September 2023 Summaries

Qdrant's latest innovation, binary quantization (BQ), optimizes vector search by converting high-dimensional vectors into binary values, reducing memory usage and improving retrieval speeds by up to 40 times while allowing users to balance speed and recall accuracy during searches. This method is particularly beneficial for large vector lengths, such as OpenAI's 1536-dimensional embeddings, which can be compressed from 32-bit to 1-bit, offering significant storage efficiency and faster boolean operations. While BQ can degrade recall accuracy, especially with small embeddings, it excels in managing large datasets with high recall expectations by using a binary index that oversamples a smaller vector subset for precise search results. The implementation involves storing full vectors on disk and binary vectors in RAM, with the ability to adjust search parameters like oversampling and rescoring to optimize performance and accuracy. This approach is ideal for scenarios requiring rapid data processing but may not be suitable for smaller embeddings due to potential accuracy losses.

The Food Discovery Demo is an open-source project designed to help users explore food options through a semantic search interface, particularly when their preferences are not clearly defined. Utilizing a FastAPI backend, a React frontend, and a Qdrant instance, the demo employs a CLIP model to encode images and text into a shared vector space, allowing for effective image-based searching and recommendations. The system addresses challenges such as the cold start problem by presenting random dish selections initially and enables textual and location-based searches to enhance user interaction. Users can provide feedback through liking or disliking dishes, which updates search results accordingly, and the demo supports both positive and negative feedback to refine recommendations. The underlying dataset includes over 2 million images sourced from the Wolt dataset, and the demo's architecture facilitates both local deployment via Docker and cloud deployment through Qdrant Cloud. The project is accessible on GitHub, allowing users to fork and modify it for personalized use cases.