# Experiment to create a student dialog initiatied and led by an AI agent (GPT). # Convert to objects: prompt_transcript, user_dialog_transcript, AI_agent import os import gradio as gr import openai as gpt gpt.api_key = os.environ.get('OPENAI_API_KEY') system_role = "high school english teacher" subject = "chapter 1 of William Golding's Lord of the Flies" follow_up_subject = "chapter 1 of Lord of the Flies" AI_entity = "Learning Assistant" human_entity = "Student" #SCRIPT PROTOTYPE # The dialog script should be loaded from the contents of a file, such as csv format. # The following embedded script is an example. script = [] script_item = [] file_index = 1 with open('dialog script.csv', 'r') as f: for imported_script_line in f: script_item = [] script_item_index = 0 script_item_from_file = imported_script_line.split(",") for line_item in script_item_from_file: script_item.append(line_item.strip()) script_item_index += 1 script.append(script_item) print(script_item) print("\n\nFinal Script: ") print(script) ''' script = [ [0, "prompt", "role_definition", "You are a high school english teacher."], [1, "prompt", "subject_definition", "William Golding's 'Lord of the Flies', Chapter 1"], [2, "prompt", "evaluation_criteria", "The user response accurately reflects the themes and actions depicted in the novel, and expresses them clearly, concisely, and with good english grammar and spelling."], [3, "prompt", "question_specific", "Where are the boys and how did they get there?"], [4, "prompt", "answer_feedback"], [5, "if", "answer_sufficient", 8], [6, "prompt", "question_followup", "Do they feel fearful or excited about their situation?"], [7, "prompt", "answer_feedback"], [8, "prompt", "question_topical", "Ask the user a question about actions the boys take once they understand their situation."], [9, "prompt", "answer_feedback"], [10, "prompt", "final_feedback", "Provide the user with an evaluation of their overall performance against the evaluation criteria."], [11, "prompt", "final_evaluation", "Provide the " + AI_entity + " with an assessment of the user's performance with grading recommendations."], [12, "prompt", "completion", "Thank the user for their participation and end discussion."] ] ''' def submitPrompt(prompt): completion = gpt.ChatCompletion.create( model="gpt-3.5-turbo", messages=prompt ) AI_message = completion["choices"][0]["message"]["content"] + "\n\n" return(AI_message) # print(os.environ.get('OPENAI_API_KEY')) # with gr.Blocks(theme=gr.themes.Soft()) as demo: with gr.Blocks() as demo: def discuss(user_input, chat_transcript, step_pos, internal_dialog_text, dialog): # dialog = dialog or [] # chat_transcript = chat_transcript or "" # step_pos = step_pos or 0 while step_pos < len(script): print(step_pos) if user_input != "": print("User Input: " + user_input + "\n") # print("DIALOG BEFORE APPEND: ") # print(dialog) dialog.append({"role": "user", "content": user_input}) # print("DIALOG: ") # print(dialog) internal_dialog_text += user_input + "\n" chat_transcript += human_entity + ": " + user_input + "\n\n" # print("CHAT: " + chat_transcript) AI_response = submitPrompt(dialog) user_input = "" action = script[step_pos] step_pos += 1 print("NEXT ACTION: " + action[2] + "\n") match action[1]: case 'prompt': match action[2]: case 'role_definition': dialog.append({"role": "system", "content": action[3]}) internal_dialog_text += "Role Definition: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) internal_dialog_text += "Role Response: " + AI_response + "\n" case 'subject_definition': dialog.append({"role": "system", "content": "Subject: " + action[3]}) internal_dialog_text += "Subject Definition: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) internal_dialog_text += "Subject Definition Response: " +AI_response + "\n" case 'get_external_prompt_text': # Open a text file, such as a video transcript, and read it into the prompt. # The prompt should be a single string. # The following code is an example. prompt_transcript = "" # This should be the transcript of the video. prompt_text = "" with open('prompt transcript.txt', 'r') as f: for line in f: prompt_transcript += line prompt_text = prompt_transcript dialog.append({"role": "system", "content": prompt_text}) internal_dialog_text += "Prompt Text: " + prompt_text + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) internal_dialog_text += "Prompt Text Response: " + AI_response + "\n" case 'evaluation_criteria': dialog.append({"role": "system", "content": "Evaluation Criteria: " + action[3]}) internal_dialog_text += "Evaluation Criteria Definition: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) internal_dialog_text += "Evaluation Criteria Definition Response: " + AI_response + "" # with open('AI Dialog Transcript.txt', 'w') as f: # f.write(internal_dialog_text) # f.close() case 'question_specific': dialog.append({"role": "assistant", "content": "Ask user the question: " + action[3]}) internal_dialog_text += "Ask the user the question: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) # dialog.append({"role": "assistant", "content": AI_response}) chat_transcript += (AI_entity + ": " + action[3] + "\n\n") internal_dialog_text += AI_entity + ": " + AI_response + "\n" break case 'answer_feedback': # Add process to check for and AI-generated answer. prompt_text = "Answer yes or no to the whether the user's answer is likely to be AI generated?" internal_dialog_text += ("AI Detected Prompt: " + prompt_text + "\n\n") dialog.append({"role": "assistant", "content": prompt_text}) AI_response = submitPrompt(dialog) # print("AI ANSWER DETECTED? " + AI_response) internal_dialog_text += "AI Detected Response: " + AI_response + "\n" if "yes" in AI_response.lower(): print("AI ANSWER DETECTED\n") prompt_text = "Say to the user: Please answer in your own words." internal_dialog_text += AI_entity + ": " + prompt_text + "\n\n" dialog.append({"role": "assistant", "content": prompt_text}) AI_response = submitPrompt(dialog) print("AI_RESPONSE FROM REQUEST TO ANSWER IN OWN WORDS: " + AI_response) chat_transcript += (AI_entity + ": " + AI_response + "\n\n") internal_dialog_text += (AI_entity + ": " + AI_response + "\n") step_pos -= 1 break else: print("ANSWER_FEEDBACK\n") dialog.append({"role": "assistant", "content": "Provide feedback as if speaking to a high school student on the quality of the most recent response based on the subject and evaluation criteria."}) internal_dialog_text += "Answer Feedback Definition: " + "Provide feedback as if speaking to a high school student on the quality of the most recent response based on the subject and evaluation criteria.\n\n" AI_response = submitPrompt(dialog) # print("AI_response in Feedback: " + AI_response + "\n") dialog.append({"role": "assistant", "content": AI_response}) chat_transcript += (AI_entity + ": " + AI_response + "\n") internal_dialog_text += ("learning Assistant: " + AI_response + "\n") case 'question_followup': print("ASKING FOLLOWUP QUESTION\n") prompt_text = "Ask the user a relevant followup question \ based on the previous question, answer, subject definition, and evaluation criteria." dialog.append({"role": "assistant", "content": prompt_text}) internal_dialog_text += "Followup Question Guidance: " + prompt_text + "\n\n" AI_response = submitPrompt(dialog) print("Followup Question: " + AI_response + "\n") dialog.append({"role": "assistant", "content": AI_response}) chat_transcript += (AI_entity + ": " + AI_response + "\n") internal_dialog_text += ("learning Assistant: " + AI_response + "\n") break case 'question_topical': print("QUESTION_TOPICAL\n") dialog.append({"role": "assistant", "content": action[3]}) internal_dialog_text += "Topical Question Definition: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) # dialog.append({"role": "assistant", "content": AI_response}) print("CHAT BEFORE: " + chat_transcript + "\n\n") chat_transcript += (AI_entity + ": " + AI_response + "\n") internal_dialog_text += ("Learning Assistant: " + AI_response + "\n") break case 'final_feedback': # Do stuff here dialog.append({"role": "assistant", "content": action[3]}) internal_dialog_text += "Final Feedback Instructions: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) print("CHAT BEFORE: " + chat_transcript + "\n\n") chat_transcript += (AI_entity + ": " + AI_response + "\n") internal_dialog_text += ("Learning Assistant: " + AI_response + "\n") case 'final_evaluation': # Do stuff here dialog.append({"role": "assistant", "content": action[3]}) internal_dialog_text += "Final Evaluation Instructions: " + action[3] + "\n\n" AI_response = submitPrompt(dialog) dialog.append({"role": "assistant", "content": AI_response}) print("CHAT BEFORE: " + chat_transcript + "\n\n") # chat_transcript += (AI_entity + ": " + AI_response + "\n") internal_dialog_text += ("Learning Assistant: " + AI_response + "\n") case 'completion': chat_transcript += "Thank you for participating. You may close the browser.\n" with open('AI Dialog Transcript.txt', 'w') as f: f.write(internal_dialog_text) f.close() # break should be unnecessary. case 'if': match action[2]: case 'answer_sufficient': dialog.append({"role": "assistant", "content": "State as yes or no whether the last answer provided by the user demonstrates a strong understanding of the previous question."}) internal_dialog_text += "Answer Evaluation Instructions: " + "State as yes or no whether the last answer provided by the user demonstrates a strong understanding of the previous question.\n\n" AI_response = submitPrompt(dialog) print("IF EVALUATION: " + AI_response) dialog.append({"role": "assistant", "content": AI_response}) # chat_transcript += (AI_entity + ": " + AI_response + "\n\n") internal_dialog_text += ("learning Assistant: " + AI_response + "\n") if "yes" in AI_response.lower(): step_pos = int(action[3]) # print("POSITION UPDATED: " + str(step_pos)) # print("\n" + script[step_pos][2]) # print("Internal Text: " + internal_dialog_text) # print(chat_transcript + "\n") # print(dialog) return "", chat_transcript, chat_transcript, dialog, internal_dialog_text, step_pos # Start processing the script before establishing the session state. This may be # susceptible to concurrency collisions, though, so it would be better to figure out # how to always use Gradio session state variables. initial_chat_recording = "" initial_prompt_history = [] initial_step_num = 0 initial_internal_transcript = "" placeholder_str = "" placeholder_str, initial_chat_recording, initial_chat_recording, initial_prompt_history, initial_internal_transcript, initial_step_num = discuss("", initial_chat_recording, 0, initial_internal_transcript, initial_prompt_history) print("INITIAL TRANSCRIPT: " + initial_internal_transcript + "\n") prompt_history = gr.State(initial_prompt_history) chat_recording = gr.State(initial_chat_recording) internal_transcript = gr.State(initial_internal_transcript) step_num = gr.State(initial_step_num) # gr.Image(shape=(200, 200) with gr.Row(): with gr.Column(scale=1, min_width=200): # gr.Image("https://svhs.co/wp-content/uploads/2018/04/SVHS.png", shape=(200, 50)) logomark = gr.Image("SVHSLogo.png", shape=(200, 50)) logomark.label = "" with gr.Column(scale=4, min_width=400): title = gr.Label("Welcome to the SVHS Learning Assistant.") title.css = "font-size: 20px;" title.label = "" # Empty space with gr.Row(): with gr.Column(scale=1, min_width=300): output = gr.Textbox(label="Dialog", value=chat_recording.value) student_answer = gr.Textbox(label="Answer") student_answer_btn = gr.Button("Answer") with gr.Column(scale=1, min_width=300): running_transcript = gr.Textbox(label="AI Transcript (Hidden From Student)", value=internal_transcript.value) student_answer_btn.click(fn=discuss, inputs=[student_answer, chat_recording, step_num, running_transcript, prompt_history], outputs= [student_answer, output, chat_recording, prompt_history, running_transcript, step_num]) # chat_recording, chat_recording, prompt_history, step_num = discuss("", chat_recording, 0, prompt_history) # examples = gr.Examples(examples=[["", chat_recording, 0, prompt_history]], inputs=[student_answer, chat_recording, step_num, prompt_history]) demo.launch(share=False)