# Welcome to Team Tonic's MultiMed from gradio_client import Client import numpy as np import base64 import gradio as gr import tempfile import requests import json import dotenv from scipy.io.wavfile import write import soundfile as sf from openai import OpenAI import time import PIL from PIL import Image import io import hashlib import datetime from utils import build_logger from transformers import AutoTokenizer, MistralForCausalLM import torch import random from textwrap import wrap import transformers from transformers import AutoConfig, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForCausalLM, MistralForCausalLM from peft import PeftModel, PeftConfig import torch import os import uuid # Global variables to hold component references components = {} dotenv.load_dotenv() seamless_client = Client("facebook/seamless_m4t") HuggingFace_Token = os.getenv("HuggingFace_Token") hf_token = os.getenv("HuggingFace_Token") base_model_id = os.getenv('BASE_MODEL_ID', 'default_base_model_id') model_directory = os.getenv('MODEL_DIRECTORY', 'default_model_directory') device = "cuda" if torch.cuda.is_available() else "cpu" def check_hallucination(assertion,citation): API_URL = "https://api-inference.huggingface.co/models/vectara/hallucination_evaluation_model" headers = {"Authorization": f"Bearer {HuggingFace_Token}"} payload = {"inputs" : f"{assertion} [SEP] {citation}"} response = requests.post(API_URL, headers=headers, json=payload,timeout=120) output = response.json() output = output[0][0]["score"] return f"**hallucination score:** {output}" # Define the API parameters VAPI_URL = "https://api-inference.huggingface.co/models/vectara/hallucination_evaluation_model" headers = {"Authorization": f"Bearer {HuggingFace_Token}"} # Function to query the API def query(payload): response = requests.post(VAPI_URL, headers=headers, json=payload) return response.json() # Function to evaluate hallucination def evaluate_hallucination(input1, input2): # Combine the inputs combined_input = f"{input1}. {input2}" # Make the API call output = query({"inputs": combined_input}) # Extract the score from the output score = output[0][0]['score'] # Generate a label based on the score if score < 0.5: label = f"🔴 High risk. Score: {score:.2f}" else: label = f"🟢 Low risk. Score: {score:.2f}" return label def save_audio(audio_input, output_dir="saved_audio"): if not os.path.exists(output_dir): os.makedirs(output_dir) # Extract sample rate and audio data sample_rate, audio_data = audio_input # Generate a unique file name file_name = f"audio_{int(time.time())}.wav" file_path = os.path.join(output_dir, file_name) # Save the audio file sf.write(file_path, audio_data, sample_rate) return file_path def process_speech(input_language, audio_input): """ processing sound using seamless_m4t """ if audio_input is None : return "no audio or audio did not save yet \nplease try again ! " print(f"audio : {audio_input}") print(f"audio type : {type(audio_input)}") out = seamless_client.predict( "S2TT", "file", None, audio_input, #audio_name "", input_language,# source language "English",# target language api_name="/run", ) out = out[1] # get the text try : return f"{out}" except Exception as e : return f"{e}" def convert_text_to_speech(input_text, source_language, target_language): """ Convert text to speech in the specified language and return the audio file path and translated text. """ client = Client("https://facebook-seamless-m4t.hf.space/--replicas/8cllp/") try: result = client.predict( "T2ST (Text to Speech translation)", # Task "text", # Input type input_text, # Input text "", # Empty string for audio name source_language, # Source language target_language, # Target language api_name="/run" # API name ) # Check if result contains files and select the first one if isinstance(result, list) and len(result) > 1: # Select the first audio file from the result original_audio_file = result[1] # Assuming the first element is the audio file # Iterate over the result to find the last text item translated_text = "" for item in result: if isinstance(item, str): translated_text = item if original_audio_file: # Generate a new file name with a random UUID new_file_name = f"audio_output_{uuid.uuid4()}.wav" new_file_path = os.path.join(os.path.dirname(original_audio_file), new_file_name) # Rename the file os.rename(original_audio_file, new_file_path) return new_file_path, translated_text else: return "No audio file generated.", translated_text else: return "Unexpected result format or insufficient data received.", "" except Exception as e: # Return a concise error message return f"Error in text-to-speech conversion: {str(e)}", "" def process_image(image_input): # Initialize the Gradio client with the URL of the Gradio server client = Client("https://adept-fuyu-8b-demo.hf.space/--replicas/pqjvl/") # Assuming image_input is a URL path to the image image_path = image_input # Call the predict method of the client result = client.predict( image_path, # URL of the image True, # Additional parameter for the server (e.g., enable detailed captioning) fn_index=2 ) return result def query_vectara(text): user_message = text # Read authentication parameters from the .env file CUSTOMER_ID = os.getenv('CUSTOMER_ID') CORPUS_ID = os.getenv('CORPUS_ID') API_KEY = os.getenv('API_KEY') # Define the headers api_key_header = { "customer-id": CUSTOMER_ID, "x-api-key": API_KEY } # Define the request body in the structure provided in the example request_body = { "query": [ { "query": user_message, "queryContext": "", "start": 1, "numResults": 25, "contextConfig": { "charsBefore": 0, "charsAfter": 0, "sentencesBefore": 2, "sentencesAfter": 2, "startTag": "%START_SNIPPET%", "endTag": "%END_SNIPPET%", }, "rerankingConfig": { "rerankerId": 272725718, "mmrConfig": { "diversityBias": 0.35 } }, "corpusKey": [ { "customerId": CUSTOMER_ID, "corpusId": CORPUS_ID, "semantics": 0, "metadataFilter": "", "lexicalInterpolationConfig": { "lambda": 0 }, "dim": [] } ], "summary": [ { "maxSummarizedResults": 5, "responseLang": "auto", "summarizerPromptName": "vectara-summary-ext-v1.2.0" } ] } ] } # Make the API request using Gradio response = requests.post( "https://api.vectara.io/v1/query", json=request_body, # Use json to automatically serialize the request body verify=True, headers=api_key_header ) if response.status_code == 200: query_data = response.json() if query_data: sources_info = [] # Extract the summary. summary = query_data['responseSet'][0]['summary'][0]['text'] # Iterate over all response sets for response_set in query_data.get('responseSet', []): # Extract sources # Limit to top 5 sources. for source in response_set.get('response', [])[:5]: source_metadata = source.get('metadata', []) source_info = {} for metadata in source_metadata: metadata_name = metadata.get('name', '') metadata_value = metadata.get('value', '') if metadata_name == 'title': source_info['title'] = metadata_value elif metadata_name == 'author': source_info['author'] = metadata_value elif metadata_name == 'pageNumber': source_info['page number'] = metadata_value if source_info: sources_info.append(source_info) result = {"summary": summary, "sources": sources_info} return f"{json.dumps(result, indent=2)}" else: return "No data found in the response." else: return f"Error: {response.status_code}" # Functions to Wrap the Prompt Correctly def wrap_text(text, width=90): lines = text.split('\n') wrapped_lines = [textwrap.fill(line, width=width) for line in lines] wrapped_text = '\n'.join(wrapped_lines) return wrapped_text def multimodal_prompt(user_input, system_prompt="You are an expert medical analyst:"): # Combine user input and system prompt formatted_input = f"{user_input}{system_prompt}" # Encode the input text encodeds = tokenizer(formatted_input, return_tensors="pt", add_special_tokens=False) model_inputs = encodeds.to(device) # Generate a response using the model output = model.generate( **model_inputs, max_length=max_length, use_cache=True, early_stopping=True, bos_token_id=model.config.bos_token_id, eos_token_id=model.config.eos_token_id, pad_token_id=model.config.eos_token_id, temperature=0.1, do_sample=True ) # Decode the response response_text = tokenizer.decode(output[0], skip_special_tokens=True) return response_text # Instantiate the Tokenizer tokenizer = AutoTokenizer.from_pretrained("stabilityai/stablelm-3b-4e1t", token=hf_token, trust_remote_code=True, padding_side="left") # tokenizer = AutoTokenizer.from_pretrained("Tonic/stablemed", trust_remote_code=True, padding_side="left") tokenizer.pad_token = tokenizer.eos_token tokenizer.padding_side = 'left' # Load the PEFT model peft_config = PeftConfig.from_pretrained("Tonic/stablemed", token=hf_token) peft_model = AutoModelForCausalLM.from_pretrained("stabilityai/stablelm-3b-4e1t", token=hf_token, trust_remote_code=True) peft_model = PeftModel.from_pretrained(peft_model, "Tonic/stablemed", token=hf_token) class ChatBot: def __init__(self): self.history = [] def predict(self, user_input, system_prompt="You are an expert medical analyst:"): formatted_input = f"{system_prompt}{user_input}" user_input_ids = tokenizer.encode(formatted_input, return_tensors="pt") response = peft_model.generate(input_ids=user_input_ids, max_length=512, pad_token_id=tokenizer.eos_token_id) response_text = tokenizer.decode(response[0], skip_special_tokens=True) return response_text bot = ChatBot() def process_summary_with_stablemed(summary): system_prompt = "You are a medical instructor . Assess and describe the proper options to your students in minute detail. Propose a course of action for them to base their recommendations on based on your description." response_text = bot.predict(summary, system_prompt) return response_text # Main function to handle the Gradio interface logic def process_and_query(input_language=None, audio_input=None, image_input=None, text_input=None): try: combined_text = "" image_description = "" markdown_output = "" image_text = "" translated_text = "" audio_file_path = "" # Debugging print statement print(f"Image Input Type: {type(image_input)}, Audio Input Type: {type(audio_input)}") # Process image input if image_input is not None: # Convert image_input to a file path image_file_path = save_image(image_input) image_text = process_image(image_file_path) combined_text += "\n\n**Image Input:**\n" + image_text # Process audio input elif audio_input is not None: audio_file_path = save_audio(audio_input) audio_text = process_speech(input_language, audio_file_path) combined_text += "\n\n**Audio Input:**\n" + audio_text # Process text input elif text_input is not None and text_input.strip(): combined_text += "The user asks the following to his health adviser: " + text_input # Check if combined text is empty else: return "Error: Please provide some input (text, audio, or image)." # Append the original image description in Markdown if image_text: markdown_output += "\n### Original Image Description\n" markdown_output += image_text + "\n" # Use the text to query Vectara vectara_response_json = query_vectara(combined_text) # Parse the Vectara response vectara_response = json.loads(vectara_response_json) summary = vectara_response.get('summary', 'No summary available') sources_info = vectara_response.get('sources', []) # Format Vectara response in Markdown markdown_output = "### Vectara Response Summary\n" markdown_output += f"* **Summary**: {summary}\n" markdown_output += "### Sources Information\n" for source in sources_info: markdown_output += f"* {source}\n" # Process the summary with Stablemed final_response = process_summary_with_stablemed(summary) # Convert translated text to speech and get both audio file and text target_language = "English" # Set the target language for the speech audio_file_path, translated_text = convert_text_to_speech(final_response, target_language, input_language) # Evaluate hallucination hallucination_label = evaluate_hallucination(final_response, summary) # Add final response and hallucination label to Markdown output markdown_output += "\n### Processed Summary with StableMed\n" markdown_output += final_response + "\n" markdown_output += "\n### Hallucination Evaluation\n" markdown_output += f"* **Label**: {hallucination_label}\n" markdown_output += "\n### Translated Text\n" markdown_output += translated_text + "\n" return markdown_output, audio_output except Exception as e: return f"Error occurred during processing: {e}. No hallucination evaluation.", None welcome_message = """ # 👋🏻Welcome to ⚕🗣️😷MultiMed - Access Chat ⚕🗣️😷 🗣️📝 This is an educational and accessible conversational tool. ### How To Use ⚕🗣️😷MultiMed⚕: 🗣️📝Interact with ⚕🗣️😷MultiMed⚕ in any language using image, audio or text! 📚🌟💼 that uses [Tonic/stablemed](https://huggingface.co/Tonic/stablemed) and [adept/fuyu-8B](https://huggingface.co/adept/fuyu-8b) with [Vectara](https://huggingface.co/vectara) embeddings + retrieval. do [get in touch](https://discord.gg/GWpVpekp). You can also use 😷MultiMed⚕️ on your own data & in your own way by cloning this space. 🧬🔬🔍 Simply click here: Duplicate Space ### Join us : 🌟TeamTonic🌟 is always making cool demos! Join our active builder's🛠️community on 👻Discord: [Discord](https://discord.gg/GWpVpekp) On 🤗Huggingface: [TeamTonic](https://huggingface.co/TeamTonic) & [MultiTransformer](https://huggingface.co/MultiTransformer) On 🌐Github: [Polytonic](https://github.com/tonic-ai) & contribute to 🌟 [PolyGPT](https://github.com/tonic-ai/polygpt-alpha)" """ languages = [ "Afrikaans", "Amharic", "Modern Standard Arabic", "Moroccan Arabic", "Egyptian Arabic", "Assamese", "Asturian", "North Azerbaijani", "Belarusian", "Bengali", "Bosnian", "Bulgarian", "Catalan", "Cebuano", "Czech", "Central Kurdish", "Mandarin Chinese", "Welsh", "Danish", "German", "Greek", "English", "Estonian", "Basque", "Finnish", "French", "West Central Oromo", "Irish", "Galician", "Gujarati", "Hebrew", "Hindi", "Croatian", "Hungarian", "Armenian", "Igbo", "Indonesian", "Icelandic", "Italian", "Javanese", "Japanese", "Kamba", "Kannada", "Georgian", "Kazakh", "Kabuverdianu", "Halh Mongolian", "Khmer", "Kyrgyz", "Korean", "Lao", "Lithuanian", "Luxembourgish", "Ganda", "Luo", "Standard Latvian", "Maithili", "Malayalam", "Marathi", "Macedonian", "Maltese", "Meitei", "Burmese", "Dutch", "Norwegian Nynorsk", "Norwegian Bokmål", "Nepali", "Nyanja", "Occitan", "Odia", "Punjabi", "Southern Pashto", "Western Persian", "Polish", "Portuguese", "Romanian", "Russian", "Slovak", "Slovenian", "Shona", "Sindhi", "Somali", "Spanish", "Serbian", "Swedish", "Swahili", "Tamil", "Telugu", "Tajik", "Tagalog", "Thai", "Turkish", "Ukrainian", "Urdu", "Northern Uzbek", "Vietnamese", "Xhosa", "Yoruba", "Cantonese", "Colloquial Malay", "Standard Malay", "Zulu" ] def clear(): # Return default values for each component return "English", None, None, "", None def create_interface(): with gr.Blocks(theme='ParityError/Anime') as iface: # Display the welcome message gr.Markdown(welcome_message) # Add a 'None' or similar option to represent no selection input_language_options = ["None"] + languages input_language = gr.Dropdown(input_language_options, label="Select the language", value="English", interactive=True) with gr.Accordion("Use Voice", open=False) as voice_accordion: audio_input = gr.Audio(label="Speak") audio_output = gr.Markdown(label="Output text") # Markdown component for audio gr.Examples([["audio1.wav"],["audio2.wav"],],inputs=[audio_input]) with gr.Accordion("Use a Picture", open=False) as picture_accordion: image_input = gr.Image(label="Upload image") image_output = gr.Markdown(label="Output text") # Markdown component for image gr.Examples([["image1.png"], ["image2.jpeg"], ["image3.jpeg"],],inputs=[image_input]) with gr.Accordion("MultiMed", open=False) as multimend_accordion: text_input = gr.Textbox(label="Use Text", lines=3, placeholder="I have had a sore throat and phlegm for a few days and now my cough has gotten worse!") gr.Examples([ ["What is the proper treatment for buccal herpes?"], ["I have had a sore throat and hoarse voice for serveral days and now a strong cough recently "], ["How does cellular metabolism work TCA cycle"], ["What special care must be provided to children with chicken pox?"], ["When and how often should I wash my hands?"], ["بکل ہرپس کا صحیح علاج کیا ہے؟"], ["구강 헤르페스의 적절한 치료법은 무엇입니까?"], ["Je, ni matibabu gani sahihi kwa herpes ya buccal?"], ],inputs=[text_input]) text_output = gr.Markdown(label="MultiMed") audio_output = gr.Audio(label="Audio Out", type="filepath") text_button = gr.Button("Use MultiMed") text_button.click(process_and_query, inputs=[input_language, audio_input, image_input, text_input], outputs=[text_output, audio_output]) clear_button = gr.Button("Clear") clear_button.click(clear, inputs=[], outputs=[input_language, audio_input, image_input, text_output, audio_output]) return iface iface = create_interface() iface.launch(show_error=True, debug=True)