Our system works as follows: The user sends a prompt via Teams, which is sent to our Teams Client. The logic for the Teams Client is implemented in https://github.com/leukschrauber/ITProjectManagement/blob/main/teams/teamsBot.js. Then, the Teams Client sends the prompt to our HTTP API Backend, which runs on Wildfly 30. The entrance point for the API is implemented in https://github.com/leukschrauber/ITProjectManagement/blob/main/service/src/main/java/at/uni/innsbruck/htibot/rest/HTIBotApiImpl.java. To generate an answer, our Backend calls the Microsoft OpenAI Service to retrieve a meaningful answer. To do so, we do not only send the User prompt, but also a meaningful system description with instructions for the bot. (Example: "You are a polite helpdesk bot, that helps the user with this question"). This task is about the optimization of these system prompts.
Find optimal English, French, Italian and German system messages for:
A Bot that is supposed to send an answer using FAQ data retrieved. (KNOWLEDGEABLE_BOT_MESSAGE)
A Bot that is not supposed to answer, because no FAQ is related to the question. (NO_CLUE_BOT_MESSAGE).
A Bot that is supposed to close the conversation and generate an incident report based on the conversation. (CLOSING_BOT_MESSAGE)
A Bot that is supposed to translate from one language to another. (TRANSLATING_BOT_MESSAGE)
Currently, there are already system messages implemented, but they are not working as expected always. Try to find optimal messages by researching onto how to craft these messages. You may also look into how people craft prompts for GPT Builder.
Our system works as follows: The user sends a prompt via Teams, which is sent to our Teams Client. The logic for the Teams Client is implemented in https://github.com/leukschrauber/ITProjectManagement/blob/main/teams/teamsBot.js. Then, the Teams Client sends the prompt to our HTTP API Backend, which runs on Wildfly 30. The entrance point for the API is implemented in https://github.com/leukschrauber/ITProjectManagement/blob/main/service/src/main/java/at/uni/innsbruck/htibot/rest/HTIBotApiImpl.java. To generate an answer, our Backend calls the Microsoft OpenAI Service to retrieve a meaningful answer. To do so, we do not only send the User prompt, but also a meaningful system description with instructions for the bot. (Example: "You are a polite helpdesk bot, that helps the user with this question"). This task is about the optimization of these system prompts.
Currently, these system prompts are all saved in the package: https://github.com/leukschrauber/ITProjectManagement/tree/main/service/src/main/java/at/uni/innsbruck/htibot/dl/botinstructions
Find optimal English, French, Italian and German system messages for: A Bot that is supposed to send an answer using FAQ data retrieved. (KNOWLEDGEABLE_BOT_MESSAGE) A Bot that is not supposed to answer, because no FAQ is related to the question. (NO_CLUE_BOT_MESSAGE). A Bot that is supposed to close the conversation and generate an incident report based on the conversation. (CLOSING_BOT_MESSAGE) A Bot that is supposed to translate from one language to another. (TRANSLATING_BOT_MESSAGE)
Currently, there are already system messages implemented, but they are not working as expected always. Try to find optimal messages by researching onto how to craft these messages. You may also look into how people craft prompts for GPT Builder.