The system employs the Neo4j Needle Design System to ensure consistency, accessibility, and a polished design. Key features include automated graph generation and visualization, a robust chat interface with RAG, explainable answers, and Ragas metrics for response evaluation, as well as real-time updates via SSEs, table filtering, and reprocessing capabilities. Overall, the application combines modern design principles and usability with advanced features, empowering users to effortlessly build, refine, and interact with knowledge graphs, making data insights more accessible and actionable.

This week's episode of This Week in Neo4j covers various topics, including Knowledge Graph Construction. The LLM Knowledge Graph Builder has received new features in 2025, which include community summaries, parallel retrievers, and expanded model support for better knowledge graph construction from text. In contrast, Deepseek-R1 7B is compared to Mistral 7B for their effectiveness in generating meaningful graphs and reducing hallucinations. Meanwhile, BloodHound Viewer has been developed as a Chrome extension that enhances the BloodHound Community Edition with additional features such as query history navigation and improved layout controls. Additionally, strategies for managing graph assets in Retrieval-Augmented Generation (RAG) applications are discussed, focusing on optimized schema design, efficient querying, automation, security, and AI integrations. The episode also highlights various upcoming events, including Neo4j Live: Cypher Workbench and Stunning Graph Visualizations with NVL, as well as conferences and meetups in different locations.

Graph databases and relational databases have different structures, query languages, and performance characteristics. Relational databases excel at handling structured data operations with simple relationships, traditional business operations, and fixed schema requirements. Graph databases are better suited for connected data exploration, knowledge graphs, variable data structures, and pattern-based queries, making them ideal for applications that require flexible schemas, efficient handling of interconnected data, and seamless horizontal scaling. Understanding the differences between these database types is crucial for selecting the right tool for specific data access patterns.

Function calling is a powerful tool that allows Large Language Models (LLMs) to expand their capabilities, enabling them to generate structured output that adheres to function names and arguments. This concept is crucial for GenAI applications, allowing them to retrieve data from various sources, such as databases or APIs, and provide more interactive engagements. The article demonstrates how to use LangChain, LangGraph, and Pydantic to effectively implement function calling in agentic workflows. It showcases a cooking assistant application that uses Neo4j as the underlying database and leverages function calling to retrieve data from the database via Cypher queries. The example application also highlights the importance of tool validation using Pydantic models and demonstrates how to handle errors and exceptions. Function calling enables LLMs to access various tools, such as predefined Cypher queries or Text2Cypher, which can be used to generate novel responses or retrieve data from external sources. By abstracting away the subgraph details and function logic, function calling reduces the tokens being processed by the LLM, resulting in more efficient and scalable applications. The article concludes that while this approach may require additional effort, it offers a powerful way for GenAI applications to expand their capabilities and provide more advanced data retrieval methods.

The text discusses the importance of modeling data in a way that aligns with how we naturally think about it, rather than just following traditional database structures. It introduces the concept of knowledge graphs and graph databases as a solution to this problem, allowing for more flexible and scalable representation of complex data relationships. The guide outlines seven steps to build a knowledge graph: defining the use case, choosing a database management system, modeling the graph structure, applying an organizing principle, preparing data for ingestion, testing the graph, and maintaining and evolving it over time. It highlights examples of successful knowledge graph implementations in various domains, such as NASA's Lessons Learned Database and Cisco's metadata-driven knowledge graph. The guide emphasizes the importance of starting with a focused use case, following best practices, and continuously adapting to changing business needs.

This guide outlines how to automate the start and stop of Neo4j AuraDB instances using GitHub Actions, enabling cost optimization by ensuring databases are only running when needed. The automation involves setting up GitHub Actions workflows, integrating API credentials securely, and utilizing a Python script to control database states. By following these steps, users can efficiently manage their Aura instances with scheduled or manual triggers, reducing operational overhead and improving cost-efficiency.

The combination of Claude, Model Context Protocol (MCP), and Neo4j represents a significant step toward democratizing graph database technology by removing the coding barrier. This approach makes graph databases more accessible to a wider audience, including educators, researchers, and business analysts, who can now model complex relationships without writing code. By using a conversation-driven development method, users can extract information from documents, create nodes in the database, and visualize the resulting graph structure. The benefits of this no-code MCP approach include accessibility, speed, natural language interface, documentation, visualization, and iterative development, while also considering limitations such as complex queries, security, performance, and validation. This technology has the potential to simplify complex database tasks and open up new possibilities for data-driven applications.

The Gen AI Toolbox for Databases has been launched in collaboration with LangChain, now including a Neo4j integration. This integration brings knowledge graph capabilities to users, expanding the functionality of database management and Gen AI applications. Agentic architectures differ from traditional retrieval-augmented generation (RAG) approaches, where the LLM is equipped with tools for information retrieval and taking action on behalf of the user. The toolbox enables developers to build agentic applications that integrate database-based tools easily with the Google Gen AI Toolbox. It supports various databases, including Neo4j, PostgreSQL, MySQL, SQL Server, Spanner, and others. The toolbox provides a range of features to help applications hit production more quickly, such as end-user authentication in tools and built-in observability through OpenTelemetry. The Neo4j integration allows users to define sources and tools for Cypher execution, providing a flexible and powerful representation of connected information. This enables the creation of agentic LangChain applications with tools that use GraphRAG patterns combining full-text and graph search. The toolbox is open source, making it easy to contribute and reuse tools across different use cases. It provides a scalable and flexible solution for building AI agents with database-based tools.

The Neo4j LLM Knowledge Graph Builder is a system that transforms unstructured text into structured knowledge, enabling seamless searchability, queryability, and contextual understanding. It leverages the power of Large Language Models (LLMs) to automate the extraction process, eliminating the need for manual data structuring and reducing the dependency on handcrafted rules and patterns. The system's pipeline consists of five steps: data ingestion, chunking, embedding generation, entity extraction, and post-processing, which are designed to be flexible, scalable, and adaptable to various domains. By utilizing Neo4j's graph database capabilities, the builder streamlines large-scale knowledge extraction, making it faster and more efficient. The system also enables features like configurable chunking, embedding-based similarity searches, schema consolidation, community detection, and entity embeddings, ensuring accuracy, flexibility, and adaptability for real-world applications.

Data lineage documents the flow of data through an organization's systems, tracking its origin, transformations, and final use. It provides a meta-view of data flow, enabling teams to trace dependencies and anomalies, strengthen data governance for better data quality, control, and regulatory compliance. Data lineage is particularly important in dynamic environments where knowledge graphs have emerged as an effective solution for storing and analyzing data lineage, capturing how information flows and transforms across systems. Unlike data provenance, which focuses on the origins of data, data lineage describes the full data life cycle, encompassing where data comes from, where it moves next, how it is used, and relevant dependencies. Data lineage serves multiple purposes, including impact analysis and change management, migration planning, data quality and trust, root cause analysis, data governance, regulatory compliance, and analytics for machine learning, by providing a clear view of how data moves and transforms throughout systems.

The Neo4j Text2Cypher task aims to translate natural language questions into Cypher queries. The team analyzed the evaluation results from various angles, including overall performance, performance across different factors such as data source and database type, and common mistakes made by models. They found that outliers appeared in all the evaluation metrics, and assigned complexity levels to instances based on score distribution. Fine-tuned models showed improved performance compared to baseline models, but struggled with certain datasets and databases. The team identified common error groups, including additional matches, wrong ground truth, naming mismatches, confusion between WHERE and property conditions, and node vs. node properties. They also encountered surprising errors such as ambiguous questions, schema issues, and Cypher-specific challenges. The analysis highlights key areas where models struggle, mainly with the quality of ground-truth data and evaluation metrics, and will guide improvements to the dataset, models, and evaluation process in future updates.

This week in Neo4j features a win for ISO GQL at Developer Week in San Francisco, as well as news on configuring multiple Neo4j drivers within Symfony, the release of Grapho XR for Meta Quest, and a new destination connector for Unstructured. There's also a behind-the-scenes look at a Knowledge Integration App called "House of Wisdom", which uses AI to provide contextually rich conversations grounded in books and research reports. Additionally, there are updates on upcoming events, including livestreams, conferences, meetups, and webinars, as well as a featured community member Djordje Benn-Maksimovic who researches robust and secure machine learning for the public sector.