{"id":22790,"date":"2023-10-25T10:34:07","date_gmt":"2023-10-25T08:34:07","guid":{"rendered":"https:\/\/golem.ai\/?p=22790"},"modified":"2026-02-17T16:47:08","modified_gmt":"2026-02-17T15:47:08","slug":"avanatge-analytique-vs-llm","status":"publish","type":"post","link":"https:\/\/miralia.ai\/en\/blog\/avanatge-analytique-vs-llm","title":{"rendered":"Avoiding LLM hallucinations through analytical AI, is it possible?"},"content":{"rendered":"

Introduction<\/h2>\n\n\n\n

The success of ChatGPT has generated massive interest in generative AI, especially in the language processing field. LLM models are rapidly improving, and there are increased initiatives to use them within companies.<\/p>\n\n\n\n

For customer support, automating the processing of incoming requests is a popular topic. For example, automatically generating draft responses can have several advantages: it saves time in processing requests and ensures consistency in responses across agents.<\/p>\n\n\n\n

However, generative AI can produce false information, known as hallucinations, which can have critical consequences. Spreading incorrect information can lead to customer dissatisfaction and impact the company's reputation.<\/p>\n\n\n\n

On the other hand, analytical AI, such as the one we have developed at Golem.ai and that is used in our product InboxCare<\/a>, does not suffer from hallucinations. Data processing is predictable and explainable, it synchronises with CRMs, and provides a dynamic response draft.<\/p>\n\n\n\n

This response draft is generated in the form of a \"template\": based on the category we detect in a message, we select the corresponding template.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

InboxCare can, of course, detect the name of the customer and personalize the response accordingly, but it can also use the information detected in a message to enhance its response. For example, detecting an order number allows it to retrieve the corresponding status from a database and thus enrich the draft response.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

The templates, by definition, have limited creativity. It is possible to customize the response dynamically based on some extracted information (such as the client's name or order status), but the tone cannot be changed.<\/p>\n\n\n\n

To illustrate these differences, let's imagine a two-axis graph, with hallucination on one side and response creativity on the other. Let's place the LLM approach and the template approach on this graph:<\/p>\n\n\n\n

<\/div>\n\n\n\n
\"\"<\/figure>\n\n\n\n
<\/div>\n\n\n\n

As mentioned before, the LLM is on the top right corner : It can be very creative, but often suffers from hallucinations. The templates on the other hands are at the bottom, closer to the center : They allow for less creativity, but do not suffer from hallucinations.<\/p>\n\n\n\n

With this study, our goal is to make the most of both worlds: Use the extracted data from our analytical AI to minimize the hallucinations of an LLM, while still benefiting from its creativity in generating responses:<\/strong><\/p>\n\n\n\n

<\/div>\n\n\n\n
\"\"<\/figure>\n\n\n\n

<\/p>\n\n\n\n

Framework of our study<\/h2>\n\n\n\n

To test our theory, we have created a few fake emails addressed to a support client agent of a company called \"ACME,\" specializing in interior furniture, on the following subjects:<\/p>\n\n\n\n