Spaces:
Running
on
Zero
Running
on
Zero
Thanush
commited on
Commit
Β·
0b85ef5
1
Parent(s):
10736b1
Enhance medical consultation app with LangChain memory management and improved patient context tracking
Browse files
app.py
CHANGED
@@ -2,9 +2,13 @@ import gradio as gr
|
|
2 |
import torch
|
3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
4 |
import spaces
|
|
|
|
|
|
|
|
|
5 |
|
6 |
# Model configuration - Using correct Me-LLaMA model identifier
|
7 |
-
ME_LLAMA_MODEL = "clinicalnlplab/me-llama-13b"
|
8 |
|
9 |
# System prompts for different phases
|
10 |
CONSULTATION_PROMPT = """You are a professional virtual doctor. Your goal is to collect detailed information about the user's health condition, symptoms, medical history, medications, lifestyle, and other relevant data.
|
@@ -26,7 +30,8 @@ MEDICINE_PROMPT = """You are a specialized medical assistant. Based on the patie
|
|
26 |
|
27 |
Be concise, practical, and focus only on general symptom relief. Do not diagnose. Include a disclaimer that you are not a licensed medical professional.
|
28 |
|
29 |
-
Patient information: {patient_info}
|
|
|
30 |
|
31 |
# Global variables
|
32 |
me_llama_model = None
|
@@ -34,13 +39,121 @@ me_llama_tokenizer = None
|
|
34 |
conversation_turns = 0
|
35 |
patient_data = []
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
def build_me_llama_prompt(system_prompt, history, user_input):
|
38 |
-
"""Format the conversation for Me-LLaMA chat model."""
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
# Use standard Llama-2 chat format since Me-LLaMA is based on Llama-2
|
40 |
-
prompt = f"<s>[INST] <<SYS>>\n{
|
41 |
|
42 |
-
# Add
|
43 |
-
|
|
|
44 |
prompt += f"{user_msg} [/INST] {assistant_msg} </s><s>[INST] "
|
45 |
|
46 |
# Add the current user input
|
@@ -80,12 +193,12 @@ def load_model_if_needed():
|
|
80 |
print("Fallback model loaded successfully!")
|
81 |
|
82 |
@spaces.GPU
|
83 |
-
def generate_medicine_suggestions(patient_info):
|
84 |
-
"""Use Me-LLaMA to generate medicine and remedy suggestions."""
|
85 |
load_model_if_needed()
|
86 |
|
87 |
-
# Create a
|
88 |
-
prompt = f"<s>[INST] {MEDICINE_PROMPT.format(patient_info=patient_info)} [/INST] "
|
89 |
|
90 |
inputs = me_llama_tokenizer(prompt, return_tensors="pt")
|
91 |
|
@@ -109,7 +222,7 @@ def generate_medicine_suggestions(patient_info):
|
|
109 |
|
110 |
@spaces.GPU
|
111 |
def generate_response(message, history):
|
112 |
-
"""Generate response using
|
113 |
global conversation_turns, patient_data
|
114 |
|
115 |
try:
|
@@ -119,12 +232,12 @@ def generate_response(message, history):
|
|
119 |
# Track conversation turns
|
120 |
conversation_turns += 1
|
121 |
|
122 |
-
# Store patient data
|
123 |
patient_data.append(message)
|
124 |
|
125 |
-
# Phase 1-3: Information gathering
|
126 |
if conversation_turns < 4:
|
127 |
-
# Build consultation prompt
|
128 |
prompt = build_me_llama_prompt(CONSULTATION_PROMPT, history, message)
|
129 |
|
130 |
inputs = me_llama_tokenizer(prompt, return_tensors="pt")
|
@@ -149,13 +262,20 @@ def generate_response(message, history):
|
|
149 |
full_response = me_llama_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
150 |
response = full_response.split('[/INST]')[-1].split('</s>')[0].strip()
|
151 |
|
|
|
|
|
|
|
152 |
return response
|
153 |
|
154 |
-
# Phase 4+: Summary and medicine suggestions
|
155 |
else:
|
156 |
-
#
|
|
|
|
|
|
|
|
|
157 |
summary_prompt = build_me_llama_prompt(
|
158 |
-
CONSULTATION_PROMPT + "\n\nNow
|
159 |
history,
|
160 |
message
|
161 |
)
|
@@ -180,34 +300,81 @@ def generate_response(message, history):
|
|
180 |
summary_response = me_llama_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
181 |
summary = summary_response.split('[/INST]')[-1].split('</s>')[0].strip()
|
182 |
|
183 |
-
#
|
184 |
-
full_patient_info = "
|
185 |
-
medicine_suggestions = generate_medicine_suggestions(full_patient_info)
|
186 |
|
187 |
# Combine both responses
|
188 |
final_response = (
|
189 |
-
f"**MEDICAL SUMMARY:**\n{summary}\n\n"
|
190 |
f"**MEDICATION AND HOME CARE SUGGESTIONS:**\n{medicine_suggestions}\n\n"
|
|
|
191 |
f"**DISCLAIMER:** This is AI-generated advice for informational purposes only. Please consult a licensed healthcare provider for proper medical diagnosis and treatment."
|
192 |
)
|
193 |
|
|
|
|
|
|
|
194 |
return final_response
|
195 |
|
196 |
except Exception as e:
|
197 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
|
199 |
-
# Create the Gradio interface
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
211 |
|
212 |
if __name__ == "__main__":
|
213 |
-
demo.launch()
|
|
|
2 |
import torch
|
3 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
4 |
import spaces
|
5 |
+
from langchain.memory import ConversationBufferWindowMemory
|
6 |
+
from langchain.schema import HumanMessage, AIMessage
|
7 |
+
import json
|
8 |
+
from datetime import datetime
|
9 |
|
10 |
# Model configuration - Using correct Me-LLaMA model identifier
|
11 |
+
ME_LLAMA_MODEL = "clinicalnlplab/me-llama-13b"
|
12 |
|
13 |
# System prompts for different phases
|
14 |
CONSULTATION_PROMPT = """You are a professional virtual doctor. Your goal is to collect detailed information about the user's health condition, symptoms, medical history, medications, lifestyle, and other relevant data.
|
|
|
30 |
|
31 |
Be concise, practical, and focus only on general symptom relief. Do not diagnose. Include a disclaimer that you are not a licensed medical professional.
|
32 |
|
33 |
+
Patient information: {patient_info}
|
34 |
+
Previous conversation context: {memory_context}"""
|
35 |
|
36 |
# Global variables
|
37 |
me_llama_model = None
|
|
|
39 |
conversation_turns = 0
|
40 |
patient_data = []
|
41 |
|
42 |
+
# LangChain Memory Configuration
|
43 |
+
class MedicalMemoryManager:
|
44 |
+
def __init__(self, k=10): # Keep last 10 conversation turns
|
45 |
+
self.conversation_memory = ConversationBufferWindowMemory(k=k, return_messages=True)
|
46 |
+
self.patient_context = {
|
47 |
+
"symptoms": [],
|
48 |
+
"medical_history": [],
|
49 |
+
"medications": [],
|
50 |
+
"allergies": [],
|
51 |
+
"lifestyle_factors": [],
|
52 |
+
"timeline": [],
|
53 |
+
"severity_scores": {},
|
54 |
+
"session_start": datetime.now().isoformat()
|
55 |
+
}
|
56 |
+
|
57 |
+
def add_interaction(self, human_input, ai_response):
|
58 |
+
"""Add human-AI interaction to memory"""
|
59 |
+
self.conversation_memory.chat_memory.add_user_message(human_input)
|
60 |
+
self.conversation_memory.chat_memory.add_ai_message(ai_response)
|
61 |
+
|
62 |
+
# Extract and categorize medical information
|
63 |
+
self._extract_medical_info(human_input)
|
64 |
+
|
65 |
+
def _extract_medical_info(self, user_input):
|
66 |
+
"""Extract medical information from user input and categorize it"""
|
67 |
+
user_lower = user_input.lower()
|
68 |
+
|
69 |
+
# Extract symptoms (simple keyword matching)
|
70 |
+
symptom_keywords = ["pain", "ache", "hurt", "sore", "cough", "fever", "nausea",
|
71 |
+
"headache", "dizzy", "tired", "fatigue", "vomit", "swollen",
|
72 |
+
"rash", "itch", "burn", "cramp", "bleed", "shortness of breath"]
|
73 |
+
|
74 |
+
for keyword in symptom_keywords:
|
75 |
+
if keyword in user_lower and keyword not in [s.lower() for s in self.patient_context["symptoms"]]:
|
76 |
+
self.patient_context["symptoms"].append(user_input)
|
77 |
+
break
|
78 |
+
|
79 |
+
# Extract timeline information
|
80 |
+
time_keywords = ["days", "weeks", "months", "hours", "yesterday", "today", "started", "began"]
|
81 |
+
if any(keyword in user_lower for keyword in time_keywords):
|
82 |
+
self.patient_context["timeline"].append(user_input)
|
83 |
+
|
84 |
+
# Extract severity (look for numbers 1-10)
|
85 |
+
import re
|
86 |
+
severity_match = re.search(r'\b([1-9]|10)\b.*(?:pain|severity|scale)', user_lower)
|
87 |
+
if severity_match:
|
88 |
+
self.patient_context["severity_scores"][datetime.now().isoformat()] = severity_match.group(1)
|
89 |
+
|
90 |
+
# Extract medications
|
91 |
+
med_keywords = ["taking", "medication", "medicine", "pills", "prescribed", "drug"]
|
92 |
+
if any(keyword in user_lower for keyword in med_keywords):
|
93 |
+
self.patient_context["medications"].append(user_input)
|
94 |
+
|
95 |
+
# Extract allergies
|
96 |
+
allergy_keywords = ["allergic", "allergy", "allergies", "reaction"]
|
97 |
+
if any(keyword in user_lower for keyword in allergy_keywords):
|
98 |
+
self.patient_context["allergies"].append(user_input)
|
99 |
+
|
100 |
+
def get_memory_context(self):
|
101 |
+
"""Get formatted memory context for the model"""
|
102 |
+
messages = self.conversation_memory.chat_memory.messages
|
103 |
+
context = []
|
104 |
+
|
105 |
+
for msg in messages[-6:]: # Last 6 messages (3 exchanges)
|
106 |
+
if isinstance(msg, HumanMessage):
|
107 |
+
context.append(f"Patient: {msg.content}")
|
108 |
+
elif isinstance(msg, AIMessage):
|
109 |
+
context.append(f"Doctor: {msg.content}")
|
110 |
+
|
111 |
+
return "\n".join(context)
|
112 |
+
|
113 |
+
def get_patient_summary(self):
|
114 |
+
"""Get structured patient information summary"""
|
115 |
+
summary = {
|
116 |
+
"conversation_turns": len(self.conversation_memory.chat_memory.messages) // 2,
|
117 |
+
"session_duration": datetime.now().isoformat(),
|
118 |
+
"key_symptoms": self.patient_context["symptoms"][-3:], # Last 3 symptoms mentioned
|
119 |
+
"timeline_info": self.patient_context["timeline"][-2:], # Last 2 timeline mentions
|
120 |
+
"medications": self.patient_context["medications"],
|
121 |
+
"allergies": self.patient_context["allergies"],
|
122 |
+
"severity_scores": self.patient_context["severity_scores"]
|
123 |
+
}
|
124 |
+
return json.dumps(summary, indent=2)
|
125 |
+
|
126 |
+
def reset_session(self):
|
127 |
+
"""Reset memory for new consultation"""
|
128 |
+
self.conversation_memory.clear()
|
129 |
+
self.patient_context = {
|
130 |
+
"symptoms": [],
|
131 |
+
"medical_history": [],
|
132 |
+
"medications": [],
|
133 |
+
"allergies": [],
|
134 |
+
"lifestyle_factors": [],
|
135 |
+
"timeline": [],
|
136 |
+
"severity_scores": {},
|
137 |
+
"session_start": datetime.now().isoformat()
|
138 |
+
}
|
139 |
+
|
140 |
+
# Initialize memory manager
|
141 |
+
memory_manager = MedicalMemoryManager()
|
142 |
+
|
143 |
def build_me_llama_prompt(system_prompt, history, user_input):
|
144 |
+
"""Format the conversation for Me-LLaMA chat model with memory context."""
|
145 |
+
# Get memory context from LangChain
|
146 |
+
memory_context = memory_manager.get_memory_context()
|
147 |
+
|
148 |
+
# Enhance system prompt with memory context
|
149 |
+
enhanced_system_prompt = f"{system_prompt}\n\nPrevious conversation context:\n{memory_context}"
|
150 |
+
|
151 |
# Use standard Llama-2 chat format since Me-LLaMA is based on Llama-2
|
152 |
+
prompt = f"<s>[INST] <<SYS>>\n{enhanced_system_prompt}\n<</SYS>>\n\n"
|
153 |
|
154 |
+
# Add only recent history to avoid token limit issues
|
155 |
+
recent_history = history[-3:] if len(history) > 3 else history
|
156 |
+
for user_msg, assistant_msg in recent_history:
|
157 |
prompt += f"{user_msg} [/INST] {assistant_msg} </s><s>[INST] "
|
158 |
|
159 |
# Add the current user input
|
|
|
193 |
print("Fallback model loaded successfully!")
|
194 |
|
195 |
@spaces.GPU
|
196 |
+
def generate_medicine_suggestions(patient_info, memory_context):
|
197 |
+
"""Use Me-LLaMA to generate medicine and remedy suggestions with memory context."""
|
198 |
load_model_if_needed()
|
199 |
|
200 |
+
# Create a prompt with both patient info and memory context
|
201 |
+
prompt = f"<s>[INST] {MEDICINE_PROMPT.format(patient_info=patient_info, memory_context=memory_context)} [/INST] "
|
202 |
|
203 |
inputs = me_llama_tokenizer(prompt, return_tensors="pt")
|
204 |
|
|
|
222 |
|
223 |
@spaces.GPU
|
224 |
def generate_response(message, history):
|
225 |
+
"""Generate response using Me-LLaMA with LangChain memory management."""
|
226 |
global conversation_turns, patient_data
|
227 |
|
228 |
try:
|
|
|
232 |
# Track conversation turns
|
233 |
conversation_turns += 1
|
234 |
|
235 |
+
# Store patient data (legacy support)
|
236 |
patient_data.append(message)
|
237 |
|
238 |
+
# Phase 1-3: Information gathering with memory
|
239 |
if conversation_turns < 4:
|
240 |
+
# Build consultation prompt with memory context
|
241 |
prompt = build_me_llama_prompt(CONSULTATION_PROMPT, history, message)
|
242 |
|
243 |
inputs = me_llama_tokenizer(prompt, return_tensors="pt")
|
|
|
262 |
full_response = me_llama_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
263 |
response = full_response.split('[/INST]')[-1].split('</s>')[0].strip()
|
264 |
|
265 |
+
# Add interaction to memory
|
266 |
+
memory_manager.add_interaction(message, response)
|
267 |
+
|
268 |
return response
|
269 |
|
270 |
+
# Phase 4+: Summary and medicine suggestions with full memory context
|
271 |
else:
|
272 |
+
# Get comprehensive patient summary from memory
|
273 |
+
patient_summary = memory_manager.get_patient_summary()
|
274 |
+
memory_context = memory_manager.get_memory_context()
|
275 |
+
|
276 |
+
# First, get summary from consultation with memory context
|
277 |
summary_prompt = build_me_llama_prompt(
|
278 |
+
CONSULTATION_PROMPT + "\n\nNow provide a comprehensive summary based on all the information gathered. Include when professional care may be needed.",
|
279 |
history,
|
280 |
message
|
281 |
)
|
|
|
300 |
summary_response = me_llama_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
301 |
summary = summary_response.split('[/INST]')[-1].split('</s>')[0].strip()
|
302 |
|
303 |
+
# Get medicine suggestions using memory context
|
304 |
+
full_patient_info = f"Patient Summary: {patient_summary}\n\nDetailed Summary: {summary}"
|
305 |
+
medicine_suggestions = generate_medicine_suggestions(full_patient_info, memory_context)
|
306 |
|
307 |
# Combine both responses
|
308 |
final_response = (
|
309 |
+
f"**COMPREHENSIVE MEDICAL SUMMARY:**\n{summary}\n\n"
|
310 |
f"**MEDICATION AND HOME CARE SUGGESTIONS:**\n{medicine_suggestions}\n\n"
|
311 |
+
f"**PATIENT CONTEXT SUMMARY:**\n{patient_summary}\n\n"
|
312 |
f"**DISCLAIMER:** This is AI-generated advice for informational purposes only. Please consult a licensed healthcare provider for proper medical diagnosis and treatment."
|
313 |
)
|
314 |
|
315 |
+
# Add final interaction to memory
|
316 |
+
memory_manager.add_interaction(message, final_response)
|
317 |
+
|
318 |
return final_response
|
319 |
|
320 |
except Exception as e:
|
321 |
+
error_msg = f"I apologize, but I'm experiencing technical difficulties. Please try again. Error: {str(e)}"
|
322 |
+
# Still try to add to memory even on error
|
323 |
+
try:
|
324 |
+
memory_manager.add_interaction(message, error_msg)
|
325 |
+
except:
|
326 |
+
pass
|
327 |
+
return error_msg
|
328 |
+
|
329 |
+
def reset_consultation():
|
330 |
+
"""Reset the consultation and memory for a new patient."""
|
331 |
+
global conversation_turns, patient_data, memory_manager
|
332 |
+
|
333 |
+
conversation_turns = 0
|
334 |
+
patient_data = []
|
335 |
+
memory_manager.reset_session()
|
336 |
+
|
337 |
+
return "New consultation started. Please tell me about your symptoms or health concerns."
|
338 |
|
339 |
+
# Create the Gradio interface with memory reset option
|
340 |
+
with gr.Blocks(theme="soft") as demo:
|
341 |
+
gr.Markdown("# π₯ Complete Medical Assistant - Me-LLaMA 13B with Memory")
|
342 |
+
gr.Markdown("Comprehensive medical consultation powered by Me-LLaMA 13B with LangChain memory management. One model handles both consultation and medicine suggestions with full context awareness.")
|
343 |
+
|
344 |
+
with gr.Row():
|
345 |
+
with gr.Column(scale=4):
|
346 |
+
chatbot = gr.Chatbot(height=500)
|
347 |
+
msg = gr.Textbox(
|
348 |
+
placeholder="Tell me about your symptoms or health concerns...",
|
349 |
+
label="Your Message"
|
350 |
+
)
|
351 |
+
|
352 |
+
with gr.Column(scale=1):
|
353 |
+
reset_btn = gr.Button("π Start New Consultation", variant="secondary")
|
354 |
+
gr.Markdown("**Memory Features:**\n- Tracks symptoms & timeline\n- Remembers medications & allergies\n- Maintains conversation context\n- Provides comprehensive summaries")
|
355 |
+
|
356 |
+
# Examples
|
357 |
+
gr.Examples(
|
358 |
+
examples=[
|
359 |
+
"I have a persistent cough and sore throat for 3 days",
|
360 |
+
"I've been having severe headaches and feel dizzy",
|
361 |
+
"My stomach hurts and I feel nauseous after eating"
|
362 |
+
],
|
363 |
+
inputs=msg
|
364 |
+
)
|
365 |
+
|
366 |
+
# Event handlers
|
367 |
+
def respond(message, chat_history):
|
368 |
+
bot_message = generate_response(message, chat_history)
|
369 |
+
chat_history.append((message, bot_message))
|
370 |
+
return "", chat_history
|
371 |
+
|
372 |
+
def reset_chat():
|
373 |
+
reset_msg = reset_consultation()
|
374 |
+
return [(None, reset_msg)], ""
|
375 |
+
|
376 |
+
msg.submit(respond, [msg, chatbot], [msg, chatbot])
|
377 |
+
reset_btn.click(reset_chat, [], [chatbot, msg])
|
378 |
|
379 |
if __name__ == "__main__":
|
380 |
+
demo.launch()
|