May 2025 Summaries

Combining computer vision with knowledge graphs, the Memgraph Community Call showcased how images can be processed and queried semantically through a GraphRAG pipeline, transforming raw images into queryable knowledge. The session, led by Dino Duranovic and Ante Javor, illustrated the integration of image recognition with text-to-image embedding via CLIP, enabling natural language queries and zero-shot classification. Knowledge graphs were highlighted as a powerful tool for structuring visual information, allowing for accurate searches, intelligent navigation, and context-based query expansion. The live demo demonstrated the practical application of these technologies, where images were processed into a graph structure using the Gemini API and CLIP embeddings, enabling both image and text queries. The Q&A session underscored the advantages of using knowledge graphs over vector databases for contextual search and expandability, and addressed considerations for modeling and querying within this framework.

Memgraph, a graph database optimized for performance and developer experience, enhances the standard openCypher query language to offer more intuitive and expressive querying capabilities. As natural language interfaces become prevalent due to large language models (LLMs), there's a debate on whether to limit query advancements for LLM compatibility. The article explores two approaches for integrating natural language interfaces with graph databases: direct Cypher generation by LLMs and tool invocation where predefined tools mapped to Cypher logic are used. While direct generation offers flexibility, it poses challenges like bias, performance issues, and safety concerns due to the complexity of prompt engineering and the dynamic nature of databases like Memgraph. Memgraph advocates for a tool-based approach, where LLMs access well-defined tools, providing stability, safety, and control over database interactions. This approach allows developers to optimize queries, maintain performance, and ensure reliable AI systems without the need for continuous retraining. Memgraph's AI Toolkit supports this strategy by offering a collection of tools for building robust AI agents, emphasizing the importance of structure and predictability in dynamic environments.

GraphRAG leverages the synergy between Large Language Models (LLMs) and knowledge graphs to enhance personalization by addressing key limitations of LLMs, such as contextual relevance, reasoning across relationships, and personalization. Knowledge graphs, which organize entities and their relationships, enable multi-hop reasoning, improved retrieval accuracy, dynamic updates, efficient information navigation, and logical storytelling, making them ideal for personalizing LLMs. These graphs structure data in a way that allows for precise queries and up-to-date information, which is crucial for industries like healthcare and e-commerce. The integration of knowledge graphs into systems like GraphRAG enhances the ability of LLMs to provide insightful, contextually relevant, and personalized responses, as demonstrated by Precina Health, which used this approach for personalized diabetes treatment plans.

In a recent demonstration during a Memgraph Community Call, Josip Mrden, Head of Solutions at Memgraph, showcased the company's innovative use of vector search capabilities to transform unstructured data into queryable knowledge. Highlighting the potential for educational tools and intelligent information retrieval applications, Mrden built an interactive Q&A app and an automatic pub quiz generator using Memgraph's graph database and vector search technology. The demo, initially a side project, utilized a tech stack including Streamlit, OpenAI, and LLMs, to effectively connect unstructured and structured data, enabling fast prototyping and creating interactive learning experiences. Memgraph combines graph and vector search functionalities, offering a unified platform that enhances performance, hierarchical knowledge composition, and reduces complexity, making it a compelling choice for GraphRAG applications. Future improvements to the system include role-based access controls, multi-tenancy, and expanded data ingestion capabilities.

G.V() is a newly integrated client tool for Memgraph users that enhances the ability to query, visualize, and explore graph databases with a range of developer-oriented features. It provides an intuitive interface for executing Cypher queries with smart features like autocomplete and syntax validation, and it allows users to save and organize these queries into collections. G.V() offers various graph visualization formats, including Graph View, Table View, JSON View, and Summary View, which can be customized using stylesheets and filters. The tool also supports query profiling and provides an interactive Graph Data Explorer for seamless data exploration and editing. Moreover, it enables multiple query editing through multi-tab support and offers parameterized queries for flexible reporting. The platform is actively being developed, and users can try its capabilities through a free version called G.V() Lite or a month-long trial of G.V() Pro.

Large Language Models (LLMs) such as ChatGPT present challenges for enterprise use because they are not trained on proprietary data and have limited context awareness. Two primary solutions to enhance their relevance are Fine-Tuning and Retrieval-Augmented Generation (RAG). Fine-Tuning involves further training the model with specific domain data to produce highly customized responses, though it is costly and requires significant expertise and resources. RAG, alternatively, keeps data separate from the model by using a retrieval system to provide the necessary context for queries, making it more adaptable and easier to implement with dynamic data. While Fine-Tuning is suited for static and repetitive queries, RAG is ideal for situations where real-time information is crucial. In some cases, combining both methods could offer the best results by leveraging fine-tuning for domain understanding and RAG for real-time updates. The choice between these approaches depends on specific use cases, budget constraints, and technical expertise.

Data migration, particularly involving graph data, poses significant challenges, but Memgraph's integration with various data sources demonstrates a streamlined approach using its External Procedures and large language models like ChatGPT. By leveraging Memgraph Query Modules, which allow custom data integration logic in languages such as Python, Rust, and C++, and the code-generation capabilities of ChatGPT, integrations with over 15 data sources were achieved in just one day. These advancements include connectors for DuckDB, Spark, Iceberg, S3, and more, facilitating seamless data migration from diverse systems like relational databases, streaming platforms, and data lakes. Memgraph's new migration toolkit simplifies the process of bringing data into its ecosystem, enhancing compatibility with modern data environments and reducing technical barriers, thereby allowing teams to focus on product development and graph modeling without needing custom loaders or pipeline restructuring.

Multi-tenancy in graph databases, as explored by Marko Budiselic, involves multiple users or organizations sharing the same system while maintaining logical isolation, a concept that is crucial for database management systems. This approach simplifies administration, reduces overhead, and ensures isolation, allowing multiple users and applications to operate concurrently without interference. In databases like PostgreSQL, multi-tenancy can be implemented through schema-based isolation or row-level security, while in graph databases like Memgraph, it involves creating separate instances or enforcing access controls. The tradeoffs include challenges in resource utilization, reliability, and security, with the need for careful management of CPU and RAM to prevent resource hogging by rogue tenants. Real-world examples, such as Slack, illustrate how multi-tenancy can enable isolated yet interactive environments, while Memgraph's enterprise version offers significant cost savings and reduced maintenance through hosting multiple graph databases within a single server instance.