Update app.py

app.py

@@ -11,10 +11,16 @@ import shutil
 import re
 import psutil
 import subprocess
-import multiprocessing
-from functools import partial
+import logging
+import torch
+import gc
+from diskcache import Cache
 import time
 
+# Configure logging
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
 # Persistent directory
 persistent_dir = "/data/hf_cache"
 os.makedirs(persistent_dir, exist_ok=True)
@@ -40,6 +46,9 @@ 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")
 
@@ -47,58 +56,47 @@ def file_hash(path: str) -> str:
     with open(path, "rb") as f:
         return hashlib.md5(f.read()).hexdigest()
 
-def extract_page_range(file_path: str, start_page: int, end_page: int) -> str:
-    """Extract text from a range of PDF pages."""
-    try:
-        text_chunks = []
-        with pdfplumber.open(file_path) as pdf:
-            for page in pdf.pages[start_page:end_page]:
-                page_text = page.extract_text() or ""
-                text_chunks.append(f"=== Page {start_page + pdf.pages.index(page) + 1} ===\n{page_text.strip()}")
-        return "\n\n".join(text_chunks)
-    except Exception:
-        return ""
-
 def extract_all_pages(file_path: str, progress_callback=None) -> str:
-    """Extract text from all pages of a PDF using parallel processing."""
     try:
         with pdfplumber.open(file_path) as pdf:
             total_pages = len(pdf.pages)
-        
-        if total_pages == 0:
-            return ""
-        
-        # Use 6 processes (adjust based on CPU cores)
-        num_processes = min(6, multiprocessing.cpu_count())
-        pages_per_process = max(1, total_pages // num_processes)
-        
-        # Create page ranges for parallel processing
-        ranges = [(i * pages_per_process, min((i + 1) * pages_per_process, total_pages))
-                  for i in range(num_processes)]
-        if ranges[-1][1] != total_pages:
-            ranges[-1] = (ranges[-1][0], total_pages)
-        
-        # Process page ranges in parallel
-        with multiprocessing.Pool(processes=num_processes) as pool:
-            extract_func = partial(extract_page_range, file_path)
+            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
+        processed_pages = 0
+
+        def extract_batch(start: int, end: int) -> List[tuple]:
             results = []
-            for idx, result in enumerate(pool.starmap(extract_func, ranges)):
-                results.append(result)
+            with pdfplumber.open(file_path) as pdf:
+                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=6) 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:
-                    processed_pages = min((idx + 1) * pages_per_process, total_pages)
-                    progress_callback(processed_pages, total_pages)
-        
-        return "\n\n".join(filter(None, results))
+                    progress_callback(min(processed_pages, total_pages), total_pages)
+
+        return "\n\n".join(filter(None, text_chunks))
     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:
-        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()
+        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)
@@ -117,46 +115,41 @@ def convert_file_to_json(file_path: str, file_type: str, progress_callback=None)
             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)
+
+        cache[cache_key] = 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()
-        print(f"[{tag}] CPU: {cpu}% | RAM: {mem.used // (1024**2)}MB / {mem.total // (1024**2)}MB")
+        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(", ")
-            print(f"[{tag}] GPU: {used}MB / {total}MB | Utilization: {util}%")
+            logger.info("[%s] GPU: %sMB / %sMB | Utilization: %s%%", tag, used, total, util)
     except Exception as e:
-        print(f"[{tag}] GPU/CPU monitor failed: {e}")
+        logger.error("[%s] GPU/CPU monitor failed: %s", tag, e)
 
 def clean_response(text: str) -> str:
-    """Clean TxAgent response to group findings under tool-derived headings."""
     text = sanitize_utf8(text)
-    # Remove tool call artifacts, None, and reasoning
     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)
-    # Remove extra whitespace and non-markdown content
     text = re.sub(r"\n{3,}", "\n\n", text)
-    text = re.sub(r"[^\n#\-\*\w\s\.\,\:\(\)]+", "", text)  # Keep markdown-relevant characters
-    
-    # Define tool-to-heading mapping
+    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",
-        # Add other tools from new_tool.json if applicable
     }
-    
-    # Parse sections and findings
+
     sections = {}
     current_section = None
     current_tool = None
@@ -165,22 +158,18 @@ def clean_response(text: str) -> str:
         line = line.strip()
         if not line:
             continue
-        # Detect tool tag
         tool_match = re.match(r"\[TOOL:\s*(\w+)\]", line)
         if tool_match:
             current_tool = tool_match.group(1)
             continue
-        # Detect section heading
         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
-        # Detect finding
         finding_match = re.match(r"-\s*.+", line)
         if finding_match and current_section and not re.match(r"-\s*No issues identified", line):
-            # Assign to tool-derived heading if tool is specified
             if current_tool and current_tool in tool_to_heading:
                 heading = tool_to_heading[current_tool]
                 if heading not in sections:
@@ -188,20 +177,17 @@ def clean_response(text: str) -> str:
                 sections[heading].append(line)
             else:
                 sections[current_section].append(line)
-    
-    # Combine non-empty sections
+
     cleaned = []
     for heading, findings in sections.items():
-        if findings:  # Only include sections with findings
+        if findings:
             cleaned.append(f"### {heading}\n" + "\n".join(findings))
-    
+
     text = "\n\n".join(cleaned).strip()
-    if not text:
-        text = ""  # Return empty string if no valid findings
-    return text
+    return text if text else ""
 
 def init_agent():
-    print("🔁 Initializing model...")
+    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")
@@ -213,14 +199,14 @@ def init_agent():
         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,
+        enable_checker=False,  # Disabled for speed
         step_rag_num=4,
         seed=100,
         additional_default_tools=[],
     )
     agent.init_model()
     log_system_usage("After Load")
-    print("✅ Agent Ready")
+    logger.info("Agent Ready")
     return agent
 
 def create_ui(agent):
@@ -231,24 +217,24 @@ def create_ui(agent):
         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()
 
-        def analyze(message: str, history: List[dict], files: List):
+        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})
-            history.append({"role": "assistant", "content": "⏳ Extracting text from files..."})
             yield history, None
 
             extracted = ""
             file_hash_value = ""
             if files:
-                # Progress callback for extraction
-                total_pages = 0
-                processed_pages = 0
                 def update_extraction_progress(current, total):
-                    nonlocal processed_pages, total_pages
-                    processed_pages = current
-                    total_pages = total
-                    animation = ["🌀", "🔄", "⚙️", "🔃"][(int(time.time() * 2) % 4)]
-                    history[-1] = {"role": "assistant", "content": f"Extracting text... {animation} Page {processed_pages}/{total_pages}"}
+                    progress(current / total, desc=f"Extracting text... Page {current}/{total}")
                     return history, None
 
                 with ThreadPoolExecutor(max_workers=6) as executor:
@@ -257,92 +243,56 @@ def create_ui(agent):
                     extracted = "\n".join(results)
                     file_hash_value = file_hash(files[0].name) if files else ""
 
-            history.pop()  # Remove extraction message
-            history.append({"role": "assistant", "content": "✅ Text extraction complete."})
-            yield history, None
+                history.append({"role": "assistant", "content": "✅ Text extraction complete."})
+                yield history, None
 
-            # Split extracted text into chunks of ~6,000 characters
             chunk_size = 6000
             chunks = [extracted[i:i + chunk_size] for i in range(0, len(extracted), chunk_size)]
             combined_response = ""
-
-            prompt_template = """
-You are a medical analysis assistant. Analyze the following patient record excerpt for clinical oversights and provide a concise, evidence-based summary in markdown format. Group findings under appropriate headings based on the tool used (e.g., drug-related findings under 'Drugs'). For each finding, include:
-- Clinical context (why the issue was missed or relevant details from the record).
-- Potential risks if unaddressed (e.g., disease progression, adverse events).
-- Actionable recommendations (e.g., tests, referrals, medication adjustments).
-Output ONLY the markdown-formatted findings, with bullet points under each heading. Precede each finding with a tool tag (e.g., [TOOL: get_abuse_info_by_drug_name]) to indicate the tool used. Do NOT include reasoning, tool calls, or intermediate steps. If no issues are found for a tool or category, state "No issues identified" for that section. Ensure the output is specific to the provided text and avoids generic responses.
-Example Output:
-### Drugs
-[TOOL: get_abuse_info_by_drug_name]
-- Opioid use disorder not addressed. Missed due to lack of screening. Risks: overdose. Recommend: addiction specialist referral.
-### Missed Diagnoses
-- Elevated BP noted without diagnosis. Missed due to inconsistent visits. Risks: stroke. Recommend: BP monitoring, antihypertensives.
-### Incomplete Assessments
-- Chest pain not evaluated. Time constraints likely cause. Risks: cardiac issues. Recommend: ECG, stress test.
-### Urgent Follow-up
-- Abnormal creatinine not addressed. Delayed lab review. Risks: renal failure. Recommend: nephrology referral.
-Patient Record Excerpt (Chunk {0} of {1}):
-{chunk}
-"""
+            batch_size = 2
 
             try:
-                # Process each chunk and stream results in real-time
-                for chunk_idx, chunk in enumerate(chunks, 1):
-                    # Update UI with chunk progress
-                    animation = ["🔍", "📊", "🧠", "🔎"][(int(time.time() * 2) % 4)]
-                    history.append({"role": "assistant", "content": f"Analyzing records... {animation} Chunk {chunk_idx}/{len(chunks)}"})
-                    yield history, None
-
-                    prompt = prompt_template.format(chunk_idx, len(chunks), chunk=chunk[:4000])  # Truncate to avoid token limits
-                    chunk_response = ""
-                    for chunk_output in agent.run_gradio_chat(
-                        message=prompt,
-                        history=[],
-                        temperature=0.2,
-                        max_new_tokens=1024,
-                        max_token=4096,
-                        call_agent=False,
-                        conversation=[],
-                    ):
-                        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)
+                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"
-                                        # Update UI with partial response
-                                        if history[-1]["content"].startswith("Analyzing"):
-                                            history[-1] = {"role": "assistant", "content": f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response.strip()}"}
-                                        else:
-                                            history[-1]["content"] = f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response.strip()}"
-                                        yield history, None
-                        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"
-                                # Update UI with partial response
-                                if history[-1]["content"].startswith("Analyzing"):
-                                    history[-1] = {"role": "assistant", "content": f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response.strip()}"}
-                                else:
-                                    history[-1]["content"] = f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response.strip()}"
-                                yield history, None
-
-                    # Append completed chunk response to combined response
-                    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"
-
-                # Finalize UI with complete response
+                            batch_responses.append(chunk_response)
+                            torch.cuda.empty_cache()
+                            gc.collect()
+
+                    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
+
                 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."})
 
-                # Generate report file
                 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:
@@ -350,7 +300,7 @@ Patient Record Excerpt (Chunk {0} of {1}):
                 yield history, report_path if report_path and os.path.exists(report_path) else None
 
             except Exception as e:
-                print("🚨 ERROR:", e)
+                logger.error("Analysis error: %s", e)
                 history.append({"role": "assistant", "content": f"❌ Error occurred: {str(e)}"})
                 yield history, None
 
@@ -359,13 +309,17 @@ Patient Record Excerpt (Chunk {0} of {1}):
     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
-    )
+    try:
+        logger.info("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
+        )
+    finally:
+        if torch.distributed.is_initialized():
+            torch.distributed.destroy_process_group()