May 2024 Summaries

This article provides an overview of commonly used evaluation metrics in search and recommendation systems, including precision@K, recall@K, MAP@K, MRR@K, and NDCG@K. These metrics can be categorized into not rank-aware vs. rank-aware metrics, with the latter considering both the number of relevant items and their position in the list of results. The article also demonstrates how to calculate each metric using Python's pytrec_eval library and provides a minimal example dataset for illustration purposes.

This article demonstrates how to implement a local Retrieval-Augmented Generation (RAG)-based chatbot in Python using open source components such as Ollama for language models and Weaviate vector database via Docker. The process involves setting up the local LLM and embedding models with Ollama, hosting a local vector database instance with Docker, and building a local RAG pipeline. This approach ensures data privacy by keeping everything on-premises without any dependencies on external services or API keys.

Weaviate 1.25 introduces dynamic vector indexing, enabling efficient scaling as data grows by dynamically switching from flat indexes to HNSW. Raft improves schema management and makes multi-node clusters more reliable. New modules allow users to host and run open source embedding and language models locally. Batch vectorization speeds up data import, while automatic tenant creation simplifies the process. Search improvements include hybrid search with vector similarity search and grouping methods, as well as keyword search with grouping.

In March, a roadshow in San Francisco featured tech leaders discussing AI applications in production, with a focus on intelligent document processing and automation technologies. Instabase leverages AI to streamline operations across various industries, using innovative approaches and strategic implementation of technologies like Weaviate. Key takeaways from the discussion include the use of generative AI for complex problem-solving, flexibility in technology deployment, and active product evolution to support customers' needs. The full video provides further insights into Instabase and Weaviate's work in GenAI and engineering.

The text discusses how a full application that discovers analogies between Wikipedia articles was built by combining serverless infrastructure from Modal with the search and storage capabilities of Weaviate. Modal is used for computing vector embeddings, while Weaviate provides both a scalable, managed deployment and all the knobs needed to configure indexing to maximize speed. The application uses async indexing, product quantization, vector index configuration, text search configuration, and batch imports to improve performance. The author recommends using Modal and Weaviate together for data- and compute-intensive applications of generative models and artificial intelligence.