November 2025 Summaries

Memgraph's experimental server introduces the concepts of Elicitation and Sampling within the Model Context Protocol (MCP) to enhance user interaction and optimize database queries. Elicitation allows servers to request additional structured information from users via clients, creating an interactive workflow where users can dynamically provide input. Sampling involves the server delegating language model calls to the client, enabling the analysis and optimization of Cypher queries by recommending appropriate indexes. This approach facilitates a user-controlled optimization process, where users can choose whether to create suggested indexes, thereby ensuring transparency and control. The experimental server's workflow, demonstrated through examples, showcases its ability to analyze queries, identify missing indexes, and prompt users for index creation, ultimately aiming to improve system adaptability and lay the groundwork for autonomous GraphRAG workflows.

The blog post explores the challenges and solutions involved in integrating large language models (LLMs) with enterprise systems, particularly emphasizing the innovative role of the Model Context Protocol (MCP) in streamlining this process. MCP is portrayed as a standardized framework that facilitates seamless interoperability between AI tools and enterprise systems, much like a USB-C for AI, reducing the need for custom integrations that often become unsustainable at scale. Memgraph's recent updates, including the integration of MCP support into Memgraph Lab, are highlighted as pivotal advancements that allow the coordination of multi-server workflows without custom wiring. The post further delves into the functionalities of Memgraph Lab as an MCP client, enabling it to serve as a centralized hub for executing complex agentic workflows across multiple servers, exemplified through use cases in customer management and supply chain automation. The article also touches on the challenges of developing MCP clients, such as the rapid evolution of MCP standards and the complexity of managing multiple servers, while outlining the future roadmap for Memgraph Lab, including enhanced multi-server connectivity and team collaboration features.

The Memgraph MCP Server, a lightweight implementation of the Model Context Protocol, has rapidly gained traction since its release, allowing seamless interaction between Memgraph and modern large language models (LLMs) through natural language queries. Recent updates include a redesigned modular server architecture that offers pluggable implementations for different server variations, full integration with the Memgraph Toolbox providing a suite of tools for advanced graph operations, and a strict read-only mode to prevent unintended data modifications by LLMs. Additionally, the server now supports both streamable HTTP and STDIO modes, enhancing compatibility across different environments, and introduces a new developer-friendly installation path using the uv tool, making local server deployment more straightforward without relying on Docker. The development team is also exploring new experimental features like FastMCP’s sampling and elicitation to further enhance the server's capabilities, inviting user feedback to refine the experience for applications such as knowledge graph assistants or real-time data analysis.

Relational databases have long been foundational for organizations, but transitioning from SQL to graph models is gaining traction due to the latter's ability to reveal context and connectedness that SQL struggles to express. The SQL2Graph process, supported by tools like HyGM and Memgraph, offers an interactive approach to schema modeling, allowing for a guided transition rather than a complete manual overhaul. This transition involves automated and incremental modes, enabling users to either quickly migrate or maintain fine-grained control over complex schemas. Keeping SQL and graph data in sync poses challenges, especially in dynamic environments, but solutions like Kafka-based change data capture aid in maintaining real-time updates. The choice of LLM models for generating Cypher queries is critical, with agent-enabled models offering improved quality and adaptability. Crucially, SQL2Graph allows users to retain their existing SQL schemas, proposing graph models based on current structures, and facilitates iterative refinement to better represent data relationships, thereby enhancing LLM applications through graph-native retrieval without altering the original relational schema.

Memgraph Lab's recent introduction of the MCP Client, compliant with the Model Context Protocol (MCP), marks a significant advancement in AI workflow interoperability by enabling seamless integration of diverse data sources. This tool acts as a bridge connecting Memgraph with various MCP servers, such as GitHub, AWS, and Tavily, allowing users to query and interact with these systems directly from a single interface. The primary challenge it addresses is the fragmentation caused by siloed systems, which require complex custom integrations that slow down development and limit real-time data application capabilities. By facilitating data exchange and orchestration across platforms like Stripe, Elasticsearch, and Slack, the MCP Client transforms Memgraph Lab into a unified workspace, enhancing the ability to create context-aware AI applications. The tool's practical applications span diverse fields, including fraud detection, log correlation, medical research, and social media analysis, all of which benefit from interconnected data insights. Future updates promise further enhancements with features like collaborative sessions and deeper LLM model integration, positioning Memgraph Lab as a comprehensive environment for AI and data exploration.

As organizations grapple with the challenge of utilizing unstructured data, transforming it into entity graphs offers a significant opportunity, given that 80 to 90 percent of enterprise data is unstructured, yet only 18 percent of organizations leverage it effectively. To navigate the complexities of building a GraphRAG pipeline, it is crucial to address technical considerations such as schema design, embedding storage, access control, entity extraction, and balancing accuracy with cost. The Unstructured2Graph tool within the Memgraph AI Toolkit, highlighted in a recent Community Call, facilitates this transformation by enabling flexible schema evolution with Hybrid Graph Modeling and embedding management for efficient vector searches. It employs label-based access controls and role-based enhancements to ensure secure data handling and contextual entity extraction through LightRAG. By optimizing resources and leveraging GPU acceleration, Memgraph aims to strike a balance between speed, precision, and cost, providing a scalable and efficient path for creating connected, queryable entity graphs.

Context engineering has emerged as a critical challenge for AI teams, with the difficulty lying in structuring, retrieving, and reasoning over data rather than in the large language models themselves. Memgraph's AI Toolkit, including SQL2Graph and Unstructured2Graph, addresses this by simplifying the transformation of structured and unstructured data into graph-ready formats, enabling efficient data conversion for Graph-based Retrieval-Augmented Generation (GraphRAG). SQL2Graph converts SQL schemas from relational databases like MySQL and PostgreSQL into graph models with the help of the Hypothetical Graph Model (HyGM), allowing for either automatic or incremental migration. Unstructured2Graph turns unstructured documents into connected knowledge graphs, integrating tools like LightRAG for LLM-based entity extraction. These innovations aim to facilitate AI engineers and data scientists in building intelligent systems by providing a seamless transition of data into graph domains, paving the way for future collaborative environments through the upcoming MCP Client in Memgraph Lab.

Modern Retrieval-Augmented Generation (RAG) systems rely on embeddings to capture both the semantic meaning of text and the structural relationships between entities in graphs, with tools like GraphRAG combining language and graph topology to enhance search and reasoning capabilities. Memgraph addresses the challenges of scaling embedding computation and management by implementing efficient, real-time processing that leverages GPU acceleration, resulting in significant performance improvements over CPU processing. Embeddings, represented as high-dimensional vectors, are crucial for tasks such as retrieval, clustering, and link prediction, but their computation, storage, and freshness present practical challenges, especially at scale. Memgraph optimizes storage with its Advanced Vector Search initiative, reducing duplication by using indexes as primary storage, thus improving performance and simplifying maintenance. The platform provides infrastructure to support scalable embedding computation, ensuring that GraphRAG pipelines remain efficient and contextually aware, enhancing the capabilities of AI systems in processing and retrieving relevant information.

The SQL2Graph Agent is a migration tool designed to automate the transformation of relational databases into graph structures, particularly for use with Memgraph. It addresses the challenge of shifting from relational to graph reasoning by automating schema analysis, graph model generation, and data validation through a combination of deterministic and AI-powered strategies. This agent leverages Hypothetical Graph Modeling (HyGM) and employs large language models (LLMs) to facilitate pattern recognition and reasoning, offering two modes: a fully automated mode and an incremental mode that incorporates user feedback. The system ensures data integrity and schema consistency through iterative validation processes, making it easier for developers to convert relational data into graph models without in-depth knowledge of graph databases. The agent is Python-based, utilizing LangGraph workflows for orchestration and Memgraph Mage migration modules, and it supports seamless integration with existing relational systems like MySQL and PostgreSQL.

HyGM, or Hypothetical Graph Modeling, is a framework designed to automate the conversion of relational schemas into graph models, enhancing the workflow of GraphRAG projects by iteratively proposing, refining, and validating graph schemas. It operates at the core of the SQL2Graph Migration Agent, providing the intelligence necessary to evaluate relational schemas and propose efficient graph structures. HyGM views schemas as hypotheses rather than fixed designs, enabling iterative feedback loops that allow for continuous refinement and improvement through human input and automated suggestions. This approach supports two strategies: a deterministic, rule-based method ideal for well-structured schemas, and an LLM-powered strategy that utilizes language models to comprehend semantic meanings for more complex cases. The process includes validation to ensure the generated models align with existing data, ultimately leading to an efficient and adaptable graph modeling workflow that integrates domain expertise and automated intelligence.

Transforming unstructured text into structured data is crucial for uncovering hidden connections and insights within documents such as PDFs, reports, and meeting notes, which traditional search tools may miss. Unstructured2Graph, part of the Memgraph AI Toolkit, facilitates this transformation by converting unstructured information into an entity graph, a network of nodes and edges representing people, places, organizations, and concepts. This process involves several steps, including text extraction, cleaning, and entity and relationship extraction using tools like spaCy and large language models, before organizing the data into a graph database like Memgraph. The resulting entity graphs support enhanced search, recommendations, and analytics, proving useful in fields like competitive intelligence, scientific research, legal analysis, and retrieval-augmented generation. Unstructured2Graph offers automated tools and libraries like LightRAG to simplify the conversion process, allowing for the efficient capture of meaningful insights from complex text data. Additionally, considerations around cost and processing speed are addressed, with continual optimization efforts aimed at improving throughput and affordability.

Unstructured2Graph, a tool within the Memgraph AI Toolkit, enables the transformation of unstructured documents such as PDFs, DOCX, and HTML into a structured knowledge graph that can be queried and reasoned over by language models. It utilizes Unstructured IO for the extraction and cleaning of text, while LightRAG handles entity recognition and relationship mapping, resulting in a graph with nodes, edges, and embeddings ready for retrieval. Users can quickly set up a project in Memgraph Cloud, ingest documents, and build entity graphs without extensive local setup. The tool also allows for the optimization of graphs through embeddings and vector indexing for semantic search, and it supports the integration of structured data for a comprehensive knowledge graph. The process involves several steps, from initializing the environment and ingesting documents to optimizing and visualizing the graph, all of which are facilitated by a few Python commands.

SQL2Graph is an agentic migration tool within the Memgraph AI Toolkit that automatically converts structured data from relational databases like PostgreSQL and MySQL into graph structures in Memgraph, facilitating the transition from tables and joins to relationships and knowledge graphs. This tool, powered by HyGM agents, simplifies the process by inferring relationships and building a queryable knowledge graph, which is essential for GraphRAG projects. Users can choose between automatic and incremental migration modes and select between deterministic or AI-powered modeling strategies, allowing for flexibility in schema conversion. The tool also supports validation checks to ensure data integrity and consistency post-migration. By integrating SQL2Graph into ETL workflows, users can automate migrations and maintain up-to-date graph databases, making it a valuable resource for AI engineers seeking to enhance their data workflows with connected context and structure.

Text search technology, integrated into Memgraph, enhances the ability to discover meaningful insights within graph data by indexing and ranking text based on context, rather than merely matching exact words. This integration, powered by the Rust-based Tantivy engine, allows Memgraph to unify the exploration of both data content and its relationships, facilitating applications such as recommendation systems, fraud detection, and knowledge graph development. The text search feature is transactional, ensuring results reflect the precise state of the database, and supports durability, replication, and multi-tenancy. By using a C++/Rust interoperability layer known as cxx, Memgraph bridges the two programming languages, leveraging Rust's performance without compromising its C++ foundation. This development allows users to efficiently examine structured and unstructured data, offering a robust, reliable, and contextually aware search capability within graph databases.