local-llm-2 / app.py
Robin Genolet
feat: use langchain
90d439d
raw
history blame
18.7 kB
import re
import uuid
import pandas as pd
import streamlit as st
import re
import matplotlib.pyplot as plt
import subprocess
import sys
import io
from utils.default_values import get_system_prompt, get_guidelines_dict
from utils.epfl_meditron_utils import get_llm_response
from utils.openai_utils import get_available_engines, get_search_query_type_options
from sklearn.metrics import confusion_matrix, ConfusionMatrixDisplay
from sklearn.metrics import classification_report
DATA_FOLDER = "data/"
POC_VERSION = "0.1.0"
MAX_QUESTIONS = 10
AVAILABLE_LANGUAGES = ["DE", "EN", "FR"]
st.set_page_config(page_title='Medgate Whisper PoC', page_icon='public/medgate.png')
# Azure apparently truncates message if longer than 200, see
MAX_SYSTEM_MESSAGE_TOKENS = 200
def format_question(q):
res = q
# Remove numerical prefixes, if any, e.g. '1. [...]'
if re.match(r'^[0-9].\s', q):
res = res[3:]
# Replace doc reference by doc name
if len(st.session_state["citations"]) > 0:
for source_ref in re.findall(r'\[doc[0-9]+\]', res):
citation_number = int(re.findall(r'[0-9]+', source_ref)[0])
citation_index = citation_number - 1 if citation_number > 0 else 0
citation = st.session_state["citations"][citation_index]
source_title = citation["title"]
res = res.replace(source_ref, '[' + source_title + ']')
return res.strip()
def get_text_from_row(text):
res = str(text)
if res == "nan":
return ""
return res
def get_questions_from_df(df, lang, test_scenario_name):
questions = []
for i, row in df.iterrows():
questions.append({
"question": row[lang + ": Fragen"],
"answer": get_text_from_row(row[test_scenario_name]),
"question_id": uuid.uuid4()
})
return questions
def get_questions(df, lead_symptom, lang, test_scenario_name):
print(str(st.session_state["lead_symptom"]) + " -> " + lead_symptom)
print(str(st.session_state["scenario_name"]) + " -> " + test_scenario_name)
if st.session_state["lead_symptom"] != lead_symptom or st.session_state["scenario_name"] != test_scenario_name:
st.session_state["lead_symptom"] = lead_symptom
st.session_state["scenario_name"] = test_scenario_name
symptom_col_name = st.session_state["language"] + ": Symptome"
df_questions = df[(df[symptom_col_name] == lead_symptom)]
st.session_state["questions"] = get_questions_from_df(df_questions, lang, test_scenario_name)
return st.session_state["questions"]
def display_streamlit_sidebar():
st.sidebar.title("Local LLM PoC " + str(POC_VERSION))
st.sidebar.write('**Parameters**')
form = st.sidebar.form("config_form", clear_on_submit=True)
model_option = form.selectbox("Quickly select a model", ("llama", "meditron"))
model_repo_id = form.text_input(label="Repo", value=model_option)#value=st.session_state["model_repo_id"])
model_filename = form.text_input(label="File name", value=st.session_state["model_filename"])
model_type = form.text_input(label="Model type", value=st.session_state["model_type"])
gpu_layers = form.slider('GPU Layers', min_value=0,
max_value=100, value=st.session_state['gpu_layers'], step=1)
system_prompt = ""
#form.text_area(label='System prompt',
# value=st.session_state["system_prompt"])
temperature = form.slider('Temperature (0 = deterministic, 1 = more freedom)', min_value=0.0,
max_value=1.0, value=st.session_state['temperature'], step=0.1)
top_p = form.slider('top_p (0 = focused, 1 = broader answer range)', min_value=0.0,
max_value=1.0, value=st.session_state['top_p'], step=0.1)
form.write('Best practice is to only modify temperature or top_p, not both')
submitted = form.form_submit_button("Start session")
if submitted and not st.session_state['session_started']:
print('Parameters updated...')
restart_session()
st.session_state['session_started'] = True
st.session_state["model_repo_id"] = model_repo_id
st.session_state["model_filename"] = model_filename
st.session_state["model_type"] = model_type
st.session_state['gpu_layers'] = gpu_layers
st.session_state["questions"] = []
st.session_state["lead_symptom"] = None
st.session_state["scenario_name"] = None
st.session_state["system_prompt"] = system_prompt
st.session_state['session_started'] = True
st.session_state["session_started"] = True
st.session_state["temperature"] = temperature
st.session_state["top_p"] = top_p
st.rerun()
def to_str(text):
res = str(text)
if res == "nan":
return " "
return " " + res
def set_df_prompts(path, sheet_name):
df_prompts = pd.read_excel(path, sheet_name, header=None)
for i in range(3, df_prompts.shape[0]):
df_prompts.iloc[2] += df_prompts.iloc[i].apply(to_str)
df_prompts = df_prompts.T
df_prompts = df_prompts[[0, 1, 2]]
df_prompts[0] = df_prompts[0].astype(str)
df_prompts[1] = df_prompts[1].astype(str)
df_prompts[2] = df_prompts[2].astype(str)
df_prompts.columns = ["Questionnaire", "Used Guideline", "Prompt"]
df_prompts = df_prompts[1:]
st.session_state["df_prompts"] = df_prompts
def handle_nbq_click(c):
question_without_source = re.sub(r'\[.*\]', '', c)
question_without_source = question_without_source.strip()
st.session_state['doctor_question'] = question_without_source
def get_doctor_question_value():
if 'doctor_question' in st.session_state:
return st.session_state['doctor_question']
return ''
def update_chat_history(dr_question, patient_reply):
print("update_chat_history" + str(dr_question) + " - " + str(patient_reply) + '...\n')
if dr_question is not None:
dr_msg = {
"role": "Doctor",
"content": dr_question
}
st.session_state["chat_history_array"].append(dr_msg)
if patient_reply is not None:
patient_msg = {
"role": "Patient",
"content": patient_reply
}
st.session_state["chat_history_array"].append(patient_msg)
return st.session_state["chat_history_array"]
def get_chat_history_string(chat_history):
res = ''
for i in chat_history:
if i["role"] == "Doctor":
res += '**Doctor**: ' + str(i["content"].strip()) + " \n "
elif i["role"] == "Patient":
res += '**Patient**: ' + str(i["content"].strip()) + " \n\n "
else:
raise Exception('Unknown role: ' + str(i["role"]))
return res
def restart_session():
print("Resetting params...")
st.session_state["emg_class_enabled"] = False
st.session_state["enable_llm_summary"] = False
st.session_state["num_variants"] = 3
st.session_state["lang_index"] = 0
st.session_state["llm_message"] = ""
st.session_state["llm_messages"] = []
st.session_state["triage_prompt_variants"] = ['''You are a telemedicine triage agent that decides between the following:
Emergency: Patient health is at risk if he doesn't speak to a Doctor urgently
Telecare: Patient can likely be treated remotely
General Practitioner: Patient should visit a GP for an ad-real consultation''',
'''You are a Doctor assistant that decides if a medical case can likely be treated remotely by a Doctor or not.
The remote Doctor can write prescriptions and request the patient to provide a picture.
Provide the triage recommendation first, and then explain your reasoning respecting the format given below:
Treat remotely: <your reasoning>
Treat ad-real: <your reasoning>''',
'''You are a medical triage agent working for the telemedicine Company Medgate based in Switzerland.
You decide if a case can be treated remotely or not, knowing that the remote Doctor can write prescriptions and request pictures.
Provide the triage recommendation first, and then explain your reasoning respecting the format given below:
Treat remotely: <your reasoning>
Treat ad-real: <your reasoning>''']
st.session_state['nbqs'] = []
st.session_state['citations'] = {}
st.session_state['past_messages'] = []
st.session_state["last_request"] = None
st.session_state["last_proposal"] = None
st.session_state['doctor_question'] = ''
st.session_state['patient_reply'] = ''
st.session_state['chat_history_array'] = []
st.session_state['chat_history'] = ''
st.session_state['feed_summary'] = ''
st.session_state['summary'] = ''
st.session_state["selected_guidelines"] = ["General"]
st.session_state["guidelines_dict"] = get_guidelines_dict()
st.session_state["triage_recommendation"] = ''
st.session_state["session_events"] = []
def init_session_state():
print('init_session_state()')
st.session_state['session_started'] = False
st.session_state['guidelines_ignored'] = False
st.session_state['model_index'] = 1
st.session_state["model_repo_id"] = "TheBloke/meditron-7B-GGUF"
st.session_state["model_filename"] = "meditron-7b.Q5_K_S.gguf"
st.session_state["model_type"] = "llama"
st.session_state['gpu_layers'] = 1
default_gender_index = 0
st.session_state['gender'] = get_genders()[default_gender_index]
st.session_state['gender_index'] = default_gender_index
st.session_state['age'] = 30
st.session_state['patient_medical_info'] = ''
default_search_query = 0
st.session_state['search_query_type'] = get_search_query_type_options()[default_search_query]
st.session_state['search_query_type_index'] = default_search_query
st.session_state['engine'] = get_available_engines()[0]
st.session_state['temperature'] = 0.0
st.session_state['top_p'] = 1.0
st.session_state['feed_chat_transcript'] = ''
st.session_state["llm_model"] = True
st.session_state["hugging_face_models"] = True
st.session_state["local_models"] = True
restart_session()
st.session_state['system_prompt'] = get_system_prompt()
st.session_state['system_prompt_after_on_change'] = get_system_prompt()
st.session_state["summary"] = ''
def get_genders():
return ['Male', 'Female']
def display_session_overview():
st.subheader('History of LLM queries')
st.write(st.session_state["llm_messages"])
st.subheader("Session costs overview")
df_session_overview = pd.DataFrame.from_dict(st.session_state["session_events"])
st.write(df_session_overview)
if "prompt_tokens" in df_session_overview:
prompt_tokens = df_session_overview["prompt_tokens"].sum()
st.write("Prompt tokens: " + str(prompt_tokens))
prompt_cost = df_session_overview["prompt_cost_chf"].sum()
st.write("Prompt CHF: " + str(prompt_cost))
completion_tokens = df_session_overview["completion_tokens"].sum()
st.write("Completion tokens: " + str(completion_tokens))
completion_cost = df_session_overview["completion_cost_chf"].sum()
st.write("Completion CHF: " + str(completion_cost))
completion_cost = df_session_overview["total_cost_chf"].sum()
st.write("Total costs CHF: " + str(completion_cost))
total_time = df_session_overview["response_time"].sum()
st.write("Total compute time (ms): " + str(total_time))
def remove_question(question_id):
st.session_state["questions"] = [value for value in st.session_state["questions"] if
str(value["question_id"]) != str(question_id)]
st.rerun()
def get_prompt_from_lead_symptom(df_config, df_prompt, lead_symptom, lang, fallback=True):
de_lead_symptom = lead_symptom
if lang != "DE":
df_lead_symptom = df_config[df_config[lang + ": Symptome"] == lead_symptom]
de_lead_symptom = df_lead_symptom["DE: Symptome"].iloc[0]
print("DE lead symptom: " + de_lead_symptom)
for i, row in df_prompt.iterrows():
if de_lead_symptom in row["Questionnaire"]:
return row["Prompt"]
warning_text = "No guidelines found for lead symptom " + lead_symptom + " (" + de_lead_symptom + ")"
if fallback:
st.toast(warning_text + ", using generic prompt", icon='🚨')
return st.session_state["system_prompt"]
st.toast(warning_text, icon='🚨')
return ""
def get_scenarios(df):
return [v for v in df.columns.values if v.startswith('TLC') or v.startswith('GP')]
def get_gender_age_from_test_scenario(test_scenario):
try:
result = re.search(r"([FM])(\d+)", test_scenario)
res_age = int(result.group(2))
gender = result.group(1)
res_gender = None
if gender == "M":
res_gender = "Male"
elif gender == "F":
res_gender = "Female"
else:
raise Exception('Unexpected gender')
return res_gender, res_age
except:
st.error("Unable to extract name, gender; using 30M as default")
return "Male", 30
def get_freetext_to_reco(reco_freetext_cased, emg_class_enabled=False):
reco_freetext = ""
if reco_freetext_cased:
reco_freetext = reco_freetext_cased.lower()
if reco_freetext.startswith('treat remotely') or reco_freetext.startswith('telecare'):
return 'TELECARE'
if reco_freetext.startswith('treat ad-real') or reco_freetext.startswith('gp') \
or reco_freetext.startswith('general practitioner'):
return 'GP'
if reco_freetext.startswith('emergency') or reco_freetext.startswith('emg') \
or reco_freetext.startswith('urgent'):
if emg_class_enabled:
return 'EMERGENCY'
return 'GP'
if "gp" in reco_freetext or 'general practitioner' in reco_freetext \
or "nicht über tele" in reco_freetext or 'durch einen arzt erford' in reco_freetext \
or "persönliche untersuchung erfordert" in reco_freetext:
return 'GP'
if ("telecare" in reco_freetext or 'telemed' in reco_freetext or
'can be treated remotely' in reco_freetext):
return 'TELECARE'
if ('emergency' in reco_freetext or 'urgent' in reco_freetext or
'not be treated remotely' in reco_freetext or "nicht tele" in reco_freetext):
return 'GP'
warning_msg = 'Cannot extract reco from LLM text: ' + reco_freetext
st.toast(warning_msg)
print(warning_msg)
return 'TRIAGE_IMPOSSIBLE'
def get_structured_reco(row, index, emg_class_enabled):
freetext_reco_col_name = "llm_reco_freetext_" + str(index)
freetext_reco = row[freetext_reco_col_name].lower()
return get_freetext_to_reco(freetext_reco, emg_class_enabled)
def add_expected_dispo(row, emg_class_enabled):
disposition = row["disposition"]
if disposition == "GP" or disposition == "TELECARE":
return disposition
if disposition == "EMERGENCY":
if emg_class_enabled:
return "EMERGENCY"
return "GP"
raise Exception("Missing disposition for row " + str(row.name) + " with summary " + row["case_summary"])
def get_test_scenarios(df):
res = []
for col in df.columns.values:
if str(col).startswith('GP') or str(col).startswith('TLC'):
res.append(col)
return res
def get_transcript(df, test_scenario, lang):
transcript = ""
for i, row in df.iterrows():
transcript += "\nDoctor: " + row[lang + ": Fragen"]
transcript += ", Patient: " + str(row[test_scenario])
return transcript
def get_expected_from_scenario(test_scenario):
reco = test_scenario.split('_')[0]
if reco == "GP":
return "GP"
elif reco == "TLC":
return "TELECARE"
else:
raise Exception('Unexpected reco: ' + reco)
def plot_report(title, expected, predicted, display_labels):
st.markdown('#### ' + title)
conf_matrix = confusion_matrix(expected, predicted, labels=display_labels)
conf_matrix_plot = ConfusionMatrixDisplay(confusion_matrix=conf_matrix, display_labels=display_labels)
conf_matrix_plot.plot()
st.pyplot(plt.gcf())
report = classification_report(expected, predicted, output_dict=True)
df_report = pd.DataFrame(report).transpose()
st.write(df_report)
df_rp = df_report
df_rp = df_rp.drop('support', axis=1)
df_rp = df_rp.drop(['accuracy', 'macro avg', 'weighted avg'])
try:
ax = df_rp.plot(kind="bar", legend=True)
for container in ax.containers:
ax.bar_label(container, fontsize=7)
plt.xticks(rotation=45)
plt.legend(loc=(1.04, 0))
st.pyplot(plt.gcf())
except Exception as e:
# Out of bounds
pass
def get_complete_prompt(generic_prompt, guidelines_prompt):
complete_prompt = ""
if generic_prompt:
complete_prompt += generic_prompt
if generic_prompt and guidelines_prompt:
complete_prompt += ".\n\n"
if guidelines_prompt:
complete_prompt += guidelines_prompt
return complete_prompt
def run_command(args):
"""Run command, transfer stdout/stderr back into Streamlit and manage error"""
cmd = ' '.join(args)
result = subprocess.run(cmd, capture_output=True, text=True)
print(result)
def get_diarized_f_path(audio_f_name):
# TODO p2: Quick hack, cleaner with os or regexes
base_name = audio_f_name.split('.')[0]
return DATA_FOLDER + base_name + ".txt"
def display_llm_output():
st.header("LLM")
form = st.form('llm')
llm_message = form.text_area('Message', value=st.session_state["llm_message"])
api_submitted = form.form_submit_button('Submit')
if api_submitted:
llm_response = get_llm_response(
st.session_state["model_repo_id"],
st.session_state["model_filename"],
st.session_state["model_type"],
st.session_state["gpu_layers"],
llm_message)
st.write(llm_response)
st.write('Done displaying LLM response')
def main():
print('Running Local LLM PoC Streamlit app...')
session_inactive_info = st.empty()
if "session_started" not in st.session_state or not st.session_state["session_started"]:
init_session_state()
display_streamlit_sidebar()
else:
display_streamlit_sidebar()
session_inactive_info.empty()
display_llm_output()
display_session_overview()
if __name__ == '__main__':
main()