Update app.py

app.py

@@ -1,12 +1,10 @@
-#import sys, os, json, gradio as gr, pandas as pd, pdfplumber, hashlib, shutil, re, time
+# ✅ Fully optimized app.py for Hugging Face Space with persistent 150GB storage
+
+import sys, os, json, gradio as gr, pandas as pd, pdfplumber, hashlib, shutil, re, time
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from threading import Thread
 
-# Setup
-current_dir = os.path.dirname(os.path.abspath(__file__))
-src_path = os.path.join(current_dir, "src")
-sys.path.insert(0, src_path)
-
+# Use /data for persistent HF storage
 base_dir = "/data"
 model_cache_dir = os.path.join(base_dir, "txagent_models")
 tool_cache_dir = os.path.join(base_dir, "tool_cache")
@@ -17,7 +15,7 @@ vllm_cache_dir = os.path.join(base_dir, "vllm_cache")
 for d in [model_cache_dir, tool_cache_dir, file_cache_dir, report_dir, vllm_cache_dir]:
     os.makedirs(d, exist_ok=True)
 
-# Hugging Face & Transformers cache
+# Set persistent HF + VLLM cache
 os.environ.update({
     "HF_HOME": model_cache_dir,
     "TRANSFORMERS_CACHE": model_cache_dir,
@@ -26,13 +24,21 @@ os.environ.update({
     "CUDA_LAUNCH_BLOCKING": "1"
 })
 
-from txagent.txagent import TxAgent
+# Force local loading only
+LOCAL_TXAGENT_PATH = os.path.join(model_cache_dir, "mims-harvard", "TxAgent-T1-Llama-3.1-8B")
+LOCAL_RAG_PATH = os.path.join(model_cache_dir, "mims-harvard", "ToolRAG-T1-GTE-Qwen2-1.5B")
 
-MEDICAL_KEYWORDS = {'diagnosis', 'assessment', 'plan', 'results', 'medications',
-                    'allergies', 'summary', 'impression', 'findings', 'recommendations'}
+# Manual download using snapshot_download (only if needed)
+# from huggingface_hub import snapshot_download
+# snapshot_download("mims-harvard/TxAgent-T1-Llama-3.1-8B", local_dir=LOCAL_TXAGENT_PATH, local_dir_use_symlinks=False)
+# snapshot_download("mims-harvard/ToolRAG-T1-GTE-Qwen2-1.5B", local_dir=LOCAL_RAG_PATH, local_dir_use_symlinks=False)
+
+sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), "src")))
+from txagent.txagent import TxAgent
 
-def sanitize_utf8(text): return text.encode("utf-8", "ignore").decode("utf-8")
 def file_hash(path): return hashlib.md5(open(path, "rb").read()).hexdigest()
+def sanitize_utf8(text): return text.encode("utf-8", "ignore").decode("utf-8")
+MEDICAL_KEYWORDS = {"diagnosis", "assessment", "plan", "results", "medications", "summary", "findings"}
 
 def extract_priority_pages(file_path, max_pages=20):
     try:
@@ -42,7 +48,7 @@ def extract_priority_pages(file_path, max_pages=20):
                 pages.append(f"=== Page {i+1} ===\n{(page.extract_text() or '').strip()}")
             for i, page in enumerate(pdf.pages[3:max_pages], start=4):
                 text = page.extract_text() or ""
-                if any(re.search(rf'\b{kw}\b', text.lower()) for kw in MEDICAL_KEYWORDS):
+                if any(re.search(rf'\\b{kw}\\b', text.lower()) for kw in MEDICAL_KEYWORDS):
                     pages.append(f"=== Page {i} ===\n{text.strip()}")
             return "\n\n".join(pages)
     except Exception as e:
@@ -59,43 +65,36 @@ def convert_file_to_json(file_path, file_type):
             result = json.dumps({"filename": os.path.basename(file_path), "content": text, "status": "initial"})
             Thread(target=full_pdf_processing, args=(file_path, h)).start()
         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")
-            result = json.dumps({"filename": os.path.basename(file_path), "rows": df.fillna("").astype(str).values.tolist()})
+            df = pd.read_csv(file_path, encoding_errors="replace", header=None, dtype=str)
+            result = json.dumps({"filename": os.path.basename(file_path), "rows": df.fillna('').astype(str).values.tolist()})
         elif file_type in ["xls", "xlsx"]:
-            try:
-                df = pd.read_excel(file_path, engine="openpyxl", header=None, dtype=str)
-            except:
-                df = pd.read_excel(file_path, engine="xlrd", header=None, dtype=str)
-            result = json.dumps({"filename": os.path.basename(file_path), "rows": df.fillna("").astype(str).values.tolist()})
+            df = pd.read_excel(file_path, engine="openpyxl", header=None, dtype=str)
+            result = json.dumps({"filename": os.path.basename(file_path), "rows": df.fillna('').astype(str).values.tolist()})
         else:
             return 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)}"})
+        return json.dumps({"error": str(e)})
 
-def full_pdf_processing(file_path, file_hash_value):
+def full_pdf_processing(file_path, h):
     try:
-        cache_path = os.path.join(file_cache_dir, f"{file_hash_value}_full.json")
+        cache_path = os.path.join(file_cache_dir, f"{h}_full.json")
         if os.path.exists(cache_path): return
         with pdfplumber.open(file_path) as pdf:
-            full_text = "\n".join([f"=== Page {i+1} ===\n{(page.extract_text() or '').strip()}" for i, page in enumerate(pdf.pages)])
-        result = json.dumps({"filename": os.path.basename(file_path), "content": full_text, "status": "complete"})
-        with open(cache_path, "w", encoding="utf-8") as f: f.write(result)
-        with open(os.path.join(report_dir, f"{file_hash_value}_report.txt"), "w", encoding="utf-8") as out: out.write(full_text)
-    except Exception as e:
-        print("PDF processing error:", e)
+            full_text = "\n".join([f"=== Page {i+1} ===\n{(p.extract_text() or '').strip()}" for i, p in enumerate(pdf.pages)])
+        with open(cache_path, "w", encoding="utf-8") as f: f.write(json.dumps({"content": full_text}))
+    except: pass
 
 def init_agent():
-    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)
+        shutil.copy(os.path.abspath("data/new_tool.json"), 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",
+        model_name=LOCAL_TXAGENT_PATH,
+        rag_model_name=LOCAL_RAG_PATH,
         tool_files_dict={"new_tool": target_tool_path},
         force_finish=True,
         enable_checker=True,
@@ -105,82 +104,59 @@ def init_agent():
     agent.init_model()
     return agent
 
-# Lazy load agent only on first use
+# Lazy load
 agent_container = {"agent": None}
 def get_agent():
     if agent_container["agent"] is None:
         agent_container["agent"] = init_agent()
     return agent_container["agent"]
 
-def create_ui(get_agent_func):
+def create_ui():
     with gr.Blocks(theme=gr.themes.Soft()) as demo:
-        gr.Markdown("<h1 style='text-align: center;'>🩺 Clinical Oversight Assistant</h1><h3 style='text-align: center;'>Identify potential oversights in patient care</h3>")
-
-        chatbot = gr.Chatbot(label="Analysis", height=600, type="messages")
+        gr.Markdown("""<h1 style='text-align:center;'>🩺 Clinical Oversight Assistant</h1>""")
+        chatbot = gr.Chatbot(label="Analysis", height=600)
+        msg_input = gr.Textbox(placeholder="Ask a question about the patient...")
         file_upload = gr.File(file_types=[".pdf", ".csv", ".xls", ".xlsx"], file_count="multiple")
-        msg_input = gr.Textbox(placeholder="Ask about potential oversights...")
         send_btn = gr.Button("Analyze", variant="primary")
         state = gr.State([])
-        download_output = gr.File(label="Download Report")
 
         def analyze(message, history, conversation, files):
             try:
-                extracted_data, file_hash_value = "", ""
+                extracted, hval = "", ""
                 if files:
-                    with ThreadPoolExecutor(max_workers=4) as pool:
+                    with ThreadPoolExecutor(max_workers=3) as pool:
                         futures = [pool.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower()) for f in files]
-                        extracted_data = "\n".join([sanitize_utf8(f.result()) for f in as_completed(futures)])
-                        file_hash_value = file_hash(files[0].name)
-
-                prompt = f"""Review these medical records and identify EXACTLY what might have been missed:
-1. List potential missed diagnoses
-2. Flag any medication conflicts
-3. Note incomplete assessments
-4. Highlight abnormal results needing follow-up
+                        extracted = "\n".join([sanitize_utf8(f.result()) for f in as_completed(futures)])
+                        hval = file_hash(files[0].name)
 
-Medical Records:\n{extracted_data[:15000]}
-
-### Potential Oversights:\n"""
+                prompt = f"""Review these medical records and identify exactly what might have been missed:
+1. Missed diagnoses
+2. Medication conflicts
+3. Incomplete assessments
+4. Abnormal results needing follow-up
 
+Medical Records:\n{extracted[:15000]}
+"""
                 final_response = ""
-                for chunk in get_agent_func().run_gradio_chat(
-                    message=prompt,
-                    history=[],
-                    temperature=0.2,
-                    max_new_tokens=1024,
-                    max_token=4096,
-                    call_agent=False,
-                    conversation=conversation
-                ):
-                    if isinstance(chunk, str):
-                        final_response += chunk
-                    elif isinstance(chunk, list):
-                        final_response += "".join([c.content for c in chunk if hasattr(c, "content")])
-
+                for chunk in get_agent().run_gradio_chat(prompt, history=[], temperature=0.2, max_new_tokens=1024, max_token=4096, call_agent=False, conversation=conversation):
+                    if isinstance(chunk, str): final_response += chunk
+                    elif isinstance(chunk, list): final_response += "".join([c.content for c in chunk if hasattr(c, 'content')])
                 cleaned = final_response.replace("[TOOL_CALLS]", "").strip()
-                if not cleaned:
-                    cleaned = "No oversights found. Consider further review."
-
                 updated_history = history + [{"role": "user", "content": message}, {"role": "assistant", "content": cleaned}]
-
-                report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") if file_hash_value and os.path.exists(os.path.join(report_dir, f"{file_hash_value}_report.txt")) else None
-                yield updated_history, report_path
+                return updated_history, None
             except Exception as e:
-                updated_history = history + [{"role": "user", "content": message}, {"role": "assistant", "content": f"❌ Error: {str(e)}"}]
-                yield updated_history, None
-
-        send_btn.click(analyze, inputs=[msg_input, chatbot, state, file_upload], outputs=[chatbot, download_output])
-        msg_input.submit(analyze, inputs=[msg_input, chatbot, state, file_upload], outputs=[chatbot, download_output])
+                return history + [{"role": "user", "content": message}, {"role": "assistant", "content": f"❌ Error: {str(e)}"}], None
 
+        send_btn.click(analyze, inputs=[msg_input, chatbot, state, file_upload], outputs=[chatbot, gr.File()])
+        msg_input.submit(analyze, inputs=[msg_input, chatbot, state, file_upload], outputs=[chatbot, gr.File()])
     return demo
 
 if __name__ == "__main__":
-    print("Launching interface...")
-    ui = create_ui(get_agent)
+    ui = create_ui()
     ui.queue(api_open=False).launch(
         server_name="0.0.0.0",
         server_port=7860,
         show_error=True,
         allowed_paths=["/data/reports"],
         share=False
-    )
+    )