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 import threading import torch from diskcache import Cache import time # Persistent directory 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" os.environ["OMP_NUM_THREADS"] = str(os.cpu_count() // 2) # Optimize CPU threading 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 # Initialize cache with 10GB limit cache = Cache(file_cache_dir, size_limit=10 * 1024**3) 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, progress_callback=None) -> str: try: with pdfplumber.open(file_path) as pdf: total_pages = len(pdf.pages) if total_pages == 0: return "" batch_size = 10 # Process 10 pages per thread batches = [(i, min(i + batch_size, total_pages)) for i in range(0, total_pages, batch_size)] text_chunks = [""] * total_pages # Pre-allocate for page order processed_pages = 0 def extract_batch(start: int, end: int) -> List[tuple]: results = [] with pdfplumber.open(file_path) as pdf: # Reopen per thread for page in pdf.pages[start:end]: page_num = start + pdf.pages.index(page) page_text = page.extract_text() or "" results.append((page_num, f"=== Page {page_num + 1} ===\n{page_text.strip()}")) return results with ThreadPoolExecutor(max_workers=min(6, os.cpu_count())) as executor: futures = [executor.submit(extract_batch, start, end) for start, end in batches] for future in as_completed(futures): for page_num, text in future.result(): text_chunks[page_num] = text processed_pages += batch_size if progress_callback: progress_callback(min(processed_pages, total_pages), total_pages) return "\n\n".join(filter(None, text_chunks)) except Exception as e: return f"PDF processing error: {str(e)}" def convert_file_to_json(file_path: str, file_type: str, progress_callback=None) -> str: try: file_h = file_hash(file_path) cache_key = f"{file_h}_{file_type}" if cache_key in cache: return cache[cache_key] if file_type == "pdf": text = extract_all_pages(file_path, progress_callback) 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"]: df = pd.read_excel(file_path, engine="openpyxl", 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}"}) cache[cache_key] = 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"\[.*?\]|\bNone\b|To analyze the patient record excerpt.*?medications\.|Since the previous attempts.*?\.|I need to.*?medications\.|Retrieving tools.*?\.", "", text, flags=re.DOTALL) text = re.sub(r"\n{3,}", "\n\n", text) text = re.sub(r"[^\n#\-\*\w\s\.\,\:\(\)]+", "", text) tool_to_heading = { "get_abuse_info_by_drug_name": "Drugs", "get_dependence_info_by_drug_name": "Drugs", "get_abuse_types_and_related_adverse_reactions_and_controlled_substance_status_by_drug_name": "Drugs", "get_info_for_patients_by_drug_name": "Drugs", } sections = {} current_section = None current_tool = None lines = text.splitlines() for line in lines: line = line.strip() if not line: continue tool_match = re.match(r"\[TOOL:\s*(\w+)\]", line) if tool_match: current_tool = tool_match.group(1) continue section_match = re.match(r"###\s*(Missed Diagnoses|Medication Conflicts|Incomplete Assessments|Urgent Follow-up)", line) if section_match: current_section = section_match.group(1) if current_section not in sections: sections[current_section] = [] continue finding_match = re.match(r"-\s*.+", line) if finding_match and current_section and not re.match(r"-\s*No issues identified", line): if current_tool and current_tool in tool_to_heading: heading = tool_to_heading[current_tool] if heading not in sections: sections[heading] = [] sections[heading].append(line) else: sections[current_section].append(line) cleaned = [] for heading, findings in sections.items(): if findings: cleaned.append(f"### {heading}\n" + "\n".join(findings)) text = "\n\n".join(cleaned).strip() return text if text else "" 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=False, # Disable checker for speed step_rag_num=4, seed=100, additional_default_tools=[], ) def preload_models(): agent.init_model() log_system_usage("After Load") preload_thread = threading.Thread(target=preload_models) preload_thread.start() preload_thread.join() print("✅ Agent Ready") return agent def create_ui(agent): with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.Markdown("

🩺 Clinical Oversight Assistant

") 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") progress_bar = gr.Progress() prompt_template = """ Analyze the patient record excerpt for clinical oversights. Provide a concise, evidence-based summary in markdown with findings grouped under tool-derived headings (e.g., 'Drugs'). For each finding, include clinical context, risks, and recommendations. Precede findings with a tool tag (e.g., [TOOL: get_abuse_info_by_drug_name]). Output only markdown bullet points under headings. If no issues, state "No issues identified". Patient Record Excerpt (Chunk {0} of {1}): {chunk} """ def analyze(message: str, history: List[dict], files: List, progress=gr.Progress()): history.append({"role": "user", "content": message}) yield history, None, None extracted = "" file_hash_value = "" if files: def update_extraction_progress(current, total): progress(current / total, desc=f"Extracting text... Page {current}/{total}") return history, None, None with ThreadPoolExecutor(max_workers=6) as executor: futures = [executor.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower(), update_extraction_progress) 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) if files else "" history.append({"role": "assistant", "content": "✅ Text extraction complete."}) yield history, None, None chunk_size = 6000 chunks = [extracted[i:i + chunk_size] for i in range(0, len(extracted), chunk_size)] combined_response = "" batch_size = 2 try: for batch_idx in range(0, len(chunks), batch_size): batch_chunks = chunks[batch_idx:batch_idx + batch_size] batch_prompts = [prompt_template.format(i + 1, len(chunks), chunk=chunk[:4000]) for i, chunk in enumerate(batch_chunks)] batch_responses = [] progress((batch_idx + 1) / len(chunks), desc=f"Analyzing chunks {batch_idx + 1}-{min(batch_idx + batch_size, len(chunks))}/{len(chunks)}") with ThreadPoolExecutor(max_workers=len(batch_chunks)) as executor: futures = [executor.submit(agent.run_gradio_chat, prompt, [], 0.2, 512, 2048, False, []) for prompt in batch_prompts] for future in as_completed(futures): chunk_response = "" for chunk_output in future.result(): if chunk_output is None: continue if isinstance(chunk_output, list): for m in chunk_output: if hasattr(m, 'content') and m.content: cleaned = clean_response(m.content) if cleaned and re.search(r"###\s*\w+", cleaned): chunk_response += cleaned + "\n\n" elif isinstance(chunk_output, str) and chunk_output.strip(): cleaned = clean_response(chunk_output) if cleaned and re.search(r"###\s*\w+", cleaned): chunk_response += cleaned + "\n\n" batch_responses.append(chunk_response) for chunk_idx, chunk_response in enumerate(batch_responses, batch_idx + 1): if chunk_response: combined_response += f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response}\n" else: combined_response += f"--- Analysis for Chunk {chunk_idx} ---\nNo oversights identified for this chunk.\n\n" history[-1] = {"role": "assistant", "content": combined_response.strip()} yield history, None, None if combined_response.strip() and not all("No oversights identified" in chunk for chunk in combined_response.split("--- Analysis for Chunk")): history[-1]["content"] = combined_response.strip() else: history.append({"role": "assistant", "content": "No oversights identified in the provided records."}) report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") if file_hash_value else None if report_path: with open(report_path, "w", encoding="utf-8") as f: f.write(combined_response) yield history, report_path if report_path and os.path.exists(report_path) else None, None except Exception as e: print("🚨 ERROR:", e) history.append({"role": "assistant", "content": f"❌ Error occurred: {str(e)}"}) yield history, None, None send_btn.click(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output, progress_bar]) msg_input.submit(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output, progress_bar]) return demo 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 )