app.py

import sys
import os
import pdfplumber
import json
import gradio as gr
from typing import List
from concurrent.futures import ThreadPoolExecutor, as_completed
import hashlib
import re
import psutil
import subprocess
from collections import defaultdict
from vllm import LLM, SamplingParams  # MODIFIED: Direct vLLM for batching

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

model_cache_dir = os.path.join(persistent_dir, "txagent_models")
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, 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"
os.environ["VLLM_NO_TORCH_COMPILE"] = "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

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_all_pages(file_path: str) -> str:
    try:
        text_chunks = []
        with pdfplumber.open(file_path) as pdf:
            for page in pdf.pages:
                page_text = page.extract_text() or ""
                text_chunks.append(page_text.strip())
        return "\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_all_pages(file_path)
            result = json.dumps({"filename": os.path.basename(file_path), "content": text, "status": "initial"})
        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 clean_response(text: str) -> str:
    text = sanitize_utf8(text)
    text = re.sub(r"\[TOOL_CALLS\].*?\n|\[.*?\].*?\n|(?:get_|tool\s|retrieve\s|use\s|rag\s).*?\n", "", text, flags=re.DOTALL | re.IGNORECASE)
    text = re.sub(r"\{'meta':\s*\{.*?\}\s*,\s*'results':\s*\[.*?\]\}\n?", "", text, flags=re.DOTALL)
    text = re.sub(
        r"(?i)(to\s|analyze|will\s|since\s|no\s|none|previous|attempt|involve|check\s|explore|manually|"
        r"start|look|use|focus|retrieve|tool|based\s|overall|indicate|mention|consider|ensure|need\s|"
        r"provide|review|assess|identify|potential|records|patient|history|symptoms|medication|"
        r"conflict|assessment|follow-up|issue|reasoning|step|prompt|address|rag|thought|try|john\sdoe|nkma).*?\n",
        "", text, flags=re.DOTALL
    )
    text = re.sub(r"\n{2,}", "\n", text).strip()
    lines = []
    valid_heading = False
    for line in text.split("\n"):
        line = line.strip()
        if line.lower() in ["missed diagnoses:", "medication conflicts:", "incomplete assessments:", "urgent follow-up:"]:
            valid_heading = True
            lines.append(f"**{line[:-1]}**:")
        elif valid_heading and line.startswith("-"):
            lines.append(line)
        else:
            valid_heading = False
    return "\n".join(lines).strip()

def normalize_text(text: str) -> str:
    return re.sub(r"\s+", " ", text.lower().strip())

def consolidate_findings(responses: List[str]) -> str:
    findings = defaultdict(set)
    headings = ["Missed Diagnoses", "Medication Conflicts", "Incomplete Assessments", "Urgent Follow-up"]
    
    for response in responses:
        if not response:
            continue
        current_heading = None
        for line in response.split("\n"):
            line = line.strip()
            if not line:
                continue
            if line.lower().startswith(tuple(h.lower() + ":" for h in headings)):
                current_heading = next(h for h in headings if line.lower().startswith(h.lower() + ":"))
            elif current_heading and line.startswith("-"):
                findings[current_heading].add(normalize_text(line))
    
    output = []
    for heading in headings:
        if findings[heading]:
            output.append(f"**{heading}**:")
            original_lines = {normalize_text(r): r for r in sum([r.split("\n") for r in responses], []) if r.startswith("-")}
            output.extend(sorted(original_lines.get(n, "- " + n) for n in findings[heading]))
    return "\n".join(output).strip() if output else "No oversights identified."

def init_agent():
    print("🔁 Initializing model...")
    log_system_usage("Before Load")
    model = LLM(
        model="mims-harvard/TxAgent-T1-Llama-3.1-8B",
        max_model_len=4096,  # MODIFIED: Reduce KV cache
        enforce_eager=True,
        enable_chunked_prefill=True,
        max_num_batched_tokens=8192,
    )
    log_system_usage("After Load")
    print("✅ Model Ready")
    return model

def create_ui(model):
    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"], 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 Report")

        def analyze(message: str, history: List[dict], files: List):
            history.append({"role": "user", "content": message})
            history.append({"role": "assistant", "content": "🔄 Analyzing..."})
            yield history, None

            extracted = ""
            file_hash_value = ""
            if files:
                with ThreadPoolExecutor(max_workers=6) 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([json.loads(r).get("content", "") for r in results if "content" in json.loads(r)])
                    file_hash_value = file_hash(files[0].name) if files else ""

            chunk_size = 800
            chunks = [extracted[i:i + chunk_size] for i in range(0, len(extracted), chunk_size)]
            chunk_responses = []
            batch_size = 8
            total_chunks = len(chunks)

            prompt_template = """
Output only oversights under these headings, one point each. No tools, reasoning, or extra text.

**Missed Diagnoses**:
**Medication Conflicts**:
**Incomplete Assessments**:
**Urgent Follow-up**:

Records:
{chunk}
"""
            sampling_params = SamplingParams(
                temperature=0.1,
                max_tokens=32,  # MODIFIED: Reduce for speed
                seed=100,
            )

            try:
                for i in range(0, len(chunks), batch_size):
                    batch = chunks[i:i + batch_size]
                    prompts = [prompt_template.format(chunk=chunk) for chunk in batch]
                    log_system_usage(f"Batch {i//batch_size + 1}")
                    outputs = model.generate(prompts, sampling_params)  # MODIFIED: Batch inference
                    batch_responses = []
                    with ThreadPoolExecutor(max_workers=8) as executor:  # MODIFIED: Parallel cleanup
                        futures = [executor.submit(clean_response, output.outputs[0].text) for output in outputs]
                        batch_responses.extend(f.result() for f in as_completed(futures))
                    chunk_responses.extend([r for r in batch_responses if r])
                    processed = min(i + len(batch), total_chunks)
                    history[-1]["content"] = f"🔄 Analyzing... ({processed}/{total_chunks} chunks)"
                    yield history, None

                final_response = consolidate_findings(chunk_responses)
                history[-1]["content"] = final_response
                yield history, None

                report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") if file_hash_value else None
                if report_path and final_response != "No oversights identified.":
                    with open(report_path, "w", encoding="utf-8") as f:
                        f.write(final_response)
                yield history, report_path if report_path and os.path.exists(report_path) else None

            except Exception as e:
                print("🚨 ERROR:", e)
                history[-1]["content"] = f"❌ Error: {str(e)}"
                yield history, 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

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