app.py

import sys
import os
import pandas as pd
import pdfplumber
import json
import gradio as gr
from typing import List, Optional
from concurrent.futures import ThreadPoolExecutor, as_completed
import hashlib
import shutil
import time
from functools import lru_cache

# Environment and path setup
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "src")))

# Configure cache directories
base_dir = "/data"
model_cache_dir = os.path.join(base_dir, "txagent_models")
tool_cache_dir = os.path.join(base_dir, "tool_cache")
file_cache_dir = os.path.join(base_dir, "cache")

os.makedirs(model_cache_dir, exist_ok=True)
os.makedirs(tool_cache_dir, exist_ok=True)
os.makedirs(file_cache_dir, exist_ok=True)

os.environ["TRANSFORMERS_CACHE"] = model_cache_dir
os.environ["HF_HOME"] = model_cache_dir
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["CUDA_LAUNCH_BLOCKING"] = "1"

from txagent.txagent import TxAgent

# Utility functions
def sanitize_utf8(text: str) -> str:
    return text.encode("utf-8", "ignore").decode("utf-8")

def file_hash(path: str) -> str:
    with open(path, "rb") as f:
        return hashlib.md5(f.read()).hexdigest()

@lru_cache(maxsize=100)
def get_cached_response(prompt: str, file_hash: str) -> Optional[str]:
    """Cache for frequent queries"""
    return None  # Implement actual cache lookup if needed

def convert_file_to_json(file_path: str, file_type: str) -> str:
    try:
        h = file_hash(file_path)
        cache_path = os.path.join(file_cache_dir, f"{h}.json")

        if os.path.exists(cache_path):
            return open(cache_path, "r", encoding="utf-8").read()

        if file_type == "csv":
            df = pd.read_csv(file_path, encoding_errors="replace", header=None, 
                            dtype=str, skip_blank_lines=False, on_bad_lines="skip")
        elif file_type in ["xls", "xlsx"]:
            try:
                df = pd.read_excel(file_path, engine="openpyxl", header=None, dtype=str)
            except:
                df = pd.read_excel(file_path, engine="xlrd", header=None, dtype=str)
        elif file_type == "pdf":
            with pdfplumber.open(file_path) as pdf:
                text = "\n".join([page.extract_text() or "" for page in pdf.pages])
            result = json.dumps({"filename": os.path.basename(file_path), "content": text.strip()})
            with open(cache_path, "w", encoding="utf-8") as f:
                f.write(result)
            return result
        else:
            return json.dumps({"error": f"Unsupported file type: {file_type}"})

        if df is None or df.empty:
            return json.dumps({"warning": f"No data extracted from: {file_path}"})

        df = df.fillna("")
        content = df.astype(str).values.tolist()
        result = json.dumps({"filename": os.path.basename(file_path), "rows": content})
        with open(cache_path, "w", encoding="utf-8") as f:
            f.write(result)
        return result
    except Exception as e:
        return json.dumps({"error": f"Error reading {os.path.basename(file_path)}: {str(e)}"})

def convert_files_to_json_parallel(uploaded_files: list) -> str:
    """Process files in parallel using ThreadPool"""
    extracted_text = []
    with ThreadPoolExecutor(max_workers=4) as executor:
        futures = []
        for file in uploaded_files:
            if not hasattr(file, 'name'):
                continue
            path = file.name
            ext = path.split(".")[-1].lower()
            futures.append(executor.submit(convert_file_to_json, path, ext))
        
        for future in as_completed(futures):
            extracted_text.append(sanitize_utf8(future.result()))
    return "\n".join(extracted_text)

def init_agent():
    """Initialize the TxAgent with optimized settings"""
    # Copy default tool file if needed
    default_tool_path = os.path.abspath("data/new_tool.json")
    target_tool_path = os.path.join(tool_cache_dir, "new_tool.json")
    if not os.path.exists(target_tool_path):
        shutil.copy(default_tool_path, target_tool_path)

    model_name = "mims-harvard/TxAgent-T1-Llama-3.1-8B"
    rag_model_name = "mims-harvard/ToolRAG-T1-GTE-Qwen2-1.5B"
    
    agent = TxAgent(
        model_name=model_name,
        rag_model_name=rag_model_name,
        tool_files_dict={"new_tool": target_tool_path},
        force_finish=True,
        enable_checker=True,
        step_rag_num=8,  # Reduced from 10
        seed=100,
        additional_default_tools=[],
        torch_dtype="auto",
        device_map="auto",
        load_in_4bit=False,
        load_in_8bit=False
    )
    agent.init_model()
    return agent

def create_ui(agent: TxAgent):
    with gr.Blocks(theme=gr.themes.Soft()) as demo:
        gr.Markdown("<h1 style='text-align: center;'>📋 CPS: Clinical Patient Support System</h1>")

        chatbot = gr.Chatbot(label="CPS Assistant", height=600, type="messages")
        file_upload = gr.File(
            label="Upload Medical File",
            file_types=[".pdf", ".txt", ".docx", ".jpg", ".png", ".csv", ".xls", ".xlsx"],
            file_count="multiple"
        )
        message_input = gr.Textbox(placeholder="Ask a biomedical question or just upload the files...", show_label=False)
        send_button = gr.Button("Send", variant="primary")
        conversation_state = gr.State([])

        def handle_chat(message: str, history: list, conversation: list, uploaded_files: list, progress=gr.Progress()):
            start_time = time.time()
            try:
                history.append({"role": "user", "content": message})
                history.append({"role": "assistant", "content": "⏳ Processing your request..."})
                yield history

                # File processing with timing
                file_process_time = time.time()
                extracted_text = ""
                if uploaded_files and isinstance(uploaded_files, list):
                    extracted_text = convert_files_to_json_parallel(uploaded_files)
                print(f"File processing took: {time.time() - file_process_time:.2f}s")

                context = (
                    "You are an expert clinical AI assistant. Review this patient's history, "
                    "medications, and notes, and ONLY provide a final answer summarizing "
                    "what the doctor might have missed."
                )
                chunked_prompt = f"{context}\n\n--- Patient Record ---\n{extracted_text}\n\n[Final Analysis]"

                # Model processing with timing
                model_start = time.time()
                generator = agent.run_gradio_chat(
                    message=chunked_prompt,
                    history=[],
                    temperature=0.3,
                    max_new_tokens=768,  # Reduced from 1024
                    max_token=4096,      # Reduced from 8192
                    call_agent=False,
                    conversation=conversation,
                    uploaded_files=uploaded_files,
                    max_round=10       # Reduced from 30
                )

                final_response = []
                for update in generator:
                    if not update:
                        continue
                    if isinstance(update, str):
                        final_response.append(update)
                    elif isinstance(update, list):
                        final_response.extend(msg.content for msg in update if hasattr(msg, 'content'))
                    
                    # Yield intermediate results periodically
                    if len(final_response) % 3 == 0:  # More frequent updates
                        history[-1] = {"role": "assistant", "content": "".join(final_response).strip()}
                        yield history

                history[-1] = {"role": "assistant", "content": "".join(final_response).strip() or "❌ No response."}
                print(f"Model processing took: {time.time() - model_start:.2f}s")
                yield history

            except Exception as chat_error:
                print(f"Chat handling error: {chat_error}")
                history[-1] = {"role": "assistant", "content": "❌ An error occurred while processing your request."}
                yield history
            finally:
                print(f"Total request time: {time.time() - start_time:.2f}s")

        inputs = [message_input, chatbot, conversation_state, file_upload]
        send_button.click(fn=handle_chat, inputs=inputs, outputs=chatbot)
        message_input.submit(fn=handle_chat, inputs=inputs, outputs=chatbot)

        gr.Examples([
            ["Upload your medical form and ask what the doctor might've missed."],
            ["This patient was treated with antibiotics for UTI. What else should we check?"],
            ["Is there anything abnormal in the attached blood work report?"]
        ], inputs=message_input)

    return demo

if __name__ == "__main__":
    # Initialize agent and warm it up
    print("Initializing agent...")
    agent = init_agent()
    
    # Warm-up call
    print("Performing warm-up call...")
    try:
        warm_up = agent.run_gradio_chat(
            message="Warm up",
            history=[],
            temperature=0.1,
            max_new_tokens=10,
            max_token=100,
            call_agent=False
        )
        for _ in warm_up:
            pass
    except:
        pass
    
    # Launch Gradio interface
    print("Launching interface...")
    demo = create_ui(agent)
    demo.queue(concurrency_count=3).launch(
        server_name="0.0.0.0",
        server_port=7860,
        show_error=True,
        share=True
    )