October 2023 Summaries

Qdrant, a vector search engine, has enhanced its Recommendation API in version 1.6, offering greater flexibility and control for users. The updated API allows for the use of both vector IDs and embeddings for positive and negative examples, enabling more sophisticated recommendation strategies. The previous API required at least one positive example and used an average vector approach, which has now been supplemented with a new "best score" strategy that selects recommendations based on the closest distances to positive and negative samples. This advancement is exemplified by the Food Discovery demo, which demonstrates the API's application in recommending meals based on user preferences and dislikes, allowing for the inclusion of multiple queries and strategy switching. The demo also highlights challenges with multimodal data, such as combining text and image embeddings, which can affect the results depending on the proximity of the embeddings in the dataset. The open-source demo is available online, with precomputed embeddings for easy deployment on platforms like Qdrant Cloud.

FastEmbed is a Python library developed by Qdrant to facilitate efficient and user-friendly embedding generation for data science and machine learning applications. Targeting 80% of NLP embedding use cases, it simplifies the process by offering default workflows and a small, focused set of transformer models, such as BAAI/bge-small-en-v1.5. FastEmbed is optimized for speed and performance by quantizing models and integrating them with ONNX Runtime, enabling fast computations even on CPUs without the need for specialized hardware. The library's design minimizes installation time and resource requirements, making it suitable for environments with storage limitations. FastEmbed seamlessly integrates with Qdrant, a vector store that enhances embedding generation, storage, and retrieval, thereby providing scalable and efficient solutions for large-scale datasets. The fusion of FastEmbed with Qdrant's capabilities offers a streamlined approach to handling text embeddings, reducing latency, and supporting extensive data operations.

Zein Wen, as a Google Summer of Code 2023 participant, collaborated with mentor Arnaud Gourlay to enhance the Qdrant vector database by developing a Polygon Geo Filter, which adds flexibility to geo-data queries by allowing users to refine results using polygons in addition to existing circular and rectangular filters. This feature was designed to align with the complex geographic boundaries users encounter, particularly in applications like restaurant recommendations that require neighborhood-specific filtering. The development process involved overcoming technical challenges such as implementing efficient geometry computations using geo hash layers and ensuring consistency in ProtoBuf and JSON interfaces, which led to the decision against a separate multi-polygon filter to maintain simplicity. This experience expanded Wen's problem-solving abilities and emphasized the importance of user-centric design and effective communication within development teams. Wen expressed gratitude for the support received during the program, which fostered a deeper understanding of open-source collaboration and inspired continued contributions to Qdrant, advocating for its adoption within the tech community.