May 2023 Summaries

Large Language Models (LLMs) have revolutionized how we access and consume information, shifting from a search engine market that was predominantly retrieval-based to one now that is growingly memory-based and performs generative search. However, the wide-scale adoption of LLMs raises concerns around privacy and data security. To leverage the advantages of generative AI while simultaneously addressing these privacy concerns, the field of privacy-preserving machine learning has emerged, offering techniques and tools such as federated learning, homomorphic encryption, and locally deployed LLMs. These approaches allow for the secure execution of large language models while protecting the confidentiality of sensitive data both during model fine-tuning as well as when providing responses grounded in proprietary data.

The latest advancements in multimodal deep learning have made it possible to extract high quality data from PDF documents and add it to a Weaviate workflow. Optical Character Recognition (OCR) technology is used to convert different types of visual documents into machine-readable formats, with new models like LayoutLMv3 and Donut leveraging both text and visual information using multimodal transformers. Unstructured, an open-source company working at the cutting edge of PDF processing, allows businesses to ingest diverse data sources and convert them into data that can be passed to a Language Learning Model (LLM). This enables users to chat with their PDFs by converting private documents from their company into text format.

Weaviate has introduced two new modules that integrate Google's Pathways Language Model (PaLM) with its vector database technology. The text2vec-palm module generates vector embeddings for semantic searches, while the generative-palm module enables generative search capabilities. These features allow users to leverage PaLM's large language model and customize it according to their needs.

Generative Feedback Loops is a concept where we use data from the database to supplement the factual knowledge of generative models and then write the generated outputs back to the database for future use. This technique can be used in various applications such as generating advertisements, summarizing podcasts, categorizing tweets, and even building autonomous AI assistants like AutoGPT. By saving intermediate results, we can create a feedback loop that improves the performance of generative models over time.

Weaviate 1.19 brings a set of new features, performance improvements, and bug fixes. Key updates include: group by arbitrary property for search results; upgraded text data type with more tokenization options and deprecation of string; introduction of dedicated uuid data types; availability of the generative-cohere module for leveraging large language models on custom datasets; tunable consistency support for vector search and GraphQL Get requests; experimental gRPC API support for faster queries; and other improvements and bug fixes.

Weaviate has announced a significant upgrade to its Weaviate Cloud service, now in public beta. The fully managed vector database as a service enables AI application developers to use the open source Weaviate vector database easily and quickly. Key improvements include automated provisioning of Weaviate database clusters, availability on Google Cloud with options for other clouds and on-premises hosting, hybrid SaaS option, updated features from the latest Weaviate open source version, enhanced user experience through an interactive admin console, API key management, and planned general availability later in 2023.