December 2023 Summaries

The Keycloak authentication pages for Singlestore Helios's Portal have undergone a significant overhaul, focusing on aligning with the rest of the application built with React and TypeScript. The development process involved analyzing and optimizing the bundle size, exploring various techniques such as internationalization stubbing, icon importing, lodash modernization, and alternative library exploration like Preact. A thorough analysis using vite-bundle-visualizer led to identifying unnecessary files, icons, and dependencies like lodash. By implementing lazy loading strategies and making adjustments in import paths, the team was able to reduce the bundle size from around 1.8MB to just over 400KB, with an additional improvement through the use of Preact, resulting in a total reduction of around 23% of the original size.

In a world where companies are integrating AI into their fabric, the landscape of databases is witnessing the emergence of vector databases, which are transforming the way data is handled. Vector libraries and databases are becoming increasingly popular as companies navigate this digital odyssey, driven by the meteoric rise of generative AI and Large Language Models (LLMs). Retrieval Augmented Generation (RAG) is a software pattern that relies on meaning-based search to turn raw data into vectors, making it easier for AI's grasp. Companies are now faced with a dizzying set of choices on how to build enterprise generative AI applications using vector stores, which come in three broad categories: vector libraries, vector-only databases, and enterprise databases that also support vectors. Vector libraries, such as FAISS, NMSLIB, ANNOY, ScaNN, offer efficient similarity search and clustering capabilities but lack comprehensive database functionalities. Vector-only databases like Pinecone, Weaviate, Milvus, ChromaDB, Qdrant, and Vespa are designed for scalable, high-performance similarity search in applications like recommendation systems and AI-powered search. Enterprise databases with vectors, such as Elasticsearch, MongoDB, SingleStoreDB, Supabase, Neo4J, Redis, and PostgreSQL, offer broader data handling capabilities, versatility in RAG, and real-time computation and data serving capabilities. Companies need to evaluate these options based on their specific use cases, including multiple data type support, search methodologies, data freshness and latency, transactional or analytics use cases, prototype to production, and other requirements. With the emergence of vector databases, developers are now faced with a new set of choices for building generative AI applications at enterprise scale.

I'd like to express gratitude for key milestones of our company over the past 12 months, including participating in the World Economic Forum Annual Meeting in Davos, Switzerland, where I spoke on a panel about real-time solutions for data management. We also launched innovative products such as SQrL and SingleStore Kai, which enable advanced AI capabilities and faster analytics. Additionally, we hosted our inaugural conference, SingleStore Now, and shared our collaborations with partners like Google, Amazon Web Services, and IBM. Our company was featured in several media outlets, including Bloomberg and CBS News, highlighting the potential of real-time AI to make a positive impact on society. I'm excited to continue these discussions through our "Into The Singleverse" podcast and look forward to what 2024 brings.

Google BigQuery, a popular cloud data warehouse, is often criticized for hidden costs and limitations in handling certain workloads, particularly those requiring frequent updates or low query latency. While BigQuery offers a low initial barrier to entry, its complex pricing model can lead to unpredictable expenses as businesses scale, with costs associated with compute capacity, data processing, and additional services. SingleStoreDB emerges as an alternative, promising cost predictability and the ability to handle both OLAP and OLTP workloads efficiently without the need for ETL or data movement. It is particularly beneficial for real-time analytics, end-user applications requiring low latency, and database consolidation. ZoomInfo, a data company, is highlighted as a case study, having saved $1.2 million by switching from BigQuery to SingleStoreDB due to improved performance and reduced total cost of ownership. SingleStore offers the flexibility to deploy in various environments, providing businesses with operational freedom and financial clarity compared to BigQuery.

SingleStoreDB is a high-performance database that allows users to transact, analyze, and contextualize data in real-time. It uses Universal Storage, which employs three storage tiers: in-memory rowstore, on-disk columnar store, and object storage. The on-disk columnar store stores warm data, while cold data resides solely in object storage. However, certain situations may result in no data being stored on the disk cache, impacting query performance. These scenarios include resuming a workspace or attaching an existing database to a new workspace. To address this issue, users can "warm up" their cluster by querying their data, which involves reading blobs from object storage and persisting them to disk. This process can be automated using a notebook provided by SingleStoreDB. By warming up the disk cache, users can avoid latency penalties when querying data that resides on disk versus object storage.

In a recent webinar, panelists discussed the growing importance of real-time Large Language Models (LLM) applications and their practical use cases in handling inventory, product, and review data. SingleStoreDB was highlighted as an ideal database for providing scalable, cost-effective, and performant storage and querying of large datasets, including structured and unstructured data and vectors. The platform's compatibility with popular technologies like Kafka, Spark, and Hadoop, as well as its deployment options across major cloud platforms, makes it a flexible solution for various use cases. Additionally, advancements in vector support and semantic searches, Neum AI's role in LLM application development, the significance of Retrieval Augmented Generation (RAG) in improving LLM responses, and the challenges involved in embedding data were discussed. These topics are crucial for developing and deploying successful LLM applications that can provide accurate and relevant information to users.

The Elegance SDK is an open-source JavaScript library that enables developers to quickly build real-time AI applications using SingleStoreDB, a NoSQL database. The SDK provides a set of ready-made tools for implementing various types of functionality, including vector search, chat completions, and file embeddings generation. It supports both MySQL and MongoDB connection types, as well as the OpenAI API. The SDK can be used in existing projects or with Next.js/Express.js templates, and it includes features such as server-side rendering, client-side rendering, and React hooks for easy integration into web applications.

In our recent webinar "How to Build Custom GPTs Using OpenAI Functions" SingleStore panelist team discussed custom Generative Pre-trained Transformers (GPTs), which are AI models tailored for specific tasks or industries, fine-tuned to provide accurate and relevant responses based on specialized data. These models differ from standard GPTs in that they are trained on specific datasets, making them more efficient in niche applications like legal analysis or medical research. Custom GPTs can handle sensitive data securely, although OpenAI's privacy and security commitments for enterprise customers are still evolving. SingleStoreDB efficiently manages both structured and unstructured data, making it an ideal choice for giving custom GPTs the additional data context they need to provide powerful, scalable, real-time responses. No-code platforms allow users to create custom GPTs without in-depth programming knowledge, making the technology accessible to a wider audience. Custom GPTs can generate reports or summaries and are bound by the quality and breadth of their training data, with limitations such as file upload limits and instruction length limits that can be overcome using the Assistants API and SingleStoreDB.

Llama 2 is a state-of-the-art language model developed by Meta that offers enhancements in terms of scale, efficiency, and performance. It is the successor to the original LLaMA model and has been trained using publicly available online data to learn general language patterns and acquire a broad understanding of language structure. The model incorporates several innovative elements, including RMSNorm pre-normalization, SwiGLU activation, and Rotary embeddings, which contribute to its ability to maintain context over longer stretches of conversation and offer more precise attention to relevant details in dialogue. Llama 2 has been trained with a reinforcement learning approach to produce non-toxic and family-friendly output, and it has been fine-tuned through supervised learning and reinforcement learning with human feedback. The model is open-source and available for download under a community license, making it an economical option for businesses that want to integrate the model with their internal data and fine-tune it for specific use cases while preserving privacy. Llama 2's flexibility, versatility, safety, and customization capabilities make it a valuable asset across sectors, including chatbot applications, summarization, translation, content generation, and coding assistance. The model is designed to be as safe or safer than other models in the market and offers an alternative for those seeking to develop on a platform that supports modification and redistribution.

This walkthrough guides users through the initial setup and data ingestion process in Singlestore Helios, introducing fundamental SQL queries to analyze and transform ingested data. The journey begins with creating a workspace group, setting up a workspace, and understanding credits within the platform. Users learn how to load sample data from an AWS S3 bucket and create a database, followed by simple SELECT statements to retrieve and limit data. They also explore basic data transformation techniques using SQL queries, ultimately mastering aggregations and grouping data for meaningful insights. By completing these steps, users can unlock the potential of Singlestore Helios and delve deeper into complex analyses, data manipulations, and intricate queries to derive valuable business insights and drive informed decisions.

The text provides a step-by-step guide on how to create a GPT (Generative Pre-trained Transformer) with actions. The process involves creating a GPT, setting up an actions server, defining an OpenAPI schema, and deploying the server as a regular Python application. The guide includes detailed instructions on how to set up the server environment, write code for the server, create an OpenAPI schema, and deploy the application using ngrok or another method. The final result is a functional GPT that can execute MySQL queries and respond accordingly.