app.py

import spaces
import torch
import gradio as gr
from transformers import AutoTokenizer, LlamaForCausalLM
import bitsandbytes, flash_attn
import os

MODEL_NAME = "openai/whisper-large-v3-turbo"
BATCH_SIZE = 8
FILE_LIMIT_MB = 1000

device = 0 if torch.cuda.is_available() else "cpu"

# Initialize the transcription pipeline
pipe = pipeline(
    task="automatic-speech-recognition",
    model=MODEL_NAME,
    chunk_length_s=30,
    device=device,
)

# Load tokenizer and model for SOAP note generation
tokenizer = AutoTokenizer.from_pretrained("NousResearch/Hermes-3-Llama-3.1-8B", trust_remote_code=True)
model = LlamaForCausalLM.from_pretrained(
    "NousResearch/Hermes-3-Llama-3.1-8B",
    torch_dtype=torch.float16,
    device_map="auto",
    load_in_8bit=False,
    load_in_4bit=True,
    use_flash_attention_2=True
)

# Prompt for SOAP note generation
sys_prompt = "You are a world class clinical assistant."
task_prompt = """
Convert the following transcribed conversation into a clinical SOAP note.
The text includes dialogue between a physician and a patient. Please clearly distinguish between the physician's and the patient's statements.
Extract and organize the information into the relevant sections of a SOAP note:
- Subjective (symptoms and patient statements),
- Objective (clinical findings and observations, these might be missing if the physician has not conducted a physical exam or has not verbally stated findings),
- Assessment (diagnosis or potential diagnoses, objectively provide a top 5 most likely diagnosis based on just the subjective findings, and use the objective findings if available),
- Plan (treatment and follow-up).
Ensure the note is concise, clear, and accurately reflects the conversation.
"""

# Function to transcribe audio inputs
@spaces.GPU
def transcribe(inputs, task):
    if inputs is None:
        raise gr.Error("No audio file submitted! Please upload or record an audio file before submitting your request.")
    text = pipe(inputs, batch_size=BATCH_SIZE, generate_kwargs={"task": task}, return_timestamps=True)["text"]
    return text

# Function to generate SOAP notes using LLM
def generate_soap(transcribed_text):
    prompt = f"<|im_start|>system\n{sys_prompt}<|im_end|>\n<|im_start|>user\n{task_prompt}\n{transcribed_text}<|im_end|>\n<|im_start|>assistant"
    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to("cuda")
    generated_ids = model.generate(input_ids, max_new_tokens=2048, temperature=0.8, repetition_penalty=1.1, do_sample=True, eos_token_id=tokenizer.eos_token_id)
    response = tokenizer.decode(generated_ids[0][input_ids.shape[-1]:], skip_special_tokens=True, clean_up_tokenization_space=True)
    return response

# Gradio Interfaces for different inputs
demo = gr.Blocks(theme=gr.themes.Ocean())

# Interface for microphone or file transcription
mf_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[gr.Audio(sources="microphone", type="filepath"), gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")],
    outputs="text",
    title="Audio Transcribe",
    description="Transcribe long-form microphone or audio inputs."
)

file_transcribe = gr.Interface(
    fn=transcribe,
    inputs=[gr.Audio(sources="upload", type="filepath", label="Audio file"), gr.Radio(["transcribe", "translate"], label="Task", value="transcribe")],
    outputs="text",
    title="Audio Transcribe"
)

# SOAP Note generation interface
soap_note = gr.Interface(
    fn=generate_soap,
    inputs="text",
    outputs="text",
    title="Generate Clinical SOAP Note",
    description="Convert transcribed conversation to a clinical SOAP note with structured sections (Subjective, Objective, Assessment, Plan)."
)

# Tabbed interface integrating SOAP note below transcription
with demo:
    with gr.TabbedInterface([mf_transcribe, file_transcribe], ["Microphone", "Audio file"]) as transcribe_tab:
        transcribe_tab.outputs[0]  # Output from transcription feeds directly to SOAP note
        soap_note  # SOAP note interface placed directly below transcription output

demo.queue().launch(ssr_mode=False)