app.py

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

# 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

# Constants
MEDICAL_KEYWORDS = {
    'diagnosis', 'assessment', 'plan', 'results', 'medications',
    'allergies', 'summary', 'impression', 'findings', 'recommendations',
    'conclusion', 'history', 'examination', 'progress', 'discharge'
}
TOKENIZER = "cl100k_base"
MAX_MODEL_LEN = 2048  # Matches your model's actual limit
TARGET_CHUNK_TOKENS = 1200  # Reduced to ensure room for prompt and response
PROMPT_RESERVE = 300  # Tokens reserved for prompt structure
MEDICAL_SECTION_HEADER = "=== MEDICAL SECTION ==="

def sanitize_utf8(text: str) -> str:
    """Ensure text is UTF-8 clean."""
    return text.encode("utf-8", "ignore").decode("utf-8")

def file_hash(path: str) -> str:
    """Generate MD5 hash of file content."""
    with open(path, "rb") as f:
        return hashlib.md5(f.read()).hexdigest()

def count_tokens(text: str) -> int:
    """Count tokens using the same method as the model"""
    encoding = tiktoken.get_encoding(TOKENIZER)
    return len(encoding.encode(text))

def extract_all_pages_with_token_count(file_path: str) -> Tuple[str, int, int]:
    """
    Extract all pages from PDF with token counting.
    Returns (extracted_text, total_pages, total_tokens)
    """
    try:
        text_chunks = []
        total_pages = 0
        total_tokens = 0
        
        with pdfplumber.open(file_path) as pdf:
            total_pages = len(pdf.pages)
            
            for i, page in enumerate(pdf.pages):
                page_text = page.extract_text() or ""
                lower_text = page_text.lower()
                
                # Mark medical sections
                if any(re.search(rf'\b{kw}\b', lower_text) for kw in MEDICAL_KEYWORDS):
                    section_header = f"\n{MEDICAL_SECTION_HEADER} (Page {i+1})\n"
                    text_chunks.append(section_header + page_text.strip())
                    total_tokens += count_tokens(section_header)
                else:
                    text_chunks.append(f"\n=== Page {i+1} ===\n{page_text.strip()}")
                
                total_tokens += count_tokens(page_text)
        
        return "\n".join(text_chunks), total_pages, total_tokens
    except Exception as e:
        return f"PDF processing error: {str(e)}", 0, 0

def convert_file_to_json(file_path: str, file_type: str) -> str:
    """Convert file to JSON format with caching and token counting."""
    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, total_pages, total_tokens = extract_all_pages_with_token_count(file_path)
            result = json.dumps({
                "filename": os.path.basename(file_path),
                "content": text,
                "total_pages": total_pages,
                "total_tokens": total_tokens,
                "status": "complete"
            })
        elif file_type == "csv":
            chunks = []
            for chunk in pd.read_csv(file_path, encoding_errors="replace", header=None, dtype=str,
                                   skip_blank_lines=False, on_bad_lines="skip", chunksize=1000):
                chunks.append(chunk.fillna("").astype(str).values.tolist())
            content = [item for sublist in chunks for item in sublist]
            result = json.dumps({
                "filename": os.path.basename(file_path),
                "rows": content,
                "total_tokens": count_tokens(str(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,
                "total_tokens": count_tokens(str(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 clean_response(text: str) -> str:
    """Clean and format the model response."""
    text = sanitize_utf8(text)
    text = re.sub(r"\[TOOL_CALLS\].*", "", text, flags=re.DOTALL)
    text = re.sub(r"\['get_[^\]]+\']\n?", "", text)
    text = re.sub(r"\{'meta':\s*\{.*?\}\s*,\s*'results':\s*\[.*?\]\}\n?", "", text, flags=re.DOTALL)
    text = re.sub(r"To analyze the medical records for clinical oversights.*?begin by reviewing.*?\n", "", text, flags=re.DOTALL)
    text = re.sub(r"\n{3,}", "\n\n", text).strip()
    return text

def format_final_report(analysis_results: List[str], filename: str) -> str:
    """Combine all analysis chunks into a well-formatted final report."""
    report = []
    report.append(f"COMPREHENSIVE CLINICAL OVERSIGHT ANALYSIS")
    report.append(f"Generated: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}")
    report.append(f"File: {filename}")
    report.append("=" * 80)
    
    sections = {
        "CRITICAL FINDINGS": [],
        "MISSED DIAGNOSES": [],
        "MEDICATION ISSUES": [],
        "ASSESSMENT GAPS": [],
        "FOLLOW-UP RECOMMENDATIONS": []
    }
    
    for result in analysis_results:
        for section in sections:
            section_match = re.search(
                rf"{re.escape(section)}:?\s*\n([^*]+?)(?=\n\*|\n\n|$)", 
                result, 
                re.IGNORECASE | re.DOTALL
            )
            if section_match:
                content = section_match.group(1).strip()
                if content and content not in sections[section]:
                    sections[section].append(content)
    
    if sections["CRITICAL FINDINGS"]:
        report.append("\n🚨 **CRITICAL FINDINGS** 🚨")
        for content in sections["CRITICAL FINDINGS"]:
            report.append(f"\n{content}")
    
    for section, contents in sections.items():
        if section != "CRITICAL FINDINGS" and contents:
            report.append(f"\n**{section.upper()}**")
            for content in contents:
                report.append(f"\n{content}")
    
    if not any(sections.values()):
        report.append("\nNo significant clinical oversights identified.")
    
    report.append("\n" + "=" * 80)
    report.append("END OF REPORT")
    
    return "\n".join(report)

def split_content_by_tokens(content: str, max_tokens: int = TARGET_CHUNK_TOKENS) -> List[str]:
    """Split content into chunks that fit within token limits"""
    paragraphs = re.split(r"\n\s*\n", content)
    chunks = []
    current_chunk = []
    current_tokens = 0
    
    for para in paragraphs:
        para_tokens = count_tokens(para)
        if para_tokens > max_tokens:
            sentences = re.split(r'(?<=[.!?])\s+', para)
            for sent in sentences:
                sent_tokens = count_tokens(sent)
                if current_tokens + sent_tokens > max_tokens:
                    chunks.append("\n\n".join(current_chunk))
                    current_chunk = [sent]
                    current_tokens = sent_tokens
                else:
                    current_chunk.append(sent)
                    current_tokens += sent_tokens
        elif current_tokens + para_tokens > max_tokens:
            chunks.append("\n\n".join(current_chunk))
            current_chunk = [para]
            current_tokens = para_tokens
        else:
            current_chunk.append(para)
            current_tokens += para_tokens
    
    if current_chunk:
        chunks.append("\n\n".join(current_chunk))
    
    return chunks

def init_agent():
    """Initialize the TxAgent with proper configuration."""
    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=2,
        seed=100,
        additional_default_tools=[],
    )
    agent.init_model()
    log_system_usage("After Load")
    print("✅ Agent Ready")
    return agent

def analyze_complete_document(content: str, filename: str, agent: TxAgent) -> str:
    """Analyze complete document with strict token management"""
    chunks = split_content_by_tokens(content)
    analysis_results = []
    
    for i, chunk in enumerate(chunks):
        try:
            # Ultra-minimal prompt to maximize content space
            base_prompt = "Analyze for:\n1. Critical\n2. Missed DX\n3. Med issues\n4. Gaps\n5. Follow-up\n\nContent:\n"
            
            # Calculate available space for content
            prompt_tokens = count_tokens(base_prompt)
            max_content_tokens = MAX_MODEL_LEN - prompt_tokens - 100  # Response buffer
            
            # Ensure chunk fits
            chunk_tokens = count_tokens(chunk)
            if chunk_tokens > max_content_tokens:
                # Find last paragraph that fits
                adjusted_chunk = ""
                tokens_used = 0
                paragraphs = re.split(r"\n\s*\n", chunk)
                
                for para in paragraphs:
                    para_tokens = count_tokens(para)
                    if tokens_used + para_tokens <= max_content_tokens:
                        adjusted_chunk += "\n\n" + para
                        tokens_used += para_tokens
                    else:
                        break
                
                if not adjusted_chunk:
                    # If even one paragraph is too big, split sentences
                    sentences = re.split(r'(?<=[.!?])\s+', chunk)
                    for sent in sentences:
                        sent_tokens = count_tokens(sent)
                        if tokens_used + sent_tokens <= max_content_tokens:
                            adjusted_chunk += " " + sent
                            tokens_used += sent_tokens
                        else:
                            break
                
                chunk = adjusted_chunk.strip()
            
            prompt = base_prompt + chunk
            
            response = ""
            for output in agent.run_gradio_chat(
                message=prompt,
                history=[],
                temperature=0.1,
                max_new_tokens=300,  # Keep responses very concise
                max_token=MAX_MODEL_LEN,
                call_agent=False,
                conversation=[],
            ):
                if output:
                    if isinstance(output, list):
                        for m in output:
                            if hasattr(m, 'content'):
                                response += clean_response(m.content)
                    elif isinstance(output, str):
                        response += clean_response(output)
            
            if response:
                analysis_results.append(response)
        except Exception as e:
            print(f"Error processing chunk {i}: {str(e)}")
            continue
    
    return format_final_report(analysis_results, filename)

def create_ui(agent):
    """Create the Gradio interface."""
    with gr.Blocks(theme=gr.themes.Soft(), title="Clinical Oversight Assistant") as demo:
        gr.Markdown("""
        <h1 style='text-align: center;'>🩺 Comprehensive Clinical Oversight Assistant</h1>
        <p style='text-align: center;'>Analyze complete medical records for potential oversights</p>
        """)
        
        with gr.Row():
            with gr.Column(scale=3):
                file_upload = gr.File(
                    file_types=[".pdf", ".csv", ".xls", ".xlsx"], 
                    file_count="multiple",
                    label="Upload Medical Records"
                )
                msg_input = gr.Textbox(
                    placeholder="Optional: Add specific focus areas or questions...",
                    label="Analysis Focus"
                )
                with gr.Row():
                    send_btn = gr.Button("Analyze Complete Documents", variant="primary")
                    clear_btn = gr.Button("Clear")
                status = gr.Textbox(label="Status", interactive=False)
            
            with gr.Column(scale=7):
                report_output = gr.Textbox(
                    label="Clinical Oversight Report",
                    lines=20,
                    max_lines=50,
                    interactive=False
                )
                download_output = gr.File(
                    label="Download Full Report",
                    visible=False
                )
        
        def analyze(files: List, message: str):
            """Process files and generate analysis."""
            if not files:
                yield "", None, "⚠️ Please upload at least one file to analyze."
                return
            
            yield "", None, "⏳ Processing documents (this may take several minutes for large files)..."
            
            file_contents = []
            filenames = []
            total_tokens = 0
            
            with ThreadPoolExecutor(max_workers=4) as executor:
                futures = []
                for f in files:
                    futures.append(executor.submit(
                        convert_file_to_json, 
                        f.name, 
                        f.name.split(".")[-1].lower()
                    ))
                    filenames.append(os.path.basename(f.name))
                
                results = []
                for future in as_completed(futures):
                    result = sanitize_utf8(future.result())
                    results.append(result)
                    try:
                        data = json.loads(result)
                        if "total_tokens" in data:
                            total_tokens += data["total_tokens"]
                    except:
                        pass
                
                file_contents = results
            
            combined_filename = " + ".join(filenames)
            combined_content = "\n".join([
                json.loads(fc).get("content", "") if "content" in json.loads(fc) 
                else str(json.loads(fc).get("rows", "")) 
                for fc in file_contents
            ])
            
            yield "", None, f"🔍 Analyzing content ({total_tokens//1000}k tokens)..."
            
            try:
                full_report = analyze_complete_document(
                    combined_content, 
                    combined_filename, 
                    agent
                )
                
                file_hash_value = hashlib.md5(combined_content.encode()).hexdigest()
                report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt")
                with open(report_path, "w", encoding="utf-8") as f:
                    f.write(full_report)
                
                yield full_report, report_path if os.path.exists(report_path) else None, "✅ Analysis complete!"
            
            except Exception as e:
                error_msg = f"❌ Error during analysis: {str(e)}"
                print(error_msg)
                yield "", None, error_msg
        
        send_btn.click(
            fn=analyze,
            inputs=[file_upload, msg_input],
            outputs=[report_output, download_output, status],
            api_name="analyze"
        )
        
        clear_btn.click(
            fn=lambda: ("", None, ""),
            inputs=None,
            outputs=[report_output, download_output, status]
        )
    
    return demo

if __name__ == "__main__":
    print("🚀 Launching app...")
    try:
        import tiktoken
    except ImportError:
        print("Installing tiktoken...")
        subprocess.run([sys.executable, "-m", "pip", "install", "tiktoken"])
    
    agent = init_agent()
    demo = create_ui(agent)
    demo.queue(
        api_open=False,
        max_size=20
    ).launch(
        server_name="0.0.0.0",
        server_port=7860,
        show_error=True,
        allowed_paths=[report_dir],
        share=False
    )