March 2024 Summaries

Vector distance calculations are crucial in vector databases as they determine the similarity between data objects. Weaviate, a vector database, spends up to 60% of its CPU time on these calculations. To improve performance, leveraging algorithmic and hardware optimizations is essential. Intel's new 5th Gen Xeon Processor - Emerald Rapids enables significant speedups in vector search operations through advanced SIMD instruction sets like AVX-256 and AVX-512. By using these instructions with loop unrolling, Weaviate achieved a ~40% QPS speed up at 90% Recall on Intel's new Xeon Processor, Emerald Rapids.

Weaviate has released a beta version of its new TypeScript client, available on npm under the name weaviate-client. The update includes changes that mirror those made in Python, such as better TypeScript support and performance improvements. Developers can try out the new client by installing it via npm or yarn and providing feedback through GitHub issues.

Generative Feedback Loops (GFLs) are a concept that involves saving outputs from generative AI systems back into databases for future use. This approach has been extended beyond AirBnB listings to Amazon products, and even tested within Snowpark Container Services. GFLs can enhance data by creating new types of applications with DSPy, improving user experience design, and offering a major opportunity to enhance these experiences further. The Hurricane project is an example of using GFLs for generating Weaviate blog posts and feeding them back into the system. It involves breaking down tasks into subtasks, stateful processing, and optimizing prompts with COPRO and MIPRO. DSPyOps is a new concept that emphasizes logging and call monitoring in AI applications, similar to LLMOps but with an added focus on optimization.

Weaviate has launched its Hero Program, aiming to recognize and celebrate members who contribute significantly to their community. With over 25K members across various platforms like LinkedIn, Twitter, GitHub, Slack, Discourse, and StackOverflow, the Weaviate Community is active in discussing recent developments in Open Source and AI, sharing knowledge on building amazing applications with Weaviate, collaborating, networking, and contributing to open-source projects. The program focuses on four core pillars: Contributions & Engagement, Collaboration & Culture, Benefits & Rewards, and Education & Elevation. Three official Weaviate Heroes have been introduced, including Ron Parker, Patrice Bourgougnon, and Travis Rehl. To become a Weaviate Hero, one can start engaging in the open-source community by collaborating with other members on various platforms or apply directly if already highly engaged with the Weaviate Community.

OpenAI ChatGPT was introduced in November 2022, revolutionizing language models with its technology that had been around since a Google paper in 2017. Weaviate has been exploring semantic relationships of vector embeddings for these transformer models since then. The debut of ChatGPT led to increased innovation across industries as they integrated Generative AI use cases and features into their business offerings. Weaviate, an open-source AI Native vector database, enables businesses to build applications like chatbots, analytics-driven automated marketing, generative feedback loops, and ecommerce recommendations. The future of GenAI applications in the enterprise is promising as more innovative tools are being developed.

Retrieval Augmented Generation (RAG) is an advanced search technique that uses a Large Language Model (LLM) to help identify and summarize answers. RAG applications search for relevant sources and create personalized answers using provided data sources instead of general knowledge LLMs are trained on. Verba, the Golden RAGtriever, is an open-source tool designed to make it easy for anyone to use RAG without technical expertise. It allows users to get personalized answers from their own data and supports customization for different use cases. The architecture of Verba consists of five key steps: ReaderManager, ChunkerManager, EmbeddingManager, RetrieveManager, and GenerationManager.