spb/wp13_uqo

1% ============================================================================2\section{Introduction}3\label{sec:introduction}4% ============================================================================56The rapid maturation of large language models (LLMs) has transformed expectations about7what conversational software can accomplish in finance. Models such as the Claude, GPT, and8Gemini families now produce fluent commentary on earnings, valuations, and macroeconomic9conditions, and a growing literature documents their ability to extract signal from financial10text \citep{lopezlira2023chatgpt,kim2024financial,wu2023bloomberggpt}. Yet fluency is not the11same as rigor. A model that \emph{describes} a value-at-risk calculation in prose has not12\emph{computed} one; a model that \emph{recalls} a stale price from pre-training has not13\emph{retrieved} the current quote; and a model that narrates a Monte Carlo experiment has14not run the simulation whose numbers a portfolio manager would act upon. The central15engineering problem in applied financial AI is therefore not generation but16\emph{grounding}: how to bind a probabilistic text generator to live data, deterministic17computation, and reproducible artifacts without sacrificing the conversational interface that18makes the technology accessible.1920This paper documents the methodology and innovations of \textbf{VQuant}, a production-grade21web platform built to address precisely this problem. VQuant accepts a natural-language22request---``analyze NVDA's fundamentals, technicals, and run a Monte Carlo simulation,'' for23example---and returns a complete, multi-section quantitative report containing live data,24computed figures, and downloadable documents. The platform is implemented as a full-stack25TypeScript monorepo: a React front end, a modular Express back end, and an isolated Python26scientific-computing layer. At its core, the system orchestrates a reasoning-augmented LLM27through an \emph{agentic loop} that exposes more than two hundred financial and analytical28tools to the model, executes the model's tool calls deterministically, streams the results29back in real time, and synthesizes a final report that can be exported in five document30formats.3132The motivation for studying this system is fourfold. \emph{First}, VQuant operationalizes a33design philosophy that we believe is the correct one for high-stakes numerical domains: the34LLM is a \emph{planner and synthesizer}, never a calculator. Every number in a VQuant report35originates from a deterministic source---a market-data API, a closed-form pricing formula, or36auditable Python code---so that the analysis is reproducible and the model's well-known37propensity for numerical hallucination is structurally contained. \emph{Second}, the platform38confronts the practical constraints of agentic systems---bounded context windows, tool-call39latency, rate limits, and cost---with concrete algorithmic solutions whose logic generalizes40beyond finance. \emph{Third}, VQuant integrates an unusually broad tool ecosystem (over 26541market-data endpoints, four web-research services, and a seventeen-library Python sandbox)42under a single function-calling protocol, providing a useful case study in tool taxonomy and43prompt-level routing. \emph{Fourth}, the system closes the loop from query to deliverable by44generating publication-ready PDF, Word, \LaTeX{} Beamer, and spreadsheet outputs from a single45internal representation, illustrating a reusable pattern for automated report generation.4647Our contributions are methodological and architectural rather than empirical. We make four48specific contributions. \textbf{(1)~Agentic orchestration with constrained tool use.} We49describe a streaming tool-calling loop (Section~\ref{sec:methodology}) in which a strategically50engineered system prompt routes requests to an appropriate subset of more than two hundred51tools, batches tool execution to respect external rate limits, and surfaces the model's52intermediate reasoning to the user in real time. \textbf{(2)~Context management under a token53budget.} We formalize an adaptive-summarization algorithm that preserves coherence across54effectively unbounded conversation histories by compressing older turns once a token threshold55is crossed, while retaining the most recent exchanges verbatim. \textbf{(3)~A single-call56computation paradigm.} We argue that consolidating data acquisition, numerical estimation, and57visualization into one sandboxed Python execution---rather than the sequential58fetch-then-analyze pattern typical of naive agents---materially reduces end-to-end latency, and59we describe the sandbox and its safety properties (Section~\ref{sec:capabilities}). \textbf{(4)~A60unified multi-format document pipeline.} We detail how a single Markdown artifact is61transformed into PDF (headless-browser rendering with mathematical typesetting), DOCX (document62converter), Beamer slides (model-structured \LaTeX{} compiled to PDF), and spreadsheets with63live formulas.6465Our emphasis throughout is on \emph{mechanism}. We describe each component as it is actually66built---the byte-level streaming protocol, the recursion that drives the loop, the throttle that67batches tool calls, the two-stage machinery that keeps the context within budget, the correction68that keeps token accounting honest, and the inter-process boundary that executes model-authored69code---so that the design is reproducible rather than merely sketched. The classical quantitative70methods the platform automates (option pricing, volatility modeling, simulation, and risk71metrics) are standard and not the contribution of this paper; we record them compactly in an72appendix and otherwise keep the focus on the architecture.7374The remainder of the paper is organized as follows.75Section~\ref{sec:literature} situates VQuant within the literatures on LLM agents, tool use,76retrieval augmentation, and AI in finance.77Section~\ref{sec:architecture} presents the system architecture, data layer, and external78integrations.79Section~\ref{sec:methodology} develops the agentic orchestration methodology, including the80tool taxonomy, context-management algorithm, and single-call execution paradigm.81Section~\ref{sec:capabilities} formalizes the embedded quantitative methods and the82document-generation pipeline.83Section~\ref{sec:evaluation} discusses performance, reliability, security, and limitations.84Section~\ref{sec:conclusion} concludes and outlines directions for evaluation and extension.85