January 2026 Summaries

A tutorial outlines the process of building a real-time research paper discovery and briefing system using Exa and Fireworks AI technologies. Exa functions as a search engine to access real-time data on scientific papers from sources like arXiv and OpenReview, while Fireworks provides a scalable inference platform for embedding and summarization. The system involves several steps, including setting up API keys, configuring search parameters, extracting text content, and optionally reranking documents using Fireworks embeddings. Summaries of selected papers are generated using the DeepSeek V3.1 language model, which provides concise insights into the core ideas, methods, results, and relevance of the research. The tutorial includes a step-by-step code walkthrough and free API credits to enable users to implement the system themselves.

Building a modern search engine involves managing the complexities of ingesting and querying the ever-changing web, characterized by heterogeneous content, varying update frequencies, and sheer volume. The in-house data processing framework, exa-d, was developed to address these challenges by optimizing typed columns with declarative dependencies, enabling engineers to focus on data relationships rather than update steps. This approach allows for efficient management of data updates, whether through surgical updates or full rebuilds, without unnecessary rewrites, thanks to exa-d's ability to identify affected rows and columns precisely. The framework ensures efficient parallel execution by distributing workloads across heterogeneous resources and minimizes redundant computation by leveraging a storage model that tracks data completeness. Using Ray Data for query planning, exa-d computes only necessary updates, maintaining a dynamic and scalable search index. As the web evolves, exa-d continues to adapt, offering a robust solution for maintaining derived states over an extensive web index.

Exa has introduced its Company Search Benchmarks to enhance the precision of AI-driven company searches, expanding on its existing People Search Benchmark. This initiative aims to differentiate the retrieval model's capabilities from memorized knowledge by focusing on fresh, structured data rather than well-known entities. The benchmark involves a dataset comprising diverse company attributes like founding year, location, and funding history, specifically avoiding prominent unicorns. It evaluates both static and dynamic facts, requiring different assessment approaches, and includes a retrieval track to ensure systems return accurate company results for various query types. By open-sourcing the dataset and evaluation harness, Exa seeks to foster advancements in retrieval research, promoting a comprehensive evaluation ecosystem across different entity types and search domains in line with its mission to build perfect search.