from typing import List, Optional, Tuple, Dict import os import json from openai import OpenAI from dotenv import load_dotenv import asyncio import regex as re from gradio_modal import Modal import gradio as gr import time # Load environment variables from .env file load_dotenv() from suno import make_song, concat_snippets, update_song_links from gpt_calls import AI_Songwriter from utils.song_utils import messages_to_history History = List[Tuple[str, str]] # a type: pairs of (query, response), where query is user input and response is system output Messages = List[Dict[str, str]] # a type: list of messages with role and content client_key = os.getenv("OPEN_AI_KEY") # print(client_key) oai_client = OpenAI( api_key=client_key, ) def determine_title(section_name, generated_audios): count = sum(1 for audio in generated_audios if audio[2].startswith(section_name)) if count > 0: section_name = f"{section_name} {count + 1}" return section_name def model_chat(genre_input, query: Optional[str], history: Optional[History], messages: Optional[Messages], generated_audios: List[Tuple[str, str, str]], auto=False) -> Tuple[str, History, Messages, str, str, str, str, str, List]: if query is None: query = '' with open('ai_tools.json') as f: ai_tools = json.load(f) songwriterAssistant = AI_Songwriter(client_key=client_key) if auto: messages = messages[:-1] + [{'role': 'user', 'content': query}] #should this be a -1? for non-auto. why does the chatbot history get messed up? else: messages = messages + [{'role': 'user', 'content': query}] messages_filtered = messages response_message = oai_client.chat.completions.create( model="gpt-4o", messages=messages_filtered, tools = ai_tools, tool_choice="required", ) print(response_message, "model chat response") current_response = "" # Step 2: determine if the response from the model includes a tool call. tool_calls = response_message.choices[0].message.tool_calls if tool_calls: messages.append({ "role": response_message.choices[0].message.role, "content": response_message.choices[0].message.content, "tool_calls": tool_calls, "function_call": response_message.choices[0].message.function_call }) if len(tool_calls) > 1: for tool_call in tool_calls: tool_message = { 'role': 'tool', 'tool_call_id': tool_call.id, 'name': tool_call.function.name, 'content': "You called two different functions when you can only call one at a time. Did you mean to call revise_section_lyrics_and_instrumental but instead had two different calls for lyrics and instrumental? Communicate this failure to the user and clarify what they are asking for, then only call one tool next time." } messages.append(tool_message) # Generate a response using GPT-4o and add it as a message model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=messages, ) current_response = model_response_with_function_call.choices[0].message.content role = "assistant" messages.append({'role': role, 'content': current_response}) yield '', messages_to_history(messages), messages, '', '', '', '', None, generated_audios, [] return # If true the model will return the name of the tool / function to call and the argument(s) for tool_call in tool_calls: print(tool_call) tool_call_id = tool_call.id tool_function_name = tool_call.function.name tool_query_args = eval(tool_call.function.arguments) print(tool_function_name, tool_query_args) with open('ai_tools.json') as f: ai_tools = json.load(f) for tool in ai_tools: if tool['function']['name'] == tool_function_name: valid_keys = tool['function']['parameters']['properties'].keys() required_keys = tool['function']['parameters']['required'] break print('query args before', tool_query_args) tool_query_args = {k: v for k, v in tool_query_args.items() if k in valid_keys} print('query args after', tool_query_args) missing_keys = [] for key in required_keys: if key not in tool_query_args: missing_keys.append(key) if len(missing_keys)>0: missing_keys_str = ", ".join(missing_keys) tool_message = { 'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f"Sorry, the keys {missing_keys_str} from the function you called are missing, communicate this to the user and either get what these args should be or figure out which function to call." } new_messages = messages + [tool_message] model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=new_messages, ) # get a new response from the model where it can see the function response current_response = model_response_with_function_call.choices[0].message.content role = "assistant" new_messages = new_messages + [{'role': role, 'content': current_response}] new_history = messages_to_history(new_messages) generated_audios = update_song_links(generated_audios) yield '', new_history, new_messages, '', '', '', '', None, generated_audios, [] # Step 3: Call the function and retrieve results. Append the results to the messages list. if tool_function_name == 'ask_question': question = songwriterAssistant.ask_question(messages) question = question.replace("ask question:", "").replace("ask question ", "").replace("ask question\n", "").replace("ask question", "") ## yield question in tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': question} new_messages = messages + [tool_message] question_message = {'role': 'assistant', 'content': question} new_messages = new_messages + [question_message] new_history = messages_to_history(new_messages) generated_audios = update_song_links(generated_audios) yield '', new_history, new_messages, '', '', '', '', None, generated_audios, [] elif tool_function_name == 'clarify_arguments': tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': 'arguments to clarify: \n' + '\n'.join(tool_query_args['arguments_to_clarify'])} new_messages = messages + [tool_message] model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=new_messages, ) # get a new response from the model where it can see the function response current_response = model_response_with_function_call.choices[0].message.content role = "assistant" new_messages = new_messages + [{'role': role, 'content': current_response}] # + "\n\nWould you like to get an audio snippet? Or continue writing?"}] # new_messages = [msg for msg in new_messages if msg['content'] is not None and msg['role'] in ['user', 'assistant']] new_history = messages_to_history(new_messages) generated_audios = update_song_links(generated_audios) yield '', new_history, new_messages, '', '', '', '', None, generated_audios, [] elif tool_function_name == 'write_section': snippet_instrumental_tags = tool_query_args.pop('snippet_instrumental_tags', None) snippet_clip_to_continue_from = tool_query_args.pop('snippet_clip_to_continue_from', None) suggested_lyrics = songwriterAssistant.write_section(**tool_query_args) suggested_lyrics = suggested_lyrics.strip('`*-\n') ## yield suggested lyrics in tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': suggested_lyrics} # audio_message = {'role': 'assistant', 'content': "Here's what I've come up with:\n" + suggested_lyrics + "\n\nGenerating audio snippet..."} new_messages = messages + [tool_message] #, audio_message model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=new_messages, ) # get a new response from the model where it can see the function response current_response = model_response_with_function_call.choices[0].message.content role = "assistant" new_messages = new_messages + [{'role': role, 'content': current_response}] # + "\n\nWould you like to get an audio snippet? Or continue writing?"}] # new_messages = [msg for msg in new_messages if msg['content'] is not None and msg['role'] in ['user', 'assistant']] history = messages_to_history(new_messages) # current_section, current_lyrics, curr_tags, clip_to_continue, curr_audio buttons = ["revise lyrics", "generate audio snippet", "continue to next section"] generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = snippet_clip_to_continue_from, choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', history, new_messages, tool_query_args['section_name'], suggested_lyrics.split(':')[-1], snippet_instrumental_tags, clips_to_continue, None, generated_audios, buttons ### DO SOMETHING TO UPDATE CURRENT GENERATION for write_section elif tool_function_name == 'revise_section_lyrics': revised_lyrics = songwriterAssistant.revise_section_lyrics(**tool_query_args) # if isinstance(revised_lyrics, list): # revised_lyrics = '\n'.join(revised_lyrics) if isinstance(revised_lyrics, str) and revised_lyrics.startswith("[") and revised_lyrics.endswith("]"): try: revised_lyrics = eval(revised_lyrics) if isinstance(revised_lyrics, list): revised_lyrics = '\n'.join(revised_lyrics) except: pass # ## yield revised lyrics in tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': revised_lyrics} # audio_message = {'role': 'assistant', 'content': "Here's my revised lyrics:\n" + revised_lyrics + "\n\nGenerating audio snippet..."} new_messages = messages + [tool_message] #, audio_message] model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=new_messages, ) # get a new response from the model where it can see the function response current_response = model_response_with_function_call.choices[0].message.content buttons = ["revise lyrics again", "generate audio snippet with new lyrics", "continue to next section"] role = "assistant" new_messages = new_messages + [{'role': role, 'content': current_response + "\n\nWould you like to get an audio snippet? Or continue writing?"}] # new_messages = [msg for msg in new_messages if msg['content'] is not None and msg['role'] in ['user', 'assistant']] history = messages_to_history(new_messages) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', history, new_messages, tool_query_args['section_name'], revised_lyrics, '', clips_to_continue, None, generated_audios, buttons elif tool_function_name == 'revise_instrumental_tags': #detangle tool_query_args dict #snippet_lyrics = tool_query_args['snippet_lyrics'] + "\n[End]" snippet_instrumental_tags = tool_query_args.get('current_instrumental_tags', None) user_instrumental_feedback = tool_query_args.get('user_instrumental_feedback', None) if snippet_instrumental_tags is None or user_instrumental_feedback is None: tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': 'Arguments are missing. Please clarify your feedback on the instrumental. Note that you cannot revise the genre if you haven\'t generated a snippet.'} audio_message = {'role': 'assistant', 'content': 'It seems like some information is missing. Could you please provide your feedback on the instrumental? Note that you cannot revise the genre if you haven\'t generated a snippet.'} new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) yield '', new_history, new_messages, '', '', '', None, None, generated_audios, [] return # if 'snippet_clip_to_continue_from' not in tool_query_args: # tool_query_args['snippet_clip_to_continue_from'] = None # snippet_clip_to_continue_from = tool_query_args['snippet_clip_to_continue_from'] new_instrumental_tags = songwriterAssistant.revise_instrumental_tags(snippet_instrumental_tags, user_instrumental_feedback) if isinstance(tool_query_args['sections_written'], str): current_lyrics = tool_query_args['sections_written'] elif isinstance(tool_query_args['sections_written'], list): current_lyrics = "\n".join(tool_query_args['sections_written']) else: current_lyrics = "" import re sections_list = re.findall(r'\[.*?\]', current_lyrics) #current_lyrics = "\n".join(tool_query_args['sections_written']) song_link = make_song(current_lyrics, new_instrumental_tags) ## filter out suno link from tool query arg while "https://audiopipe.suno.ai/?item_id=" not in song_link: print("BUGGED OUT, trying again...") time.sleep(5) song_link = make_song(current_lyrics, new_instrumental_tags) clip_id = song_link.split("https://audiopipe.suno.ai/?item_id=")[1] tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'new instrumental tags: {new_instrumental_tags}, clip id: {clip_id}'} audio_message = {'role': 'assistant', 'content': f'Sure! I\'ve revised the instrumental tags: {new_instrumental_tags}\nCurrent lyrics: {current_lyrics}\n\n '} audio_message['content'] += f'\n\nWhat do you think?' new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) # current_section, current_lyrics, curr_tags, clip_to_continue, curr_audio if len(sections_list) > 0: section_name = f"Up to {sections_list[-1]}" else: section_name = "Up to latest section" section_name = determine_title(section_name, generated_audios) generated_audios.append((song_link, current_lyrics, new_instrumental_tags, section_name, "streaming")) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) buttons = ["return to original instrumental", "re-revise genre", "revise lyrics", "merge snippets", "continue to next section"] yield '', new_history, new_messages, ', '.join(sections_list), current_lyrics, new_instrumental_tags, clips_to_continue, f'', generated_audios, buttons elif tool_function_name == 'revise_section_lyrics_and_instrumental': snippet_instrumental_tags = tool_query_args.pop('current_instrumental_tags', None) user_instrumental_feedback = tool_query_args.pop('user_instrumental_feedback', None) snippet_clip_to_continue_from = tool_query_args.pop('snippet_clip_to_continue_from', None) # Revise section lyrics revised_lyrics = songwriterAssistant.revise_section_lyrics(**tool_query_args) # Revise instrumental tags new_instrumental_tags = songwriterAssistant.revise_instrumental_tags(snippet_instrumental_tags, user_instrumental_feedback) song_link = make_song(revised_lyrics, new_instrumental_tags, snippet_clip_to_continue_from) while "https://audiopipe.suno.ai/?item_id=" not in song_link: print("BUGGED OUT, trying again...") time.sleep(5) song_link = make_song(revised_lyrics, new_instrumental_tags, snippet_clip_to_continue_from) clip_id = song_link.split("https://audiopipe.suno.ai/?item_id=")[1] tool_message_instrumental = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'revised lyrics: {revised_lyrics}\nrevised instrumental tags: {new_instrumental_tags}, clip id: {clip_id}'} audio_message = {'role': 'assistant', 'content': f'Sure! I\'ve revised the lyrics and instrumental tags: {revised_lyrics}\nRevised lyrics: {revised_lyrics}\n\n '} audio_message['content'] += f'\n\nWhat do you think?' new_messages = messages + [tool_message_instrumental, audio_message] new_history = messages_to_history(new_messages) generated_audios.append((song_link, revised_lyrics, new_instrumental_tags, tool_query_args["section_name"], "streaming")) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = snippet_clip_to_continue_from, choices=[x[3] for x in generated_audios]+[""], interactive=True) buttons = ["return to original instrumental", "re-revise genre", "revise lyrics", "merge snippets", "continue to next section"] yield '', new_history, new_messages, tool_query_args["section_name"], revised_lyrics, new_instrumental_tags, clips_to_continue, f'', generated_audios, buttons elif tool_function_name == 'merge_all_snippets': updated_clip_url, updated_lyrics, updated_tags, clips_list = concat_snippets(tool_query_args['last_snippet_id']) if "still streaming" in updated_clip_url: tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'still streaming, try again later'} audio_message = {'role': 'assistant', 'content': f'Unfortunately the generated clip audio is still being streamed, so you can merge later when it is fully generated.'} new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', new_history, new_messages, "", "", "", clips_to_continue, None, generated_audios, [] else: if "https://audiopipe.suno.ai/?item_id=" in updated_clip_url: updated_clip_id = updated_clip_url.split("https://audiopipe.suno.ai/?item_id=")[1] elif "https://cdn1.suno.ai/" in updated_clip_url: updated_clip_id = updated_clip_url.split("https://cdn1.suno.ai/")[1].split(".mp3")[0] else: updated_clip_id = "unknown" #pass this info in new tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'updated clip id: {updated_clip_id}\nupdated lyrics: {updated_lyrics}\nupdated clips path: {clips_list}'} audio_message = {'role': 'assistant', 'content': f'Sure! All the clips are now merged.

updated lyrics: {updated_lyrics}

updated clips path: {clips_list}

'} sections_list = [line for line in current_lyrics.split('\n') if line.startswith('[') and line.endswith(']')] new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) if len(sections_list) > 0: section_name = "Merge up to " + sections_list[-1] else: section_name = "Merge up to latest section" section_name = determine_title(section_name, generated_audios) # current_section, current_lyrics, curr_tags, clip_to_continue, curr_audio generated_audios.append((updated_clip_url, updated_lyrics, updated_tags, section_name, "streaming")) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', new_history, new_messages, section_name, updated_lyrics, updated_tags, clips_to_continue, f'', generated_audios, [] elif tool_function_name == 'finish_full_song': ## args are sections_to_be_written, relevant_ideas, last_snippet_id, sni ## STEP 0: POP out instrumental args snippet_instrumental_tags = tool_query_args.pop('snippet_instrumental_tags', None) snippet_clip_to_continue_from = tool_query_args.pop('snippet_clip_to_continue_from', None) if isinstance(tool_query_args['sections_written'], str): current_lyrics = tool_query_args['sections_written'] elif isinstance(tool_query_args['sections_written'], list): current_lyrics = "\n".join(tool_query_args['sections_written']) else: current_lyrics = "" ## STEP 1: WRITE ALL LYRICS using songwriterAssistant remaining_lyrics = songwriterAssistant.write_all_lyrics(**tool_query_args) full_lyrics = current_lyrics + remaining_lyrics + "\n[End]" # current_section, current_lyrics, curr_tags, clip_to_continue, curr_audio yield '', history, messages, "Full Song", full_lyrics, snippet_instrumental_tags, snippet_clip_to_continue_from, None, generated_audios, [] ## STEP 2: MAKE SONG FOR REMAINING LYRICS song_link = make_song(remaining_lyrics, snippet_instrumental_tags, snippet_clip_to_continue_from) #tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'{full_lyrics}'} audio_message = {'role': 'assistant', 'content': f'New snippet: \n '} new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) generated_audios.append((song_link, remaining_lyrics, snippet_instrumental_tags, "Rest of Song", "streaming")) yield '', new_history, new_messages, "Rest of Song", full_lyrics, snippet_instrumental_tags, snippet_clip_to_continue_from, song_link, generated_audios, [] ## STEP 3: MERGE FULL SONG if snippet_clip_to_continue_from not in [None, ""]: updated_clip_url = "still streaming" while "still streaming" in updated_clip_url: if "https://audiopipe.suno.ai/?item_id=" in song_link: clip_id = song_link.split("https://audiopipe.suno.ai/?item_id=")[1] else: clip_id = updated_clip_url.split("https://cdn1.suno.ai/")[1].split(".mp3")[0] updated_clip_url, updated_lyrics, updated_tags, clips_list = concat_snippets(clip_id) else: updated_clip_url, updated_lyrics, clips_list = song_link, remaining_lyrics, [] ## YIELD UPDATED CLIP URL, LYRICS, AND CLIPS LIST if "https://audiopipe.suno.ai/?item_id=" in updated_clip_url: updated_clip_id = updated_clip_url.split("https://audiopipe.suno.ai/?item_id=")[1] elif "https://cdn1.suno.ai/" in updated_clip_url: updated_clip_id = updated_clip_url.split("https://cdn1.suno.ai/")[1].split(".mp3")[0] else: updated_clip_id = "unknown" #tool and assistant message tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'updated clip id: {updated_clip_id}\nupdated lyrics: {updated_lyrics}\nupdated clips path: {clips_list}'} audio_message = {'role': 'assistant', 'content': f'All done! Thank you for participating :) \nFinal Lyrics: {full_lyrics} \nFinal song: '} new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) generated_audios.append((updated_clip_url, updated_lyrics, updated_tags, "Full Song", "streaming")) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', new_history, new_messages, "Full Song", full_lyrics, snippet_instrumental_tags, clips_to_continue, f'', generated_audios, [] elif tool_function_name == 'get_audio_snippet': #detangle tool_query_args dict snippet_lyrics = tool_query_args['snippet_lyrics'] + "\n[End]" snippet_instrumental_tags = tool_query_args['snippet_instrumental_tags'] snippet_clip_to_continue_from = tool_query_args.get('snippet_clip_to_continue_from', None) song_link = make_song(snippet_lyrics, snippet_instrumental_tags, snippet_clip_to_continue_from) if "still streaming" in song_link: tool_message = { 'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': 'The snippet to extend is still streaming. Please try generating this audio snippet in a little bit.' } new_messages = messages + [tool_message] model_response_with_function_call = oai_client.chat.completions.create( model="gpt-4o", messages=new_messages, ) # get a new response from the model where it can see the function response current_response = model_response_with_function_call.choices[0].message.content role = "assistant" new_messages = new_messages + [{'role': role, 'content': current_response}] new_history = messages_to_history(new_messages) generated_audios = update_song_links(generated_audios) clips_to_continue = gr.Dropdown(label='Clip to continue', value = "", choices=[x[3] for x in generated_audios]+[""], interactive=True) buttons = [] yield '', new_history, new_messages, snippet_lyrics.split("\n")[0], snippet_lyrics, snippet_instrumental_tags, clips_to_continue, None, generated_audios, buttons return print("MAKE SONG IS DONE") ## filter out suno link from tool query arg clip_id = song_link.split("https://audiopipe.suno.ai/?item_id=")[1] tool_message = {'role': 'tool', 'tool_call_id': tool_call_id, 'name': tool_function_name, 'content': f'snippet lyrics: {snippet_lyrics}\ninstrumental tags: {tool_query_args["snippet_instrumental_tags"]}, clip id: {clip_id}'} audio_message_content = "Here's what I've come up with:\n" + snippet_lyrics + '\n\n' + f'

instrumental tags: {tool_query_args["snippet_instrumental_tags"]}

' audio_message_content += f'

continued from clip: {snippet_clip_to_continue_from}

' audio_message_content += "What do you think?" audio_message = {'role': 'assistant', 'content': audio_message_content} section_name = snippet_lyrics.split("\n")[0].strip('[]* ') section_name = determine_title(section_name, generated_audios) #audio_message = {'role': 'assistant', 'content': gr.Audio(value=song_link, label=section_name, interactive=False, show_label=False, waveform_options={"show_controls": False})} new_messages = messages + [tool_message, audio_message] new_history = messages_to_history(new_messages) print("AUDIO MESSAGE DONE") generated_audios.append((song_link, snippet_lyrics, snippet_instrumental_tags, section_name, "streaming")) generated_audios = update_song_links(generated_audios) buttons = ["revise lyrics", "revise genre", "merge snippets", "continue to next section"] clips_to_continue = gr.Dropdown(label='Clip to continue', value = snippet_clip_to_continue_from, choices=[x[3] for x in generated_audios]+[""], interactive=True) yield '', new_history, new_messages, snippet_lyrics.split("\n")[0], snippet_lyrics, snippet_instrumental_tags, clips_to_continue, f'', generated_audios, buttons else: print(f"Error: function {tool_function_name} does not exist") else: # Model did not identify a function to call, result can be returned to the user current_response = response_message.choices[0].message.content role = "assistant" new_messages = messages + [{'role': role, 'content': current_response}] # new_messages = [msg for msg in new_messages if msg['content'] is not None and msg['role'] in ['user', 'assistant']] history = messages_to_history(new_messages) yield '', history, new_messages, '[...]'