import sys import os import pandas as pd import json import gradio as gr from typing import List from concurrent.futures import ThreadPoolExecutor import hashlib import shutil import re import psutil import subprocess import logging import torch import gc from diskcache import Cache import time import asyncio # Try importing pypdfium2 and pytesseract, fall back to pdfplumber try: import pypdfium2 as pdfium import pytesseract from PIL import Image HAS_PYPDFIUM2 = True except ImportError: HAS_PYPDFIUM2 = False import pdfplumber # Configure logging logging.basicConfig(level=logging.INFO) logging.getLogger("pdfminer").setLevel(logging.ERROR) logger = logging.getLogger(__name__) # 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" 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() async def extract_all_pages_async(file_path: str, progress_callback=None, force_ocr=False) -> str: try: extracted_text = "" total_pages = 0 text_chunks = [] if HAS_PYPDFIUM2: pdf = pdfium.PdfDocument(file_path) total_pages = len(pdf) if total_pages == 0: return "" def extract_page(i): page = pdf[i] text = page.get_textpage().get_text_range() or "" if (not text.strip() or len(text) < 100) and force_ocr and 'pytesseract' in sys.modules: logger.info("Falling back to OCR for page %d", i + 1) bitmap = page.render(scale=2).to_pil() text = pytesseract.image_to_string(bitmap, lang="eng") return (i, f"=== Page {i + 1} ===\n{text.strip()}") with ThreadPoolExecutor(max_workers=4) as executor: futures = [executor.submit(extract_page, i) for i in range(total_pages)] for future in as_completed(futures): page_num, text = future.result() text_chunks.append((page_num, text)) logger.debug("Page %d extracted: %s...", page_num + 1, text[:50]) if progress_callback: progress_callback(page_num + 1, total_pages) text_chunks.sort(key=lambda x: x[0]) extracted_text = "\n\n".join(chunk[1] for chunk in text_chunks if chunk[1].strip()) pdf.close() else: with pdfplumber.open(file_path) as pdf: total_pages = len(pdf.pages) if total_pages == 0: return "" for i, page in enumerate(pdf.pages): text = page.extract_text() or "" text_chunks.append((i, f"=== Page {i + 1} ===\n{text.strip()}")) logger.debug("Page %d extracted: %s...", i + 1, text[:50]) if progress_callback: progress_callback(i + 1, total_pages) extracted_text = "\n\n".join(chunk[1] for chunk in text_chunks if chunk[1].strip()) logger.info("Extracted %d pages, total length: %d chars", total_pages, len(extracted_text)) if len(extracted_text) < 1000 and not force_ocr and HAS_PYPDFIUM2 and 'pytesseract' in sys.modules: logger.info("Text too short, retrying with OCR") return await extract_all_pages_async(file_path, progress_callback, force_ocr=True) return extracted_text except Exception as e: logger.error("PDF processing error: %s", 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: logger.info("Using cached extraction for %s", file_path) return cache[cache_key] if file_type == "pdf": text = asyncio.run(extract_all_pages_async(file_path, progress_callback, force_ocr=False)) 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}"}) cache[cache_key] = result logger.info("Cached extraction for %s, size: %d bytes", file_path, len(result)) return result except Exception as e: logger.error("Error processing %s: %s", os.path.basename(file_path), 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() logger.info("[%s] CPU: %.1f%% | RAM: %dMB / %dMB", tag, cpu, mem.used // (1024**2), mem.total // (1024**2)) 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(", ") logger.info("[%s] GPU: %sMB / %sMB | Utilization: %s%%", tag, used, total, util) except Exception as e: logger.error("[%s] GPU/CPU monitor failed: %s", tag, e) def clean_response(text: str) -> str: text = sanitize_utf8(text) text = text.replace("[", "").replace("]", "").replace("None", "") text = text.replace("\n\n\n", "\n\n") sections = {} current_section = None seen_lines = set() for line in text.splitlines(): line = line.strip() if not line or line in seen_lines: continue seen_lines.add(line) 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) sections.setdefault(current_section, []) continue if current_section and line.startswith("- "): sections[current_section].append(line) cleaned = [f"### {heading}\n" + "\n".join(findings) for heading, findings in sections.items() if findings] result = "\n\n".join(cleaned).strip() logger.debug("Cleaned response length: %d chars", len(result)) return result or "No oversights identified" def summarize_findings(all_responses: List[str]) -> str: combined_response = "\n\n".join(all_responses) if not combined_response or all("No oversights identified" in resp.lower() for resp in all_responses): return "### Comprehensive Clinical Oversight Summary\nNo critical oversights were identified across the provided patient records after thorough analysis." sections = { "Missed Diagnoses": [], "Medication Conflicts": [], "Incomplete Assessments": [], "Urgent Follow-up": [] } current_section = None seen_findings = set() for line in combined_response.splitlines(): line = line.strip() if not line: 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) continue if current_section and line.startswith("- ") and line not in seen_findings: sections[current_section].append(line) seen_findings.add(line) summary_lines = [] for heading, findings in sections.items(): if findings: summary_lines.append(f"### {heading}") for finding in findings: summary_lines.append(f"{finding}\n - **Risks**: Potential adverse outcomes if not addressed.\n - **Recommendation**: Immediate clinical review and follow-up.") result = "### Comprehensive Clinical Oversight Summary\n" + "\n".join(summary_lines) if summary_lines else "### Comprehensive Clinical Oversight Summary\nNo critical oversights identified." logger.debug("Summary length: %d chars", len(result)) return result def init_agent(): logger.info("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, enable_rag=False, init_rag_num=0, step_rag_num=0, seed=100, additional_default_tools=[], ) agent.init_model() log_system_usage("After Load") logger.info("Agent Ready") return agent def create_ui(agent): with gr.Blocks(theme=gr.themes.Soft()) as demo: gr.Markdown("