April 2024 Summaries

Large language models (LLMs) have gained popularity due to their advantages for organizations building generative AI applications. However, LLMs can produce incorrect responses that are presented confidently, a phenomenon known as hallucination. This can occur due to various factors such as insufficient training data, inadequate supervision, model overfitting, and knowledge cutoff dates. Hallucinations can manifest in three main types: factual inaccuracies, generated quotations or sources, and logical inconsistencies. To mitigate these hallucinations, researchers are developing strategies such as prompt engineering, retrieval augmented generation (RAG), and post-generation verification techniques. These approaches aim to enhance the reliability of AI systems, improve output quality, and ensure alignment with human values and factual accuracy. By leveraging targeted mitigation tools and strategies like RAG, organizations can significantly reduce instances of hallucinations, thereby maximizing the benefits of generative AI applications.

SingleStore Kai is a powerful solution designed to address data growth challenges in modern applications, particularly those using MongoDB. It integrates seamlessly with existing MongoDB applications, providing high-speed analytics capabilities without complex data migrations or code changes. The SingleStore Kai eStore demo application showcases its capabilities, leveraging MongoDB for transactional purposes and demonstrating significant performance improvements through ultra-fast CDC replication and innovative optimizations. By augmenting analytical requirements with MongoDB, users can leverage SingleStore for powerful features while maintaining existing transactional capabilities, making it a flexible solution for building complex applications.

The author reflects on the convergence of the data and AI communities in 2024, where traditional boundaries between them are becoming increasingly blurred. They discuss three major problems: RAG (Representational Artificial Generalism), specialized vector databases, and complex RAG startups. The author believes that general-purpose databases like SingleStore can solve the RAG problem by providing efficient ANN search capabilities. They also argue that specialized vector databases will become irrelevant as general-purpose databases improve their performance. Complex RAG startups, however, are likely to survive due to their focus on data retrieval and innovation in this area. The author highlights the need for a unified speedy layer that combines database and AI capabilities, and expresses skepticism about trusting AI with control over physical systems in the near future. They also disappointingly note that AI-powered analytics have not yet delivered on their promise of replacing human analysts. Finally, they predict that agentic apps, which enable AI to build workflows and solve problems, will become increasingly important in the enterprise world.

At the heart of our journey with SQrL lies a simple mission: to empower developers to fully leverage the capabilities of SingleStore, simplifying the creation of SQL queries and enhancing code debugging. We're introducing Code Generator within SQrL, designed to translate requirements into highly optimized SQL queries by processing requests in English along with selected tables. This feature alleviates complexity in navigating HTAP engine configurations for specific projects, allowing developers to focus on innovation rather than query optimization. Your data's security is paramount, and SQrL's Code Generator operates on schema information alone, ensuring sensitive information remains protected. By bridging the gap between complex SQL query formulation and strategic goals, we're setting a new standard for database interaction. We invite you to try out Code Generator through our Portal or by reaching out directly.

We are excited to announce an exclusive partnership between SingleStore and Agile Platform (에이플랫폼) to bring real-time, AI-driven solutions to the Korean market. SingleStore's distinctive data platform is specially designed for growing businesses dealing with apps, analytics, and AI — and its next-generation technology has been gaining significant traction in the APAC region. Agile Platform, a data-specialized firm, offers consulting services and the optimal solution, SingleStore, for enabling real-time processing, analysis and visualization based on extreme data — in both on-premises and cloud environments. Having already established a strong presence in Korea, SingleStore is proud to have leading companies including Kakao, POSCO, LG Innotek, Kbank, Kiwoon Securities, SK Telecom and many more onboard as clients. The partnership with Agile Platform will further enhance our market presence and offer Korean businesses access to a powerful, scalable data platform that effectively handles their growing data requirements. Through this partnership, we aim to empower Korean companies with the only real-time data platform that can read, write, and reason on petabyte-scale data in a few milliseconds — optimizing operations, enhancing analytics capabilities, and unlocking new data insights.

SingleStore is a general-purpose database that has achieved high performance in vector search, with benchmarking results showing it outperforms or compares favorably to specialized vector databases like Milvus and pgvector. The company's indexed vector Approximate Nearest Neighbor (ANN) search is designed for applications requiring fast querying of large vector data sets, and its modern, scalable data platform supports real-time analytics, full-text indexing, and queries over various data types. With SingleStore, developers can combine high-performance vector search with standard SQL capabilities, enabling the use cases that require both structured and unstructured data to be processed in a few milliseconds through one query statement. The company is continuously improving its performance and adding new features, including integration of new vector index libraries and improved hybrid search capabilities.

The new Fivetran destination for SingleStore enables seamless data integration from hundreds of sources into the real-time data platform, empowering businesses to leverage their data for real-time insights. Automated data integration tools like Fivetran have emerged as a solution to traditional methods' challenges, offering a streamlined approach to data integration. The connector bridges the gap between data sources and SingleStore, allowing users to effortlessly move data from various sources into the platform, while providing features such as real-time data sync and automated schema management.

SingleStore Helios SmartDR is a fully automated disaster recovery solution that replicates data and configurations to a secondary region, ensuring minimal downtime in case of a crisis. It differs from traditional disaster recovery solutions by not requiring idle computing resources in the secondary region, making it cost-effective. The solution prioritizes availability over consistency during failover, allowing customers to accept a five-minute window for potential data loss as part of their Recovery Point Objective (RPO). SmartDR offers pre-provisioning flexibility, enabling customers to test DR without disrupting production, and supports bidirectional replication, ensuring that new data entered in the secondary region is also replicated back to the primary region. The solution empowers users to operate mission-critical applications with confidence, featuring a highly available architecture, ACID transactions for real-time OLTP workloads, and fast data recovery through its online PITR feature.

SingleStore is celebrating National Poetry Month with a fun contest using its Notebooks feature. The challenge involves creating a haiku or ASCII art and submitting it for a chance to win Meta Ray Ban Smart sunglasses, an Amazon gift card, and bragging rights. To participate, developers can activate their free trial, create a workspace, and then start building a new Notebook with the desired name and access level. They can take inspiration from existing Notebooks and example projects, including one created by a developer evangelist showcasing how to create Haiku or ASCII art. The submission guidelines require sharing/pushing the project to GitHub, tagging relevant accounts on X and LinkedIn, and submitting the repository link by April 30, 2024, with any questions related to the contest or Notebook running/building asked in a dedicated Discord channel.

The SingleStore team implemented a Workspace Auto Suspend feature, which allows users to program their workspaces to suspend automatically after a certain amount of time. The feature is available in two types: Idle and Scheduled, with Scheduled suspend only available to paid subscribers. Intern Filipe Silva worked on the project during his summer internship, focusing on backend queries, frontend UI, and recurring tasks. After completion, the team sent the feature to QA for testing, which revealed some issues that delayed the initial release. The feature was later refined to improve efficiency, including a change to the recurring task approach. Key statistics show that most trial users enable auto suspend by default and set values around one hour, while paid subscribers tend to use scheduled suspend. The team learned valuable lessons about project estimation, default form values, and continuous improvement.

SingleStore introduces projections, a new feature that allows users to create secondary tables with different sort and shard keys than the primary table. Projections can be used to speed up range filters, GROUP BY queries, collocated joins, and COUNT(DISTINCT ...) operations by allowing data to be processed locally on partitions without requiring expensive cross-node data movement. Projections are a generalization of traditional secondary indexes and materialized views, but they offer more flexibility and power. They can be used sparingly, only when there is no alternative base table organization that will meet the user's needs. The future of projections includes improvements to the query optimizer and additional features such as online CREATE PROJECTION and bookmark lookups.

The text discusses the integration of SingleStore's database with OpenAI's language models to revolutionize resume evaluation in HR. This solution streamlines the recruitment process and enhances accuracy, enabling HR professionals to focus on strategic decision-making rather than manual data processing. The system uses a combination of vector indexes (FLAT, HNSW, IVF) to optimize query performance and storage efficiency for large-scale datasets. It also leverages AI capabilities to extract relevant information from resumes, generate embeddings, and perform semantic searches. The code provided demonstrates the implementation of this solution, including functions for data processing, embedding generation, and database operations. By integrating SingleStore's database with OpenAI's language models, organizations can significantly improve their talent acquisition and management capabilities.