app.py

import sys
import os
import pandas as pd
import pdfplumber
import json
import gradio as gr
from typing import List
from concurrent.futures import ThreadPoolExecutor, as_completed
import hashlib
import shutil
import re
import psutil
import subprocess

# ---------------------------------------------------------------------------------------
# Persistent directory setup
# ---------------------------------------------------------------------------------------
persistent_dir = "/data/hf_cache"
os.makedirs(persistent_dir, exist_ok=True)

model_cache_dir = os.path.join(persistent_dir, "txagent_models")
tool_cache_dir = os.path.join(persistent_dir, "tool_cache")
file_cache_dir = os.path.join(persistent_dir, "cache")
report_dir = os.path.join(persistent_dir, "reports")
vllm_cache_dir = os.path.join(persistent_dir, "vllm_cache")

for directory in [model_cache_dir, tool_cache_dir, file_cache_dir, report_dir, vllm_cache_dir]:
    os.makedirs(directory, exist_ok=True)

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

current_dir = os.path.dirname(os.path.abspath(__file__))
src_path = os.path.abspath(os.path.join(current_dir, "src"))
sys.path.insert(0, src_path)

from txagent.txagent import TxAgent

# ---------------------------------------------------------------------------------------
# Utilities
# ---------------------------------------------------------------------------------------
MEDICAL_KEYWORDS = {
    'diagnosis', 'assessment', 'plan', 'results', 'medications',
    'allergies', 'summary', 'impression', 'findings', 'recommendations'
}

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()

def extract_priority_pages(file_path: str, max_pages: int = 20) -> str:
    try:
        text_chunks = []
        with pdfplumber.open(file_path) as pdf:
            for i, page in enumerate(pdf.pages[:3]):
                text = page.extract_text() or ""
                text_chunks.append(f"=== Page {i+1} ===\n{text.strip()}")
            for i, page in enumerate(pdf.pages[3:max_pages], start=4):
                page_text = page.extract_text() or ""
                if any(re.search(rf'\b{kw}\b', page_text.lower()) for kw in MEDICAL_KEYWORDS):
                    text_chunks.append(f"=== Page {i} ===\n{page_text.strip()}")
        return "\n\n".join(text_chunks)
    except Exception as e:
        return f"PDF processing error: {str(e)}"

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):
            with open(cache_path, "r", encoding="utf-8") as f:
                return f.read()

        if file_type == "pdf":
            text = extract_priority_pages(file_path)
            result = json.dumps({"filename": os.path.basename(file_path), "content": text, "status": "initial"})
        elif file_type == "csv":
            df = pd.read_csv(file_path, encoding_errors="replace", header=None, dtype=str, 
                             skip_blank_lines=False, on_bad_lines="skip")
            content = df.fillna("").astype(str).values.tolist()
            result = json.dumps({"filename": os.path.basename(file_path), "rows": content})
        elif file_type in ["xls", "xlsx"]:
            try:
                df = pd.read_excel(file_path, engine="openpyxl", header=None, dtype=str)
            except Exception:
                df = pd.read_excel(file_path, engine="xlrd", header=None, dtype=str)
            content = df.fillna("").astype(str).values.tolist()
            result = json.dumps({"filename": os.path.basename(file_path), "rows": content})
        else:
            result = json.dumps({"error": f"Unsupported file type: {file_type}"})
        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 processing {os.path.basename(file_path)}: {str(e)}"})

def log_system_usage(tag=""):
    try:
        cpu = psutil.cpu_percent(interval=1)
        mem = psutil.virtual_memory()
        print(f"[{tag}] CPU: {cpu}% | RAM: {mem.used // (1024**2)}MB / {mem.total // (1024**2)}MB")
        result = subprocess.run(
            ["nvidia-smi", "--query-gpu=memory.used,memory.total,utilization.gpu", "--format=csv,nounits,noheader"],
            capture_output=True, text=True
        )
        if result.returncode == 0:
            used, total, util = result.stdout.strip().split(", ")
            print(f"[{tag}] GPU: {used}MB / {total}MB | Utilization: {util}%")
    except Exception as e:
        print(f"[{tag}] ⚠️ GPU/CPU monitor failed: {e}")

def init_agent():
    print("🔁 Initializing model...")
    log_system_usage("Before Load")

    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)

    agent = TxAgent(
        model_name="mims-harvard/TxAgent-T1-Llama-3.1-8B",
        rag_model_name="mims-harvard/ToolRAG-T1-GTE-Qwen2-1.5B",
        tool_files_dict={"new_tool": target_tool_path},
        force_finish=True,
        enable_checker=True,
        step_rag_num=8,
        seed=100,
        additional_default_tools=[],
    )
    agent.init_model()
    log_system_usage("After Load")
    print("✅ Agent Ready")

    return agent

# ---------------------------------------------------------------------------------------
# Gradio UI
# ---------------------------------------------------------------------------------------
def create_ui(agent):
    with gr.Blocks(theme=gr.themes.Soft()) as demo:
        gr.Markdown("<h1 style='text-align: center;'>🩺 Clinical Oversight Assistant</h1>")
        chatbot = gr.Chatbot(label="Analysis", height=600, type="messages")
        file_upload = gr.File(file_types=[".pdf", ".csv", ".xls", ".xlsx"], file_count="multiple")
        msg_input = gr.Textbox(placeholder="Ask about potential oversights...", show_label=False)
        send_btn = gr.Button("Analyze", variant="primary")
        download_output = gr.File(label="Download Full Report")

        def analyze(message: str, history: list, files: list):
            history = history + [{"role": "user", "content": message},
                                 {"role": "assistant", "content": "⏳ Analyzing records for potential oversights..."}]
            yield history, None

            extracted = ""
            file_hash_value = ""
            if files:
                with ThreadPoolExecutor(max_workers=4) as executor:
                    futures = [executor.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower()) for f in files]
                    results = [sanitize_utf8(f.result()) for f in as_completed(futures)]
                    extracted = "\n".join(results)
                    file_hash_value = file_hash(files[0].name)

            prompt = f"""Review these medical records and identify EXACTLY what might have been missed:
1. List potential missed diagnoses
2. Flag any medication conflicts
3. Note incomplete assessments
4. Highlight abnormal results needing follow-up

Medical Records:
{extracted[:12000]}

### Potential Oversights:
"""
            response = ""
            try:
                for chunk in agent.run_gradio_chat(
                    message=prompt,
                    history=[],
                    temperature=0.2,
                    max_new_tokens=1024,
                    max_token=4096,
                    call_agent=False,
                    conversation=[]
                ):
                    if chunk is None: continue
                    if isinstance(chunk, str):
                        response += chunk
                    elif isinstance(chunk, list):
                        response += "".join([c.content for c in chunk if hasattr(c, 'content')])
                    cleaned = response.strip()
                    cleaned = re.sub(r"\[TOOL_CALLS\].*?$", "", cleaned, flags=re.DOTALL).strip()
                    history[-1] = {"role": "assistant", "content": cleaned or "⏳ Processing..."}
                    yield history, None
            except Exception as e:
                history[-1] = {"role": "assistant", "content": f"❌ Error: {str(e)}"}
                yield history, None
                return

            final_output = re.sub(r"\[TOOL_CALLS\].*?$", "", response, flags=re.DOTALL).strip()
            if not final_output:
                final_output = "No clear oversights identified. Recommend comprehensive review."
            history[-1] = {"role": "assistant", "content": final_output}

            report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") if file_hash_value else None
            yield history, report_path if report_path and os.path.exists(report_path) else None

        send_btn.click(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output])
        msg_input.submit(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output])
    return demo

# ---------------------------------------------------------------------------------------
# Launch
# ---------------------------------------------------------------------------------------
if __name__ == "__main__":
    print("🚀 Launching app...")
    agent = init_agent()
    demo = create_ui(agent)
    demo.queue(api_open=False).launch(
        server_name="0.0.0.0",
        server_port=7860,
        show_error=True,
        allowed_paths=[report_dir],
        share=False
    )