Varex Agent Configuration Guide

Welcome to the Varex Agent Configuration interface, where you can customize your digital assistant to meet specific needs and preferences. This guide outlines the available settings to enhance your interaction with the agent.

General Settings

• Translator Responses: This feature enables the use of Google Translate for converting responses from the Large Language Models (LLMs) into your preferred, non-English language. Although ChatGPT and other LLMs can generate high-quality responses in multiple languages, there might be instances where a model struggles with non-English inputs. By activating this option, you can ensure seamless translation for any LLM, enhancing the multilingual capabilities of your agent.

• Include Relevant Documents: When activated, this setting incorporates citations from source documents that were referenced to generate responses. This feature is particularly useful during testing phases or for internal AI-powered applications. However, it might not always be appropriate for user's contexts, where simplicity and directness are often prioritized.

• Enable Hallucination Guard: The Hallucination Guard is a critical feature designed to enhance the accuracy and reliability of generated responses. By adding an additional check, it prevents the generation of answers not grounded in source documents. While this increases the processing time and the cost of requests, it significantly boosts the trustworthiness of the information provided.

Cache Settings

As most user's requests might be similar, it makes a lot of sense to use caching. This approach searches for similar requests from the past (if documents were not changed) and returns relevant responses. The key parameter here is the similarity threshold, which will estimate how similar requests should be to treat them similarly.

• Use Cache: Caching is an effective strategy for managing similar inquiries, especially considering the repetitive nature of many requests. This method relies on identifying and retrieving responses to past inquiries that closely match the current question, assuming no significant changes in the underlying documents or information bases. The effectiveness of caching hinges on the similarity threshold setting, which determines the degree of resemblance required between requests to be considered a match.

• Cache Threshold: The cache threshold is a critical parameter that affects the performance of the caching mechanism. A threshold value around 0.95 is typically a good starting point, indicating a high degree of similarity between requests. However, it's crucial to conduct thorough testing to adjust this value according to the specific domain and language of operation. A too-low threshold might result in disparate requests being treated as similar, leading to inappropriate responses. Conversely, a too-high threshold could reduce the efficiency of the caching system. It is advisable to begin with a high threshold and gradually adjust it downwards to find an optimal balance that maintains response accuracy without undermining the benefits of caching.

Warning: Use caching carefully. You should test the threshold carefully before going with it to production. Setting a low threshold can treat different requests as similar and lead to errors. At the same time, a very high threshold reduces caching usage. You can start with a high value (close to one) and reduce it gradually.

This configuration guide aims to empower you with the tools to tailor your Varex Agent for enhanced efficiency, accuracy, and user satisfaction.