Shaping the future through cutting-edge research and development

We adapt and develop state-of-the-art methodologies to optimize how agents interact with data retrieval tools and how contextual queries are routed between multiple agents. Methods include dynamic prompting, query expansion, self-reflection, and intelligent routing across any number of datasets, which enable LLMs to effectively leverage S&P Global’s's AI-ready data. We also develop agentic architectures built to orchestrate complex workflows such as conducting research and generating long-form content.

We explore and develop cutting-edge strategies for unstructured and structured data retrieval, enabling agentic workflows atop complex financial data. We push unstructured data retrieval far beyond naive vector search RAG, leveraging structural information and relationships from our document processing toolkit to enrich data for complex data discovery and question answering at performant inference speeds. For structured and tabular data, we combine LLM-optimized tools with a proprietary Text-to-SQL framework. Across both, we apply strategies such as reinforcement learning, self-consistency, and self-reflection to optimize our agents for financial datasets to deliver sharper customer insights across complex data sources.

We research and develop multi-tiered evaluation frameworks and benchmarks for finance and business. Our evaluations assess models across a broad spectrum of tasks, from simple data retrieval and long-context QA to advanced multi-step reasoning and program synthesis. These frameworks support complex agentic products spanning tool usage, code and SQL generation, computer use, and report generation. To measure both end-to-end solution quality and individual component performance, we employ diverse strategies—ranging from AST parsing to LLM-juries calibrated by domain experts—while rigorously auditing data relevance and output consistency. Beyond applied metrics, we actively research and challenge the fundamentals of evaluation itself. We explore critical questions such as the reliability of LLM-as-a-judge, the impact of high-quality labels, methods for efficiently predicting model performance on unseen evaluations, and the real-world properties that current benchmarks fail to address. Ultimately, our comprehensive approach enables data-driven decisions and continuous model improvement.

We analyze and improve tokenization methods, which determine how all input data should be parsed and represented to a model. Tokenization naturally impacts a model's ability to process and understand data such as human language text,  and remains an under-explored LLM  research frontier. We develop fundamental tokenization capabilities we can use to better present data to LLMs, including our rich, domain-specific business and finance data. This work includes drastically improving the speed and efficiency of training and using tokenizers, discovering LLM limitations with processing numeric data in various languages and scripts, moving beyond the restrictions of the ubiquitous BPE tokenizer, and building our own enhanced tokenizer.

We explore how to best harness the structural and relational information within and across documents, enabling models to move beyond surface-level text and achieve richer document understanding. Our research across document components and their relationships informs ingestion strategies that ensure both structural and semantic coherence as well as providing better structured more usable content downstream. This coherence and richer representation unlocks greater performance and even new capabilities compared with traditional extraction. Models need this context and structure to answer targeted questions, surface broader insights across wider knowledge bases, and ground validated assertions against a document corpus.

We are on a mission to extract and structure all information contained within documents, from plain text to the most complex visual elements. Our proprietary models extract figure data from document images and graphs, unlocking highly accurate quantitative question-answering capabilities. In our pursuit of comprehensive document AI, we are also building  the next generation of table extraction models including Vision Language Models (VLMs) to parse tables in their natural context within page images.