Real-time Assurance of Financial Guidance AI using Explicit Guardrails

随着对ChatGPT和其他大型语言模型(LLM)的巨大兴趣，商界中的许多人都清楚地看到，LLM非常吸引人，并提供了以低得多的成本为客户提供智能解决方案的可能性。但也有充分的报道称，这些模型会产生错误和幻觉。如果不解决这一问题，这些模型在许多实际的企业应用中使用是不安全的。如果可以防止幻觉，并可以保证LLMS响应的准确性，那么它们的灵活性、引人入胜的用户界面和低成本将释放许多可能性。在我们的业务中，我们特别专注于为消费者解锁数字金融指导，特别是大众市场和低收入领域，而这些市场和低收入领域目前还没有得到金融服务公司的很好的服务。在过去三个月里，开源LLM社区出现的新工具和方法开始能够增强和限制LLM的输出，如ChatGPT和GPT4\。社区主要使用它们来混合不同的数据源和内容。我们看到了一种修改这些方法的方法，以在LLM周围为特定的财务指导任务创建强大的护栏。这种保证方法承诺显著提高质量和安全，我们相信这是金融服务公司使用人工智能支持客户和他们的财务决策的推动因素。这个项目是关于设计这个保证解决方案的。然后，项目将把这个工程设计的保证原型应用于一个真实的财务指导养老金使用案例。除了这个最初的使用案例，这种保证方法还有可能应用于其他指导领域，其中护栏可以用自然语言或逻辑和代码清楚地表达出来，例如法律、税收、补贴和监管使用案例。

With the huge interest in ChatGPT, and other large language models (LLMs), it's clear to many in the business community that LLMs are very engaging and offer the possibility of providing smart solutions for customers at much lower cost. But it is also well reported that these models generate errors and hallucinations. These models are unsafe to use in many real enterprise applications without addressing this.If the hallucinations could be prevented and guarantees could be given as to the accuracy of responses by LLMs, then their flexibility, engaging UI, and low cost would unlock many possibilities. In our business we are particularly focused on unlocking digital financial guidance for consumers, particularly the mass market and low income segments which are not well served by financial services companies today.New tooling and methods emerging in the last three months from the open-source LLM community are starting to be able to augment and constrain the outputs of LLMs such as ChatGPT and GPT4\. The community is using them primarily to blend in different data sources and content. We see a way to modify these methods to create strong guardrails around LLMs for specific financial guidance tasks. This assurance approach promises to significantly improve quality and safety and we believe this is an enabler for financial services companies to using AI to support their customers and their financial decisions.This project is about engineering this assurance solution. The project will then apply this engineered assurance prototype to a real financial guidance pension use case.Beyond this initial use case, this assurance methodology has potential to be applied in other guidance domains where guardrails can be articulated clearly in natural language or logic and code, for example legal, taxes, subsidies and regulation use cases.