April 2025 Summaries

Building a GraphRAG (Graph-based Retrieval-Augmented Generation) system involves combining structured data, typically stored in knowledge graphs, with large language models (LLMs) to enhance the precision and relevance of responses. The process begins with structuring and modeling data using in-memory structures for dynamic datasets or choosing appropriate databases such as Memgraph, a real-time optimized in-memory graph database, or vector databases like Weaviate for semantic searches. Once the data is structured, key data points are identified through pivot searches using techniques like keyword, text, vector, or geo searches. Relevance is expanded by applying community detection algorithms such as Louvain or graph traversals to discover connected information, which is then appended to the user's query and sent to an LLM for a context-aware response. Memgraph's integration capabilities with tools like LangChain or LlamaIndex facilitate seamless interaction with LLMs, while hybrid approaches combining graph databases, vector databases, and search engines can be employed based on specific needs.

NASA has developed a People Knowledge Graph using graph databases and large language models (LLMs) to improve people analytics, enabling the identification of experts, team formation, and skill planning. The system, powered by Memgraph and running on secure AWS infrastructure, connects people, projects, and skills, allowing real-time insights through Cypher queries and a chatbot interface. The initiative addresses the limitations of traditional relational databases by modeling complex relationships and supporting subject matter expert discovery, project similarity analysis, and organizational insights. The graph schema includes nodes such as employees, skills, and projects, with future plans to expand the graph significantly. The project also leverages cosine similarity for project relationship analysis and stores embeddings for similarity searches. Despite its evolving nature, NASA aims to enhance data quality, automate pipelines, and scale up the People Graph to encompass over 500,000 nodes and millions of edges.

Choosing between GraphQL and REST APIs is essential for determining how an application interacts with data, especially in the context of graph databases. REST APIs, which have been the standard for over a decade, operate on the concept of resources and require multiple endpoints for data retrieval, leading to inefficiencies in handling interconnected data. In contrast, GraphQL, developed by Facebook, utilizes a single endpoint and allows clients to specify exact data needs, making it a natural fit for graph databases due to its ability to mirror graph data structures and facilitate complex relationship traversals in a single query. The integration of GraphQL with graph databases like Memgraph is particularly beneficial, as it allows for seamless data retrieval and real-time capabilities, enhancing the developer experience by simplifying complex traversals and supporting dynamic query needs without extensive backend changes. This makes GraphQL a more flexible and efficient option for applications dealing with heavily connected data, while REST remains suitable for simpler, less connected data scenarios.

Memgraph has been recognized as a Responsive Company in the Knowledge Graph Market by MarketsandMarkets' 360Quadrants, highlighting its commitment to leveraging connected data through its high-performance, in-memory graph database, Memgraph DB. This tool is designed to assist organizations in real-time insights for challenges like risk assessment and network data exploration. The growing importance of knowledge graphs is underscored by their ability to unify and analyze complex data, which is increasingly being combined with AI, machine learning, and natural language processing for applications like fraud detection and recommendation systems. Key market trends include the adoption of graph databases for scalability, the integration of AI for automated processes, and the demand for self-service data discovery tools. Memgraph's introduction of version 3.0, featuring GraphRAG and vector search capabilities, aims to enhance AI applications by providing accurate context and reducing hallucinations, thereby supporting the development of context-aware AI solutions.

The blog post provides a tutorial on creating an agentic Retrieval Augmented Generation (RAG) system utilizing graph algorithms, specifically Memgraph's PageRank, within a multi-agent workflow framework that leverages LlamaIndex. It builds upon previous setups by first incorporating Memgraph as a graph store and creating a sample dataset, then using LlamaIndex to define function agents for tasks such as data retrieval and arithmetic operations. A retriever agent is tasked with executing the PageRank algorithm to extract and rank nodes, while a calculator agent processes numerical data from these nodes. The tutorial includes setting up the environment with necessary dependencies and establishing connections to Memgraph using Python, along with detailed implementation of the agents and workflow needed for automated query execution and data processing. The post emphasizes the potential of integrating graph intelligence into agent systems and encourages further experimentation to enhance the capabilities of GenAI pipelines.

Large Language Models (LLMs) are impressive in generating plausible responses, but they have significant limitations when it comes to understanding proprietary or enterprise-specific data, often leading to confident but incorrect answers. The systemic flaw lies in their reliance on general, publicly available information and their inability to reason with context-specific data. This is where Retrieval-Augmented Generation (RAG) becomes crucial, as it grounds AI outputs in structured, verifiable data, enhancing the safety and scalability of AI systems in enterprise settings. RAG, particularly when combined with graph databases, offers dynamic, context-rich solutions that adapt to real-time data changes without the need for constant retraining. Graph RAG leverages the interconnected nature of graph databases to provide contextually relevant insights, making it particularly valuable in fields like finance and healthcare, where understanding the relationships between data points is essential. As businesses seek to create smarter and safer AI systems, the integration of RAG and graph technology is seen as a pivotal step forward, marking the beginning rather than the end of RAG's role in AI development.

In a tutorial by Matea Pesic, a multi-agent Retrieval-Augmented Generation (RAG) system is constructed using LlamaIndex and Memgraph, expanding upon a previous single-agent GraphRAG system to improve information retrieval for language models. The multi-agent approach enables specialization and collaboration among agents, thus allowing for more dynamic and capable data processing pipelines. The system integrates graph-based querying and tool-using agents by setting up Memgraph as a graph store, creating a Property Graph Index for structured knowledge retrieval, and implementing function agents for arithmetic and semantic tasks. It then combines these elements in an AgentWorkflow to handle complex queries, demonstrated through a scenario involving the 2023 Canadian federal budget. The tutorial concludes by highlighting the benefits of multi-agent systems in executing specialized tasks and hints at future developments involving Memgraph algorithms for even richer interactions.

Memgraph's high-performance graph database capabilities have been integrated into the KNIME Analytics Platform through a new extension developed by Redfield, enabling low-code graph analysis. This integration combines Memgraph's in-memory processing power with KNIME's visual, no-code interface, allowing users to analyze interconnected data without extensive coding. The extension includes nodes for connecting to, writing to, and reading from Memgraph databases using Cypher queries, facilitating seamless integration into KNIME workflows. A practical demonstration using a cocktail database showcases the extension's capabilities, highlighting its utility in exploring relationships, optimizing bar inventories, and visualizing bartender networks. By leveraging Memgraph's real-time graph analytics with KNIME's data processing tools, users can perform advanced network analysis and build recommendation engines efficiently, making it a valuable tool for teams working with connected data.

In April 2025, Memgraph announced updates to their integration with LangChain, now evolved into LangGraph, to enhance the development of agentic applications that rely on large language models (LLMs). The integration has been migrated to a dedicated repository, langchain-memgraph, and is available on PyPI, facilitating easier and faster feature updates. This package includes the MemgraphQA chain and Knowledge Graph construction capabilities. Additionally, Memgraph introduced a toolkit for agentic applications compatible with LangGraph, initially supporting Cypher query execution with plans for further expansion. The company encourages community feedback and contributions to guide future developments of the toolkit.

Large Language Models (LLMs) require improved context to function effectively, as their current limitations in processing text can lead to loss of important information and inaccuracies, particularly in long-tail dependencies. Fine-tuning these models to enhance their performance is costly, time-consuming, and often results in static models that struggle with real-time data updates. Moreover, LLMs face security risks as they may inadvertently disclose sensitive information without adequate safeguards. The article suggests using real-time knowledge graphs, such as Memgraph, to provide LLMs with dynamic and structured context, thereby enhancing the accuracy, relevance, and security of their outputs. This approach allows LLMs to generate personalized responses tailored to specific needs while ensuring that they work with the most current and relevant data, ultimately shifting the focus from fine-tuning the models themselves to refining how they access and process context.

The blog post by Matea Pesic explores the integration of LlamaIndex and Memgraph to create a single-agent retrieval-augmented generation (RAG) system, enhancing how data is retrieved and processed in AI-powered applications. Memgraph, a fast graph database, is used as a structured knowledge store, while LlamaIndex optimizes information retrieval for large language models (LLMs). The tutorial demonstrates setting up Memgraph, creating a Property Graph Index, and implementing an agent that performs both arithmetic operations and semantic retrieval. It involves using OpenAI's GPT-4 model for generating contextual responses and highlights creating a RAG pipeline to efficiently retrieve and query structured data, such as the 2023 Canadian federal budget. By leveraging these technologies, developers can build advanced, knowledge-graph-aware AI applications, with the article providing foundational examples for further development.

The Memgraph MCP Server is a newly introduced lightweight server implementation of the Model Context Protocol (MCP) designed to facilitate the connection between Memgraph and large language models (LLMs). By acting as a bridge, it allows for the seamless running of Cypher queries and retrieval of schema information from a Memgraph database, enhancing data interaction capabilities. The server is in its early stages but offers significant flexibility, enabling users to switch between different LLM providers and build complex workflows. To utilize the server, users can follow a quick start guide to set it up locally and use tools like Claude Desktop as an MCP Client to interact with the database, executing queries and understanding graph data structures. Future plans for the Memgraph MCP Server include expanding its features, supporting JavaScript environments, and integrating it with a central AI Toolkit repository to streamline the development of graph-powered applications and intelligent agents.