February 2024 Summaries

Weaviate 1.24 introduces named vectors, allowing a single object to have multiple vectors with different indexes, compression methods, and vectorizers. This version also improves binary quantization (BQ) for faster and more memory-efficient vector search. Additionally, it simplifies Docker configuration with a new image that requires no configuration, enhances backend improvements for updates, imports, and deletions, and updates the Python client to support all latest features.

Weaviate has integrated a gRPC API into its solutions, starting with the Python client library, resulting in significant speed improvements compared to previous REST + GraphQL solutions. The new gRPC-based data imports show significant reductions in time taken for both datasets tested. Query throughput is also significantly improved, with each query taking 40-70% less time to return results when using gRPC in comparison to REST + GraphQL. This results in a significant, over 2.6x, improvement in how many queries can be served.

Weaviate, a vector database, introduced reranking at the second stage in version 1.20. This feature allows users to improve search relevance by adding reranking to the second-stage of their search process. Cohere's rerank endpoint enables users to build search systems that add reranking at the last stage, and fine-tuning boosts the model's performance in unique domains. The blog post demonstrates how to fine-tune Cohere's reranker model using Weaviate's blogs dataset and DSPy's signature and chain-of-thought module. It also explains how to re-index data with a new schema that includes the fine-tuned model ID, and how to query the database with and without reranking.

A Generative Feedback Loop is a powerful use case that combines a Large Language Model (LLM) with semantic search. This technique can be used to enrich, enhance or summarize data, enabling new opportunities for searching through information and providing better search results. In this blog post, the author demonstrates how to implement a Generative Feedback Loop using open-source software like Weaviate, Ollama, and Mistral. The process involves loading product review data into Snowflake, deploying Weaviate, Ollama, and a vectorizer within Snowflake's processing engine (SPCS), generating product metadata using the latest Mistral model, and performing semantic and hybrid searches using Weaviate. This approach allows users to search for experiences, feelings, keywords, or product names of products while keeping all data within Snowflake.

The Weaviate Python client library has reached general availability with a new API that is more powerful and easier to use. This update includes collection-based interactions and default property returns in queries, making the code more concise and user-friendly. Additionally, the v4 client offers IDE assistance/autocomplete, speed improvements through gRPC usage, and enhanced typing and type safety. The new client also provides batch operations with rate limiting capabilities and improved error handling during batching operations.