## The Basics of GenAIScript ### Introduction GenAIScript is a framework that empowers teams, including non-developers, to create and use AI-enhanced scripts to support their workflows. GenAIScript provides support for authoring and debugging JavaScript scripts that incorporate calls to foundation models and LLMs [^1] in their execution. GenAIScript is a programming framework that allows its users to author AI scripts (which we call a GenAIScript), debug those scripts in a development environment that is an extension of VS Code, and package those scripts in a command-line interface that can be deployed in many contexts. Our VS Code extension supports easy authoring of a GenAIScript by writing natural language in markdown syntax plus a small amount of stylized JavaScript programming. Our framework allows users to leverage multiple LLM models, parameterize the calls to the models, execute and debug scripts, trace the construction of the LLM prompts and provide a full trace of execution from prompt construction to LLM generation to parsing the LLM result. Our framework also supports extracting multiple forms of output from LLM generations, including output in files of different types, outputs intended as edits to existing files and outputs in structured formats, such as JSON. ### Key terms **GenAIScript** A stylized JavaScript program that defines the context for the LLM call, allows arbitrary JavaScript code execution, packages the prompt input for the LLM, calls the LLM, and unpacks that LLM output based on the directions given in the prompt. **GPVM**: A runtime system that given a GenAIScript executes the GenAIScript, which involves integrating the context into a prompt, calling the specified LLM, and extracting content from the LLM result. **VS Code GenAIScript extension** An add-in to VS Code that provides users with easy methods for creating, editing, running and debugging GenAIScript. **Foundation models and LLMs** While GenAIScript currently supports different LLMs, in the future we anticipate that we will incorporate additional foundation models beyond large language models. ## Capabilities ### System behavior GenAIScript is a general-purpose AI-script authoring framework for seamlessly integrating code execution and foundation model/LLM invocations. A GenAIScript is a JavaScript program in a stylized format that allows users to easily specify the LLM context and prompt, invoked a specified model, and parse the resulting output according to user specifications. This functionality allows even users who are not programmers to inspect model results and double check them for correctness. GenAIScript can be written in any IDE but the VS Code GenAIScript add-in makes creating, executing and debugging GenAIScript especially easy. GenAIScript users can implement tools that generate and edit multiple files with a single tool and our integration with VS Code leverages existing functionality in for refactoring to allow users to easily see the results of the tool execution. The add-in supports creating a new GenAIScript, invoking a given GenAIScript, tracing the execution of the GenAIScript in establishing the LLM context and final prompt, and unparsing the LLM output into the user-specified elements. Examples of all of these capabilities can be viewed in the documents in the GenAIScript repository: [microsoft/GenAIScript: Generative AI Scripting (github.com)](https://microsoft.github.io/genaiscript/) The goal of GenAIScript is to empower a broad range of potential users to innovate with building AI-powered scripts and identify new ways to leverage AI for their daily tasks. We expect that professional developers, who are familiar with writing and using scripts to enhance their productivity will be the early adopters of GenAIScript. GenAIScript will give these users benefit because GenAIScript can do many things that existing scripts written in traditional scripting languages like JavaScript and Python cannot do. While developers can leverage other frameworks, such as langchain and Semantic Kernel, that integrate calls to LLMs into languages like Python, these frameworks require more user effort and have less IDE support than GenAIScript. Ultimately, because our goal is to make GenAIScript easy to author, modify, debug and run, we anticipate that they will be useful far beyond professional developers. A major impact of GenAIScript will be to enable non-developers to innovate and build GenAIScripts that enhance their productivity. We illustrate this point with examples below. ### Documentation To help users get started with GenAIScript, we include documentation in our repository that illustrates in code snippets the contents of several different GenAIScripts. The documentation shows both what the example GenAIScript looks like as well as what the effect is from the GenAIScript acting on a particular input. While these examples are intended to explain the technology, they are not intended to be the basis for user-written tools. ### Use cases #### Intended uses GenAIScript can be used in any context where a command line script written in another programming language might be used but the use cases are much more ambitious because the LLM can do much more than ordinary code. Here are some examples: - **Checking for potential inconsistencies in a collection of configuration files or other content.** Using the LLM, a GenAIScript can inspect configuration files and leverage the LLM's understanding of common configuration errors to detect and report them. Before LLMs, professional developers would write tools, such as lint[^2], which are complex programs that detect inconsistencies in the syntax of their code files. With GenAIScript, checking tools can be written for much richer scenarios (such as checking for inappropriate variable names), and by individuals who are not professional developers. - **Automating document translation:** Given documentation in a repository written in one natural language, a GenAIScript can be written to translate that documentation into another language. For a specific example of why GenAIScript is important for this use, consider the task of maintaining the localization of the MakeCode[^3] documentation. MakeCode documentation has nearly 2M files, which are typically markdown with a mix of code snippets. Many documents are partially translated (at the paragraph level). To check the correctness of document translations, there are 3500 registered volunteer translators for 35+ languages. One cannot just apply Bing translate for this use case, as it typically destroys the code snippets. With GenAIScript, we can have a script that goes through every documentation file, pulls the current localized version and assembles a prompt to ask the LLM to fill in the missing translations, while leaving the existing ones alone. Because the LLM model we use has already been trained on MakeCode examples and documentation it is aware of the syntax. - **Creating a short version of a longer white paper by summarizing each chapter.** LLMs are quite effective at summarizing documents. A GenAIScript can be written to take each chapter of a long document and summarize it in a section of a shorter document. - **Translating a monolog to a dialog.** Given a monolog from a video transcript, a GenAIScript can be written to rewrite the monolog into a dialog between two individuals (akin to sports announcers talking to each other) to make the video more interesting and accessible. #### Unintended uses GenAIScript is a general framework for authoring scripts. As a result, an adversary can use GenAIScript to author adversarial scripts that could be used for malicious purposes. All of the adversarial uses of GenAIScript could also be implemented in other LLM language extension frameworks such as Sematic Kernel, autogen, and langchain, so the danger from unintended uses of GenAIScript stems from possibility that it might make it easier to author adversarial scripts. This issue is present in any infrastructure that makes programming easier, including languages such as PowerShell, JavaScript, and Python, as well as IDEs such as VS Code and Visual Studio. While we cannot prevent unintended uses, we will encourage users to consider Responsible AI practices when they build GenAIScripts. We provide more details about issues related to security and trust in [security and trust](https://microsoft.github.io/genaiscript/reference/security-and-trust/). #### Foundation model best practices We strongly encourage GenAIScript users to use foundation models and LLMs that support robust Responsible AI mitigations, such as the Azure Open AI (AOAI) services. Such services continually update the safety and RAI mitigations to track our up-to-date understanding on how to deploy and use foundation models most responsibly. Here are resources to help understand and use best practices when employing foundations models for scripts and applications: - [Blog post on responsible AI features in AOAI that were presented at Ignite 2023](https://techcommunity.microsoft.com/t5/ai-azure-ai-services-blog/announcing-new-ai-safety-amp-responsible-ai-features-in-azure/ba-p/3983686) - [Transparency note for Azure OpenAI Service](https://learn.microsoft.com/en-us/legal/cognitive-services/openai/transparency-note?tabs=text) - [Microsoft Office of Responsible AI (ORA) Best Practices on using AOAI models](https://learn.microsoft.com/en-us/legal/cognitive-services/openai/overview) We recommand to review the [Content Safety](/genaiscript/reference/scripts/content-safety) documentation for more information on how to guard against harmful content and jailbreaking. ## Limitations GenAIScript is an evolving framework that will improve based on input from users. Existing limitations in the framework include integration into only one IDE (VS code), and internal support for OpenAI APIs plus a relatively small number of other LLMs. We intend to allow users to integrate calls to external services (such as RAG) in GenAIScript to provide the LLM with more context. We anticipate adding support for more foundation models as the use cases evolve. We also anticipate that the on-ramp to using GenAIScript will evolve. We have explored supporting invoking the GenAIScript framework as part of a VS Code Copilot Chat experience (hosted in VS Code Insider's Edition). We also understand that some developers would prefer to implement their GenAIScript using Python instead of JavaScript. We anticipate building a Python binding form authoring GenAIScripts in the future. ### Technical limitations, operational factors and ranges GenAIScript does not use any AI model in executing the framework itself. Individuals using GenAIScript to author their own AI scripts will be subject to the technical limitations, operational factors, and ranges of the AI LLM their script uses. ### Best practices for improving system performance GenAIScript encourages users to consult the best practices for authoring effective prompts for the specific LLM they are invoking in their tool. ## Learn more about responsible AI [Microsoft AI principles](https://www.microsoft.com/en-us/ai/responsible-ai) [Microsoft responsible AI resources](https://www.microsoft.com/en-us/ai/responsible-ai-resources) [Microsoft Azure Learning courses on responsible AI](https://docs.microsoft.com/en-us/learn/paths/responsible-ai-business-principles/) ## Learn more about the GenAIScript Read more about GenAIScript at our GitHub site, [microsoft/GenAIScript: GenAI Scripting (github.com)](https://github.com/microsoft/genaiscript/) ## Contact us Give us feedback on this document: zorn@microsoft.com, jhalleux@microsoft.com ## About this document © 2024 Microsoft Corporation. All rights reserved. This document is provided \"as-is\" and for informational purposes only. Information and views expressed in this document, including URL and other Internet Web site references, may change without notice. You bear the risk of using it. Some examples are for illustration only and are fictitious. No real association is intended or inferred. This document is not intended to be, and should not be construed as providing. legal advice. The jurisdiction in which you're operating may have various regulatory or legal requirements that apply to your AI system. Consult a legal specialist if you are uncertain about laws or regulations that might apply to your system, especially if you think those might impact these recommendations. Be aware that not all of these recommendations and resources will be appropriate for every scenario, and conversely, these recommendations and resources may be insufficient for some scenarios. - Published: March 18, 2024 - Last updated: March 18, 2024 --- [^1]: Throughout this document when we refer to LLMs we mean any foundation model that is compatible with our interfaces. [^2]: [Lint (software) - Wikipedia](https://en.wikipedia.org/wiki/Lint_(software)) [^3]: https://makecode.org/