March 2026 Summaries

Local graph search is a crucial retrieval pattern in GraphRAG, useful for extracting focused, context-rich subsets from a graph, especially when answers are not stored as single values but require understanding the surrounding relationships. This approach is particularly effective when the user's query begins with partial information, necessitating a pivot search to identify a relevant starting node, after which it expands through neighboring nodes and relationships to filter and return pertinent context. Atomic GraphRAG, Memgraph's execution model, enhances this process by integrating search, traversal, ranking, and filtering in a single query plan, reducing orchestration code and improving efficiency. This makes local graph search suitable for scenarios where the answer depends on the connections and interactions within the graph, such as recommending players based on engagement patterns or reviewing related GitHub issues, as opposed to questions requiring exact values or broad dataset synthesis.

GraphRAG is not a singular retrieval pattern but rather a system that adapts its pipeline based on the nature of business questions, with a focus on analytical questions using the Text-to-Cypher (Text2Cypher) retrieval pattern. Text2Cypher is effective for precise, structured queries within graph databases, providing exact answers like node existence, record counts, and filtered result sets. This approach is part of the broader Atomic GraphRAG framework, which aims to streamline data retrieval by integrating more logic directly into the database, reducing the complexity and sprawl of traditional systems. While Text2Cypher excels in answering specific analytical questions, it is not suitable for open-ended or exploratory queries that require broader context or synthesis. Atomic GraphRAG emphasizes maintaining a concise retrieval process within the database, allowing for efficient and maintainable pipeline execution, particularly in more complex query scenarios.

Memgraph, a complex C++ system, utilizes three distinct graph views—CMake, Ninja, and Conan—to analyze its architecture, build execution, and third-party dependencies, thereby providing insights into its structural and functional dynamics. These graphs are generated using Graphviz DOT format, converted to Cypher, and imported into Memgraph for comprehensive analysis using algorithms like PageRank and betweenness centrality. The CMake graph offers insights into target-level architecture and coupling, the Ninja graph focuses on build-step dependencies and optimization points for reducing rebuild fanout, and the Conan graph highlights package-level relationships and risks associated with third-party dependencies. This multi-layered graphical approach aids in identifying chokepoints, optimizing build processes, and managing dependency risks, ultimately facilitating safer code refactoring and more efficient builds. By treating central targets with care and focusing on optimizing high-impact nodes, engineering teams can reduce potential ripple effects and improve the maintainability of the system.

The guide explores the distinction between prompt engineering and context engineering in the development of AI systems, emphasizing the limitations of relying solely on prompt engineering—where the focus is on crafting prompts to communicate effectively with models—in real-world applications. As AI systems are increasingly integrated into enterprise environments, the need for context engineering becomes apparent, which involves structuring the AI's environment to ensure it has access to the relevant and accurate data necessary for its tasks. This shift addresses issues such as hallucinations and errors stemming from inadequate context, which prompt tweaks alone cannot resolve. Context engineering involves defining, curating, integrating, and governing the data and tools accessible to the AI, ensuring reliable and trustworthy output. GraphRAG, a method that combines knowledge graphs with retrieval augmented generation, is presented as a practical solution for implementing context engineering, offering a way to make AI systems more effective and grounded in business-specific knowledge. The guide underscores the importance of context engineering in transitioning from prompt-first to context-first AI development, aiming to improve AI reliability and utility in enterprise settings.

A leading retail bank, Capitec Bank, enhanced its fraud detection capabilities by integrating a graph-based machine learning pipeline using Memgraph to combat the rising fraud, particularly Authorized Push Payment (APP) scams. During a Memgraph Community Call, Derick Schmidt and Jan Ehlers discussed how fraud patterns often manifest in networks, necessitating a graph database approach to manage the complexity and scale of transactions that traditional systems struggle with. By transitioning from a standard machine learning setup to a graph-based approach, Capitec upgraded its pipeline to incorporate neighborhood and structure signals into the feature table, which are challenging for tabular engineering to capture. The process involved simplifying the graph schema, generating graph features, and dealing with data leakage through careful handling of fraud labels. The team used daily sampling to address class imbalance, ensuring the graph structure was preserved. This approach, connected to Capitec's fraud case management via AWS SageMaker and Kafka, resulted in efficient fraud detection with low false positive rates and scalable performance, scoring approximately 3.5 million records daily with an average runtime of two hours.

Memgraph recently transitioned from using its own co-located hardware infrastructure to primarily utilizing Hetzner's services, a move driven by the need for cost efficiency, predictability, operational simplicity, and benchmarking rigor. This shift resulted in a 50% reduction in infrastructure costs and improved build times and reliability. Hetzner's shared vCPU instances are optimized for cost-efficient velocity, significantly reducing build times and variability compared to the old hardware, while Hetzner's bare metal servers provide the stability needed for precise benchmarking. The migration alleviated the technical burdens of maintaining physical hardware and enhanced the company's ability to focus on core database development, culminating in faster, more stable CI pipelines and more scientific benchmarking processes. Despite the occasional outages, the infrastructure now offers improved networking, newer hardware, faster storage, and better support, leading to more reliable and consistent operations.

The blog post discusses the process of building a governed, queryable knowledge graph from 571 million Amazon reviews, highlighting the challenges and solutions in achieving scalability and efficiency. It introduces the Graph Development Lifecycle (GDL) managed by three key components: Graph.Build Studio for schema design, Transformers for data ingestion, and Graph Writer for data publishing to Memgraph. The article emphasizes the importance of a well-structured schema, mapping data to schema, and testing transformations on smaller datasets to avoid costly errors. It also explains how Kafka facilitates the ingestion of large datasets and describes the Atomic GraphRAG approach as a unified execution layer within the database, eliminating the need for external processing. The session underscores the significance of efficient memory management and dynamic schema loading to handle large-scale graph data and presents strategies for optimizing retrieval processes and reducing latency.

Context rot is a phenomenon that occurs when AI systems, particularly those integrating large language models (LLMs), degrade over time due to the accumulation of outdated or conflicting information without appropriate updating or filtering. This degradation manifests subtly, often leading to longer prompts, increased latency, and technically plausible yet incorrect answers. The common remedy of adding more context information exacerbates the issue by overwhelming LLMs with irrelevant data, highlighting the inefficiencies of vector-based retrieval methods. In contrast, graph-based retrieval systems, like GraphRAG, offer a more structured approach by maintaining coherent and current context through connected entities, which helps reduce confusion and improve the accuracy of AI outputs. The challenge of context rot extends to the tools and protocols used by AI systems, where evolving permissions and business logic further complicate data retrieval and processing. Addressing context rot requires disciplined context engineering, where information is carefully curated, updated, and prioritized to ensure the reliability and relevance of AI system outputs over time.

Memgraph's production telemetry stack, utilizing AWS, ClickHouse, and Grafana, is designed to efficiently handle large volumes of telemetry data for a distributed graph database, transforming raw events into actionable insights without becoming an operational burden. The architecture separates functions such as ingestion, storage, processing, and visualization to allow for independent scaling, using AWS for durable raw storage and batch processing, ClickHouse for flexible data querying and storage, and Grafana for visualization and trend analysis. This decoupled design enables rapid evolution of telemetry data alongside the product itself, allowing for scalable, cost-effective analysis of system behavior and feature adoption across diverse deployment environments. By leveraging AWS's infrastructure for ingestion and storage, ClickHouse's ability to work with semi-structured data, and Grafana's dashboard capabilities, Memgraph can maintain a reliable and adaptable telemetry system that informs engineering decisions and product development.