April 2024 Summaries

In this text, the author discusses using Weaviate, an open source search engine, to locate Waldo in a "Where's Waldo" book. The author explains how Weaviate can convert natural language questions into vectors and find other nearby vectors to pinpoint relevant information. They also mention that Weaviate supports image data and performs the same type of "near image" search across a vector space. The author demonstrates this by splitting an input image into smaller chunks, converting each chunk into a vector, and then using Weaviate to find which chunk is similar to an image of Waldo. Finally, they reassemble the original image with the identified chunk highlighted. This technique has potential applications in various fields such as agriculture, insurance, and e-commerce.

Team Weaviate attended Google Cloud Next conference where they connected with community members, made new friends, and shared their latest product features and trend predictions. They observed that vector databases are becoming more popular in the AI stack, and enterprises are showing a preference for open-source solutions like Weaviate due to data compliance and security requirements. Many industries are building AI applications using Weaviate, with financial services, retail, and media being the most prevalent. There was significant interest in optimizing AI setups through concepts such as quantization and rescoring. Google Cloud's announcements included new products like VertexAI Agent Builder and updates to Gemini. The team also hosted a breakout session on deploying Weaviate on Google Kubernetes Engine, and announced upcoming events including Google I/O and an in-person event with Google Cloud in New York City.

Unbody is an AI-as-a-service platform that automates the entire AI pipeline from data to the front-end, all in one API. The platform integrates data from various sources and allows users to select their preferred AI engine. It simplifies complex development tools by providing a unified, straightforward interface. Unbody's transformation with Weaviate has significantly enhanced its capabilities, enabling personalized, secure, and efficient data processing. Since launching at Websummit in November 2023, the platform has grown to 250 beta users and continues to develop new features and open-source projects showcasing its capabilities.

Language models have revolutionized AI interaction through prompt engineering, which is more an art than a science. DSPy introduces two new concepts for better control over generative AI systems: a programming model and optimizers or compilers. The DSPy programming model lets users define components of language model requests, while the compilers provide prompts that enhance LLM performance. Optimization strategies in machine learning have been transformed with LLMs, which no longer require gradient descent training but can be tweaked through prompting. DSPy's BootstrapFewShot, COPRO, and MIPRO optimizers help users tune instructions and create high-quality examples for better responses from LLMs.

Weaviate has been named one of Forbes' AI 50, a list of "AI startups to watch" developed by Forbes with Sequoia Capital and Meritech Capital. The selection process involves two separate judging panels that assess the companies from both technical and investment perspectives. Previous selectees include Hugging Face, Anthropic, and Databricks. Weaviate's inclusion in this list is a testament to its functional performance, reliability, ease of use, and value addition. As generative AI moves into production, the focus shifts from being interesting or cool to being functional, reliable, easy to work with, and adding value. The company aims to stay on the cutting edge and bring its customers along as it explores new applications for vector search and other emerging technologies.

This tutorial demonstrates how to build a Next.js application powered by hybrid search in Weaviate. Hybrid search combines keyword (BM25) and vector search to leverage both exact term matching and semantic context. The tutorial covers the technology that powers hybrid search, how to use it in Weaviate, and how to adjust its parameters for better results. It also provides a step-by-step guide on setting up Weaviate Cloud, importing data, creating a Next.js backend, and developing a frontend for the application.

Binary Quantization (BQ) is a vector compression algorithm that reduces memory requirements while trading off retrieval accuracy. It simplifies vector encoding by retaining only their directionality, with each dimension encoded as a single bit indicating whether it's positive or negative. This technique works well for high-dimensional vectors and can significantly reduce the amount of space required to store them. BQ also enables faster distance calculations between compressed binary vectors using bitwise operations. However, its effectiveness depends on data distribution and dimensionality.