STDM：让数据“自我思考”的交互式文档规范

STDM 使数据制品（包括文档）能够引导大型语言模型（LLM）对其进行解读，从而创造出交互式、自我导向的体验。这种方法既保留了作者的原始意图，又解锁了新的分析能力。

STDM enables data artifacts (including documents) to guide large language models (LLMs) in interpreting them, thereby creating interactive, self-directed experiences. This approach preserves the author's original intent while unlocking new analytical capabilities.

⚠️ 注意事项

v0.1 规范详述的是一个实验性概念。实施时需要考虑规范中概述的安全原则。虽然 STDM 旨在增强数据效用并保留作者意图，但它涉及引导 LLM 行为，这本身带有风险，包括可能导致意外行为。规范中包含了缓解策略，但社区在应用时应保持适当的谨慎并充分考虑安全因素。

The v0.1 specification details an experimental concept. Implementation requires consideration of the security principles outlined in the specification. While STDM aims to enhance data utility and preserve author intent, it involves directing LLM behavior which carries risks, including causing unexpected behaviour. The specification incorporates mitigation strategies, but the community should apply appropriate caution and security considerations.

什么是 STDM？

STDM 在文档、图像或数据文件中嵌入结构化指令，这些指令指导 LLM 应如何处理、分析和呈现数据。这可以创建出“智能制品”，例如：

STDM embeds structured instructions within documents, images, or data files that direct how LLMs should process, analyze, and present the data. This creates "smart artifacts" like:

• floodplain.html - 开放获取论文《当前和未来气候条件下墨累-达令盆地的洪泛区淹没》的交互式 HTML 海报。

• floodplain.pdf - 交互式 PDF 版本。

• floodplain.html - An interactive HTML poster of the open access paper "Floodplain inundation in the Murray–Darling Basin under current and future climate conditions."

• floodplain.pdf - An interactive PDF version.

支持 STDM 的数据能够：

STDM enabled data is capable of:

• 渲染具有自定义导航功能的交互式界面 (Rendering interactive interfaces with custom navigation)

• 对嵌入内容执行数据分析 (Executing data analysis on embedded content)

• 创建复杂文档的引导式探索 (Creating guided exploration of complex documents)

• 提供针对特定数据的智能辅助 (Providing intelligent assistance tailored to specific data)

• 在交互过程中保持上下文和目的 (Maintaining context and purpose across interactions)

视觉指示器

“我是思考数据”丝带：支持 STDM 的文档会在右上角用彩色的“我是思考数据”丝带进行视觉标记。此丝带：

I am Thinking Data Ribbon: STDM-enabled documents are visually marked with a colorful "I am Thinking Data" ribbon in the top-right corner. This ribbon:

• 提供一个清晰的视觉指示，表明文档包含 STDM 指令 (Provides a clear visual indicator that the document contains STDM instructions)

• 帮助用户识别具有嵌入式智能的文档 (Helps users identify documents with embedded intelligence)

• 如果需要，可以通过单击丝带中的“×”来关闭 (Can be dismissed by clicking the "×" in the ribbon if desired)

• 包含在诸如洪泛区可视化等示例中 (Is included in examples like the floodplain visualization)

核心概念与常见问题

Q1: 什么是自我思考数据清单（STDM）？

A: 自我思考数据清单（STDM）是一种数字制品（如 HTML 页面、文本文件，甚至是 PDF 或图像中的元数据），它将主要数据内容（通常是文本）与明确的、结构化的指令捆绑在一起。这些指令定义了大型语言模型（LLM）作为外部解释引擎，应如何处理、交互、呈现或执行与该制品嵌入数据相关的任务。

A: A Self-Thinking Data Manifest (STDM) is a digital artifact (like an HTML page, text file, or even metadata within a PDF or image) that bundles primary data content (often text) with explicit, structured instructions. These instructions define how a Large Language Model (LLM), acting as an external interpretation engine, should process, interact with, present, or execute tasks related to the artifact's embedded data.

• 核心思想：目标是创建自我导向的制品。STDM 不仅仅是被动数据，它还包含一个蓝图，指导 LLM 如何提供针对该数据的特定、可预测且可能具有交互性的体验。“自我思考”这个术语形象地指代了清单为 LLM 关于该数据的行动提供这些自包含方向的能力。核心原则是：“清单指导引擎对数据的‘思考过程’和用户体验。”

• 自然语言编程STDM利用现代LLM的高级指令遵循能力，通过结构化自然语言指令（如GOAL、CONSTRAINTS、REQUESTED_TOOLS、CUSTOM_UI_DEFINITION）有效地“编程”LLM在特定数据上下文中的行为。：STDM 利用了现代 LLM 先进的指令遵循能力。通过使用结构化的自然语言指令，如 GOAL（目标）、CONSTRAINTS（规则/边界）、REQUESTED_TOOLS（所需能力）和 CUSTOM_UI_DEFINITION（呈现/交互格式），作者可以有效地为该特定数据上下文“编程”LLM 的行为。LLM 解释这些高级指令，就像计算机解释代码一样。

• 颠覆“数据聊天”模式：传统上，让数据与 LLM 交互通常需要在数据之上构建单独的应用程序或聊天界面。STDM 颠覆了这种模式。它将交互逻辑、角色设定和呈现指南直接嵌入到数据制品本身。当用户将此制品提供给兼容的 LLM 解释器并明确命令运行它时，该制品本质上会带来其自定义的界面和交互模式，并由 LLM 驱动。

• Core Idea: The goal is to create self-directing artifacts. Instead of just passive data, an STDM contains a blueprint guiding an LLM on how to provide a specific, predictable, and potentially interactive experience tailored to that data. The term 'Self-Thinking' evocatively refers to the manifest's ability to provide these self-contained directions for the LLM's actions concerning the data. The core principle is: "The manifest directs the engine's 'thought process' and user experience of the data."

• Natural Language Programming: STDM leverages the advanced instruction-following capabilities of modern LLMs. By using structured natural language directives like GOAL (the objective), CONSTRAINTS (rules/boundaries), REQUESTED_TOOLS (needed capabilities), and CUSTOM_UI_DEFINITION (presentation/interaction format), authors can effectively "program" the LLM's behaviour for that specific data context. The LLM interprets these high-level instructions much like a computer interprets code.

• Inverting "Chat Over Data": Traditionally, making data interactive with an LLM often involves building a separate application or chat interface on top of the data. STDM inverts this model. It embeds the interaction logic, persona, and presentation guidelines directly within the data artifact itself. When a user provides this artifact to a compatible LLM interpreter and gives an explicit command to run it, the artifact essentially brings its own custom interface and interaction pattern with it, powered by the LLM.

Q2: STDM 与标准提示工程有何不同？

A: 虽然两者都涉及向 LLM 提供指令，但 STDM 将这些指令与数据进行了形式化和捆绑。

A: While both involve providing instructions to an LLM, STDM formalizes and bundles these instructions with data.

Q3: 为什么现在是引入 STDM 这类概念的合适时机？

A: 具有显著更大上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。的大型语言模型的出现是一个关键推动因素。

A: The emergence of Large Language Models with significantly larger context windows is a key enabler.

• 以前，LLM 的上下文窗口LLM处理输入文本时的长度限制，超出部分可能被截断或忽略，影响模型对长内容的整体理解。通常太小，无法同时容纳大量的数据内容和详细的处理指令。

• 现代 LLM 可以处理数十万甚至数百万个令牌。这使得整个文档、数据集（以文本形式）或复杂配置文件可以加载到 LLM 的上下文中，同时附带一套全面的 STDM 指令（包括目标、约束、UI 定义和后备逻辑）。

• 这种技术能力使得 STDM 的核心概念——将数据与其自身的解释指南捆绑在一起——变得切实可行。

• Previously, LLM context windows were often too small to hold both substantial data content and detailed processing instructions simultaneously.

• Modern LLMs can process hundreds of thousands, or even millions, of tokens. This allows an entire document, dataset (in textual form), or complex configuration file to be loaded into the LLM's context along with a comprehensive set of STDM instructions (including goals, constraints, UI definitions, and fallback logic).

• This technical capability makes the core STDM concept – bundling data with its own interpretation guide – practically feasible.

Q4: STDM 是否只是在促成或便利恶意提示注入？

A: STDM 承认与 LLM 处理外部指令相关的风险，包括恶意提示注入。它提出了围绕用户控制和明确意图的具体机制来缓解这些风险。

A: STDM acknowledges the risks associated with LLMs processing external instructions, including malicious prompt injection. It proposes specific mechanisms centered on user control and explicit intent to mitigate these risks.

Q5: 用户如何知道文档是否包含 STDM？

A: 此规范（v0.1）侧重于格式和 LLM 解释。包含 STDM 的制品应能被明确识别（例如，通过标签或命名约定），以促进用户在交互前的认知。其他机制包括：

A: This specification (v0.1) focuses on the format and LLM interpretation. Artifacts containing an STDM should be clearly identifiable as such (e.g., via labelling or naming conventions), promoting user awareness before interaction. Additional mechansisms include:

• LLM 解释器通知：在收到文档和明确的“运行”命令后，LLM 解释器应理想地通知用户它已检测到 STDM 指令，并在继续之前说明 GOAL（遵循安全前言指导）。

• 文件命名/元数据：作者可以使用命名约定（例如 .stdm.html）或明确的元数据字段。

• 视觉指示器：文档可以包含图标或通知，表明存在嵌入式指令。提供了一个示例横幅。

• 透明度原则：虽然指令不应干扰正常查看，但它们应可供审查。如何在不同文件类型中最好地实现这一点是一个有待探索的领域。

• LLM Interpreter Notification: Upon receiving a document and an explicit "run" command, the LLM interpreter should ideally notify the user that it has detected STDM instructions and state the GOAL before proceeding (as guided by the Safety Preamble).

• File Naming/Metadata: Authors could use naming conventions (e.g., .stdm.html) or explicit metadata fields.

• Visual Indicators: Documents could icon or notice indicating the presence of embedded instructions. An example banner is provided.

• Transparency Principle (Sec 2.3): While instructions shouldn't interfere with normal viewing, they should be accessible for review. How this is best implemented in different file types is an area for exploration.

开始使用

• 规范：完整的 STDM v0.1 规范 解释了格式、结构和推荐的实现模式。

• 实时演示：查看已渲染的规范，它本身就是一个实验性的 STDM 文档。

• 示例：查看示例目录以获取各种格式的实现示例。

  • 洪泛区淹没研究：floodplain.html - 开放获取论文《当前和未来气候条件下墨累-达令盆地的洪泛区淹没》的交互式 HTML 海报。此示例嵌入了完整的论文，并提供了 STDM 指令，使 LLM 能够逐页导航和分析内容。

• Specification: The complete STDM v0.1 Specification explains the format, structure, and recommended implementation patterns.

• Live Demo: View the rendered specification which is itself an experimental STDM document.

• Examples: Check the examples directory for implementation samples in various formats.

  • Floodplain Inundation Study: floodplain.html - An interactive HTML poster of the open access paper "Floodplain inundation in the Murray–Darling Basin under current and future climate conditions." This example embeds the complete paper and provides STDM instructions that enable LLMs to navigate and analyze the content page by page.

开发

规范的 HTML 版本是从 Markdown 源文件构建的：

The HTML version of the specification is built from the markdown source:

python build/build.py stdm_spec_v0.1.md instructions.txt template.html index.html

STDM v0.1 作为一个实验性规范，为数据与人工智能的交互开辟了一条新颖的路径。它强调作者意图、用户控制和结构化指令，旨在使数据本身变得更加智能和易于访问。随着 LLM 能力的不断发展，此类标准可能会在塑造未来人机协作处理信息的方式中发挥关键作用。

STDM v0.1, as an experimental specification, opens up a novel pathway for data-AI interaction. By emphasizing author intent, user control, and structured instructions, it aims to make data itself smarter and more accessible. As LLM capabilities continue to evolve, standards like this may play a key role in shaping how humans and machines collaboratively work with information in the future.

常见问题（FAQ）

STDM技术如何帮助我更好地理解复杂数据文档？

STDM在数据文档中嵌入结构化指令，指导LLM进行针对性处理和分析，创建交互式探索体验，既能保留作者原意，又能提供智能辅助和数据分析功能。

使用STDM技术时需要注意哪些安全风险？

STDM v0.1是实验性概念，涉及引导LLM行为可能产生意外结果。实施时必须遵循规范中的安全原则，保持适当谨慎并充分考虑安全缓解策略。

如何识别一个文档是否支持STDM功能？

支持STDM的文档会在右上角显示彩色“我是思考数据”丝带作为视觉标记。用户可通过此标识快速识别具有嵌入式智能指令的交互式文档。