November 2025 Summaries

As AI applications advance, the limitations of vector-only search systems are becoming increasingly apparent, especially in scenarios requiring complex data relationships and precision. Tensors, which are multi-dimensional numerical representations, offer a more structured alternative that preserves critical context such as sequence, position, and modality-specific structure, making them ideal for advanced retrieval tasks. Unlike vectors, tensors support richer data representations that can handle multimodal inputs across text, images, and video, and are integral to powering modern retrieval techniques. Vespa's tensor framework aims to address real-world challenges by offering a minimal, composable set of tensor operations, unified handling of dense and sparse data, and strong typing with named dimensions, thereby enhancing performance and reducing complexity. This approach enables more precise relevance scoring and personalized experiences in AI-driven applications, positioning tensors as foundational for future developments in this space.

Vector databases, crucial for modern AI systems through their ability to perform approximate nearest neighbor (ANN) searches for similarity-based retrieval, are facing limitations as retrieval-augmented generation (RAG) applications become more complex, requiring richer data representations across modalities like text, images, and video. These limitations include a lack of full-text search capabilities, inadequate integration with structured data and business logic, inflexible ranking systems, and the inability to perform real-time machine learning inference, all of which hinder personalization, hybrid relevance scoring, and real-time responsiveness. Additionally, the batch-oriented nature of many vector-native systems leads to stale results, and their inability to maintain spatial, linguistic, and temporal contexts in multimodal data further complicates their effectiveness. As the demand for precise, context-aware, and real-time results grows, the reliance on vectors alone is proving insufficient, suggesting a need for a more expressive foundation to meet the evolving needs of enterprise-scale AI applications.

Advent of Tensors 2025 is a 24-day coding challenge designed to help participants gain hands-on experience with tensors, focusing on multi-dimensional arrays and related concepts. Starting on December 1, 2025, a new challenge will be released daily, encouraging problem-solving and practical application through Vespa Tensor Expressions, with participants able to submit their solutions via a form. While the event offers prizes for those completing the most challenges, it emphasizes learning and enjoyment over competition. Participants are encouraged to engage at their own pace, join the Vespa Slack Community for discussion, and enjoy the festive atmosphere while exploring tensor computations.

During their summer internship at Vespa.ai, developers Erling Fjelstad and Edvard Wekre Dingsør embarked on a project to integrate a Model Context Protocol (MCP) server within the Vespa container, enhancing its interaction capabilities with large language models (LLMs). The MCP server, developed using both Python and Java, facilitates the querying of Vespa applications via natural language and supports tools for retrieving schema information, executing queries, and searching documentation. Initially created as a standalone server, the MCP server was later integrated into the Vespa container, requiring significant adaptation to Vespa's architecture and a deep dive into both the MCP SDK and Vespa engine codebase. The project not only provided practical experience in managing complex systems and evolving technologies but also highlighted the importance of agile development and collaboration in a tech startup environment. Despite challenges, the internship offered valuable lessons in innovation and teamwork, culminating in a fully functional tool that simplifies interactions with Vespa applications using natural language.

Protein Language Models (PLMs) are revolutionizing biologics discovery by learning the complex structures and dependencies of proteins, yet the current infrastructure struggles to manage and utilize their outputs effectively. These models generate high-dimensional embeddings, rich in biological meaning, which often go unindexed, limiting their utility in research and development. The integration of a PLM store as a memory system can transform these embeddings into a searchable database, enhancing the retrieval and analysis of biological data. Vespa.ai, a powerful platform for hybrid retrieval, offers a solution by storing vectors and metadata within a single schema, enabling real-time, context-aware queries that combine protein structure, experimental metadata, and sequence similarity. This approach allows researchers to perform complex queries, such as finding antibody variants with specific binding affinities and stability characteristics, with unprecedented speed and accuracy. Vespa.ai's ability to handle multi-modal retrieval in a single query plan positions it as a vital tool for advancing the field of AI-driven biology, ensuring that the full potential of PLMs is realized by making biological insights accessible and actionable.