from modules.gpt_modules import gpt_call from langchain.prompts import PromptTemplate def debate_judgement( judgement_who, user_debate_history, bot_debate_history ): if judgement_who == 'User-Bot': judgement_prompt_preset = "\n".join([ "!!Instruction!", "You are now the judge of this debate. Evaluate the debate according to the rules below.", "Rule 1. Decide between the USER and BOT.", "Rule 2. Summarize the debate as a whole and what each debater said.", "Rule 3. For each debater, explain what was persuasive and what made the differnce between winning and losing.", ]) judgement_prompt = "\n".join([ judgement_prompt_preset, "User: " + user_debate_history, "Bot: " + bot_debate_history, "Judgement must be logical with paragraphs.", "Do not show Rule", "Write judgement below.", "Judgement: " ]) elif judgement_who == 'User': judgement_prompt_preset = "\n".join([ "!!Instruction!", "You are now the judge of this debate. Evaluate the debate according to the rules below.", "Rule 1. Summarize the debate as a whole and each said.", "Rule 2. Explain what was persuasive and what made the differnce between winning and losing.", ]) judgement_prompt = "\n".join([ judgement_prompt_preset, "User: " + user_debate_history, "Judgement must be logical with paragraphs.", "Do not show Rule", "Write judgement below.", "Judgement: " ]) elif judgement_who == 'Bot': judgement_prompt_preset = "\n".join([ "!!Instruction!", "You are now the judge of this debate. Evaluate the debate according to the rules below.", "Rule 1. Summarize the debate as a whole and each said.", "Rule 2. Explain what was persuasive and what made the differnce between winning and losing.", ]) judgement_prompt = "\n".join([ judgement_prompt_preset, "Bot: " + bot_debate_history, "Judgement must be logical with paragraphs.", "Do not show Rule", "Write judgement below.", "Judgement: " ]) bot_response = gpt_call(judgement_prompt) return bot_response