코드는 너무 길어서 맨 아래에 첨부하겠습니다. 자세한 코드는 위에 링크한 저희 github를 참조해주세요
3. 서비스 기대효과
고민 없이 1분 만에 나에게 맞는 데이트 코스 생성
원하지 않는 코스는 수정 , 나만의 데이트 코스를 간편하게 생성
마음에 들었던 코스는 저장 후 손쉽게 관리
4. 향후 발전 방향
기능개선
GPT-4 API 도입
코스 생성 지역 추가 (e.g., 지방 여행)
지원 언어 확대 (Azure의 번역 모듈 활용)
추가 기능 개발
이전 생성 코스에 대한 평가 요청, 다음 코스 생성시에 반영
생성된 코스를 다른 사람들과 공유할 수 있는 기능
APPENDIX: LOCO-GPT
import os
import re
import json
from typing import List, Dict
import openai
from google.cloud import translate
# API Keys
os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "YOURGOOGLEAPI.json" # 수정 필요
openai.api_key = "YOUROPENAIKEY" # 수정 필요
google_trans_id = "YOURGOOGLEAPIID" # 수정 필요
def place_en_to_ko(location: str) -> str:
place_dict ={'강남/역삼/선릉': 'Gangnam / Yeoksam / Seolleung',
'강남구청': 'Gangnam-gu Office',
'개포/일원/수서': 'Gaepo/Irwon/Suseo',
'건대/군자/구의': 'Kondae/Gunja/Guui',
'금호/옥수': 'Kumho/Oksu',
'명동/을지로/충무로': 'Myeongdong/Euljiro/Chungmuro',
'방이': 'Bangi-dong',
'북촌/삼청': 'Bukchon/Samcheong',
'삼성/대치': 'Samsung/Daechi',
'상수/합정/망원': 'Sangsu/Hapjeong/Mangwon',
'서울역/회현': 'Seoul Station/Hoehyeon',
'서초/방배': 'Seocho/Bangbae',
'서촌': 'Seochon',
'선릉/삼성': 'Seolleung/Samsung',
'성수/뚝섬': 'Seongsu/Ttukseom',
'신사/논현': 'Sinsa/Nonhyeon',
'신촌/홍대/서교': 'Sinchon / Hongdae / Seogyo',
'압구정/청담': 'Apgujeong/Cheongdam',
'양재/도곡': 'Yangjae/Dogok',
'연남': 'Yeonnam',
'영동포/여의도': 'Yeongdongpo/Yeouido',
'용산/삼각지': 'Yongsan/Samgakji',
'이태원/한남': 'Itaewon/Hannam',
'잠실/송파': 'Jamsil/Songpa',
'종로/광화문': 'Jongno/Gwanghwamun',
'분당': 'Bundang-gu',
'수원/광교': 'Suwon/Gwanggyo',
'판교': 'Pangyo'}
gyeonggido_list = ['Bundang-gu', 'Suwon/Gwanggyo', 'Pangyo']
# gyeonggido-rization
if place_dict[location] in gyeonggido_list:
return place_dict[location] + ", GyeongGi-Do"
else:
return place_dict[location] + ", Seoul"
def translator(text: str, task: str, project_id=google_trans_id) -> str:
client = translate.TranslationServiceClient()
location = "us-central1"
# 아까 지정한 용어집 이름과 같이 맞춰줘야 함
glossary = client.glossary_path(project_id, location, 'loco_translator_glossary')
glossary_config = translate.TranslateTextGlossaryConfig(glossary=glossary)
parent = f"projects/{project_id}/locations/{location}"
# Translate text from English to Korean
if task == 'ko-en':
response = client.translate_text(
request={
"parent": parent,
"contents": [text],
"mime_type": "text/plain", # mime types: text/plain, text/html
"source_language_code": "ko",
"target_language_code": "en-US",
"glossary_config": glossary_config,
}
)
return response.glossary_translations[0].translated_text
elif task == 'en-ko':
response = client.translate_text(
request={
"parent": parent,
"contents": [text],
"mime_type": "text/plain", # mime types: text/plain, text/html
"source_language_code": "en-US",
"target_language_code": "ko",
"glossary_config": glossary_config,
}
)
return response.glossary_translations[0].translated_text
else:
pass
def translator_chatgpt(text: str, task: str) -> str:
TRANSLATION_TASK_PROMPT = ""
if task == 'ko-en':
TRANSLATION_TASK_PROMPT = f"Translate {text} into English."
elif task == 'en-ko':
TRANSLATION_TASK_PROMPT = f"Translate {text} into Korean."
else:
pass
LOCO_TRANSLATION_PROMPT = f"""{TRANSLATION_TASK_PROMPT}
You should not translate the following words: ACTIVITY, "activity_name", "start_time", "end_time", "description", "budget".
Do not generate both before and after translation results. Begin!:
"""
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[{"role": "user",
"content": f"{LOCO_TRANSLATION_PROMPT}."}],
temperature=1.1,
max_tokens=2048,
top_p=0.9,
frequency_penalty=0.0,
presence_penalty=0.0
)
result = response.choices[0].message.content
return result
def init():
LOCO_PREFIX_RPOMPT = """We are making LOCO, a dating application.
LOCO aims to recommend the best dating route by receiving the Meeting and Parting time, Cost, and Approximate location as input.
You must tell the user in detail where and what to do depending on the time of day.
You should not provide virtual location information, but must present location information that exists in reality.
You must not recommend physically impossible routes. For example, Human can not travel from Anam Station to Seoul Nat'l Univ. Station in 1 minutes.
You must not make recommendations beyond a given budget. For example, you shouldn't be recommending $10 spaghetti to your user if user's budget is only $5.
You must create a date path strictly according to our instructions.
Do you understand this task is?:"""
while True:
init_loco = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[{"role": "user",
"content": f"{LOCO_PREFIX_RPOMPT}"}],
temperature=1.1,
max_tokens=2048,
top_p=0.9,
frequency_penalty=0.0,
presence_penalty=0.0
)
print("---------------------------------------------------")
print("\"Human\": Do you understand what this task is?")
print("\"AI\":", init_loco.choices[0].message.content)
if "yes" in init_loco.choices[0].message.content.lower():
print("\"Human\": Great, Let's begin!")
print("---------------------------------------------------")
break
else:
continue
def postprocessing(result: str) -> str:
"""
[Exception 1] OUTPUT FORMAT:
ACTIVITY
1. ~
[Exception 1-1] "activity_name" not in result
ACTIVITY
1. "Lotte World Tower Observation Deck"
[Exception 1-2] "ACTIVITY" not in result
1. Breakfast at Bongchu Jjimdak (봉추찜닭) in Jamsil
- "activity_name": "Bongchu Jjimdak (봉추찜닭)",
[Exception 1-3] Variation 1-3
ACTIVITY
1. "activity_name": "Seoul Sky Observatory",
[Exception 1-4] Variation 1-4
ACTIVITY
1. "Breakfast at Paul Bassett"
[Exception 1-5] There is one ACTIVITY
ACTIVITY
"activity_name": "Seonjeongneung Park",
"start_time": "2023-04-09T11:00:00",
"end_time": "2023-04-09T13:00:00",
"description": "A historic park featuring the tombs of two Joseon Dynasty kings and their queens.",
"budget": "0"
"activity_name": "Starfield COEX Mall",
[Exception 2] UNINTENDED OUTPUT:
"budget": "39000"
Total budget: 101000 Won <- THIS
[Exception 3] UNINTENDED OUTPUT between activities
"budget": "0"
To end the day, we can grab some coffee and dessert at Starbucks Yangjae Station Branch. It's a cozy place to chat and relax after a long day. <- THIS
"activity_name": "Starbucks Yangjae Station Branch",
"""
# remove multiple spaces
result = re.sub(' +', '', result)
result = result.replace("- ", "")
pattern = """ACTIVITY
\"activity_name\""""
result = result.replace(pattern, "ACTIVITY\n\"activity_name\"")
# exception: 1-5
if result.count('ACTIVITY') == 1:
if "1. " in result:
pattern = re.compile(r'\d\. .*')
findall_num = pattern.findall(result)
for f in findall_num:
result = re.sub(f, "", result)
result = result.replace("\n\"activity_name\"", "ACTIVITY\n\"activity_name\"")
else:
result = result.replace("\n\"activity_name\"", "ACTIVITY\n\"activity_name\"")
# exception: 3
exception_pattern_3 = re.compile(r'\n\n.+\n')
findall_between = exception_pattern_3.findall(result)
if findall_between:
for f in findall_between:
if "\"activity_name\"" in f or 'ACTIVITY' in f:
continue
result = result.replace(f, "")
# exception: 1-4
activity_findall = re.compile(r'\d\. .*').findall(result)
for a in activity_findall:
if "activity_name" not in a:
result = re.sub(a, '\nACTIVITY', result)
# exception: 1-3
if "ACTIVITY\n1. \"activity_name\"" in result:
result = re.sub(r'\d\. ', 'ACTIVITY\n', result)
exception_pattern_1 = re.compile(r'\d\. ' + '\"*[a-zA-Z| *|:|é|\"]*\"*')
# exception: 1-1
if "activity_name" not in result:
activity_findall = exception_pattern_1.findall(result)
for a in activity_findall:
result = re.sub(exception_pattern_1, 'ACTIVITY\n\"activity_name\": ' + a, result)
result = re.sub(r'\d\. ', "", result)
# exception: 1-2
if "ACTIVITY" not in result:
activity_findall = exception_pattern_1.findall(result)
for _ in activity_findall:
result = re.sub(exception_pattern_1, 'ACTIVITY\n', result)
result = re.sub(r'\d\. ', "", result)
result = re.sub(r'\([^)]*\) *[a-zA-Z| *|가-힣]*', "", result)
# exception: 1
if exception_pattern_1.match(result):
result = re.sub('\d\. ' + '[a-zA-Z | *]*', 'ACTIVITY', result)
if "ACTIVITY\n\nACTIVITY" in result:
result = result.replace("ACTIVITY\n\nACTIVITY", "ACTIVITY")
if "\n\n\nACTIVITY" in result:
result = result.replace("\n\n\nACTIVITY", "\n\nACTIVITY")
if "ACTIVITY\n\nACTIVITY" in result:
result = result.replace("ACTIVITY\n\nACTIVITY", "ACTIVITY\n")
# exception: 2
exception_pattern_2 = re.compile(r"\"budget\": \"[a-zA-Z0-9]*\"")
findall_budget = exception_pattern_2.findall(result)
if findall_budget: # not empty
chop_idx_from = result.rfind(findall_budget[-1])
chop_idx_to = len(findall_budget[-1])
result = result[:chop_idx_from + chop_idx_to]
# e.g., ACTIVITY 1 -> ACTIVITY
result = re.sub(r'ACTIVITY ' + '[0-9]+', 'ACTIVITY', result)
# e.g., ACTIVITY 1: OR ACTIVITY1: OR ACTIVITY: OR ACTIVITY : -> ACTIVITY
result = re.sub(r'ACTIVITY' + '[0-9| ]*:', 'ACTIVITY', result)
# e.g., 1. ACTIVITY -> ACTIVITY
result = re.sub(r'\d\. ACTIVITY', 'ACTIVITY', result)
# remove (SOMETHING)
result = re.sub(re.compile(r'\([^)]*\) *'), "", result)
# remove #
# e.g., "budget": "0" # free admission -> "budget": "0"
result = re.sub(re.compile(r'# *.*'), "", result)
if "ACTIVITYACTIVITY" in result:
result = result.replace("ACTIVITYACTIVITY", "ACTIVITY")
if "ACTIVITY\nACTIVITY" in result:
result = result.replace("ACTIVITY\nACTIVITY", "ACTIVITY")
return result
def remove_verb(p: str) -> str:
p = p.replace("Artistic exploration at ", "")
p = p.replace("Breakfast at ", "")
p = p.replace("Brunch at ", "")
p = p.replace("Chimaek at ", "")
p = p.replace("Coffee at ", "")
p = p.replace("Coffee and Cake at ", "")
p = p.replace("Coffee and dessert at ", "")
p = p.replace("Coffee break at ", "")
p = p.replace("Cycling in ", "")
p = p.replace("Dinner at ", "")
p = p.replace("Explore ", "")
p = p.replace("Go to ", "")
p = p.replace("Hiking at ","")
p = p.replace("Karaoke at ", "")
p = p.replace("Lunch at ", "")
p = p.replace("Night view at ", "")
p = p.replace("Night view of ", "")
p = p.replace("Picnic at ", "")
p = p.replace("Riding the Cable Car at ", "")
p = p.replace("Relax at ", "")
p = p.replace("Shopping at ", "")
p = p.replace("Shopping in ", "")
p = p.replace("Stroll around ", "")
p = p.replace("Tea time at ", "")
p = p.replace("Visit ", "")
p = p.replace("Visit to ", "")
p = p.replace("Visiting ", "")
p = p.replace("Virtual Reality Experience at ", "")
p = p.replace("Walking tour of ", "")
p = p.replace("Walking tour in ", "")
p = p.replace("Walk along ", "")
p = p.replace("Walk in ", "")
p = p.replace("Walk around ", "")
return p
def change_route(meeting_time: str,
parting_time: str,
budget: str,
place: str,
user_request: str,
prior_places: List[str]) -> str:
m_year, m_time = meeting_time.split("T")
m_year = m_year.split('-')
meeting_time = m_time.strip() + ", " + m_year[0] + "/" + m_year[1] + "/" + m_year[2]
p_year, p_time = parting_time.split("T")
p_year = p_year.split('-')
parting_time = p_time.strip() + ", " + p_year[0] + "/" + p_year[1] + "/" + p_year[2]
print(f"time: from {meeting_time} to {parting_time}, budget: {budget}, place: {place}")
print("---------------------------------------------------")
USER_REQUEST_PROMPT = ""
# user_request: "가격이 너무 비싸요"
if user_request:
user_request = translator(user_request, task='ko-en') # 'ko-en' or 'en-ko'
USER_REQUEST_PROMPT = f"You should consider this message: {user_request}. "
# prior_places: ['서울올림픽미술관', '꼬꼬춘천치킨', '코엑스 아쿠아리움', '스타필드 코엑스몰']
if prior_places:
temp_places = ""
for idx, pp in enumerate(prior_places):
if idx == len(prior_places) - 1:
temp_places += translator(pp, task="ko-en")
else:
temp_places = temp_places + translator(pp, task="ko-en") + ", "
USER_REQUEST_PROMPT += f"You must not include the following locations: {temp_places}"
print("USER_REQUEST_PROMPT: ", USER_REQUEST_PROMPT)
LOCO_INSTRUCTION_RPOMPT = f"""Plan a date from {meeting_time} to {parting_time}, budget is {budget}, and Meeting place is {place}.
You should actively recommend names of place that exist in reality, what it actually costs.
{USER_REQUEST_PROMPT}.
You should recommend only one place and don't give me a choice.
OUTPUT FORMAT:
```
ACTIVITY
"activity_name": "Starbucks Yangjae Station Branch",
"start_time": "2023-04-09T13:00:00",
"end_time": "2023-04-09T15:00:00",
"description": "Starbucks is the world's largest multinational coffee chain.",
"budget": "10000"
```
"activity_name" should be a place only and "description" should consist of one sentence
"""
LOCO_SUFFIX_PROMPT = """You must follow the OUTPUT FORMAT, Begin!:"""
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[{"role": "user",
"content": f"{LOCO_INSTRUCTION_RPOMPT}. {LOCO_SUFFIX_PROMPT}"}],
temperature=1.1,
max_tokens=2048,
top_p=0.9,
frequency_penalty=0.0,
presence_penalty=0.0
)
result = response.choices[0].message.content
return result
def generate_route(meeting_time: str,
parting_time: str,
budget: str,
place: str,
user_request: str) -> str:
m_year, m_time = meeting_time.split("T")
m_year = m_year.split('-')
meeting_time = m_time.strip() + ", " + m_year[0] + "/" + m_year[1] + "/" + m_year[2]
p_year, p_time = parting_time.split("T")
p_year = p_year.split('-')
parting_time = p_time.strip() + ", " + p_year[0] + "/" + p_year[1] + "/" + p_year[2]
print(f"time: from {meeting_time} to {parting_time}, budget: {budget}, place: {place}")
print("---------------------------------------------------")
USER_REQUEST_PROMPT = ""
# user_request: "가격이 너무 비싸요"
if user_request:
user_request = translator(user_request, task='ko-en') # 'ko-en' or 'en-ko'
USER_REQUEST_PROMPT = f"You should consider this message: {user_request}. "
# date time <= 2 hours -> recommend a place
SINGLE_PLACE_PROMPT = ""
if int(p_time.split(":")[0]) - int(m_time.split(":")[0]) <= 3:
SINGLE_PLACE_PROMPT = "If date time is less than 2 hours, you recommend only one place. Do not give me a choice."
LOCO_INSTRUCTION_RPOMPT = f"""Plan a date from {meeting_time} to {parting_time}, budget is {budget}, and Meeting place is {place}.
You should actively recommend names of place that exist in reality, what it actually costs.
{USER_REQUEST_PROMPT}. {SINGLE_PLACE_PROMPT}
OUTPUT FORMAT:
```
ACTIVITY
"activity_name": "Starbucks Yangjae Station Branch",
"start_time": "2023-04-09T13:00:00",
"end_time": "2023-04-09T15:00:00",
"description": "Starbucks is the world's largest multinational coffee chain.",
"budget": "10000"
```
"activity_name" should be a place only, "budget" should be only numbers and "description" should consist of one sentence
"""
LOCO_SUFFIX_PROMPT = """You must follow the OUTPUT FORMAT, Begin!:"""
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[{"role": "user",
"content": f"{LOCO_INSTRUCTION_RPOMPT}. {LOCO_SUFFIX_PROMPT}"}],
temperature=1.1,
max_tokens=2048,
top_p=0.9,
frequency_penalty=0.0,
presence_penalty=0.0
)
result = response.choices[0].message.content
return result
class LOCO:
def __init__(self):
print("Initializing LOCO")
init()
def inference(self,
meeting_time: str,
parting_time: str,
place: str,
budget: int,
user_request: str,
prior_places: List[str]) -> List[Dict]:
# preprocessing
place = place_en_to_ko(place)
budget = str(budget) + " Won"
if prior_places: # user-requested try (not empty)
print("Change a meeting place")
print("---------------------------------------------------")
# change a meeting place
result = change_route(meeting_time,
parting_time,
budget,
place,
user_request,
prior_places)
else: # first try
# generate schedule
print("Generate Routes")
result = generate_route(meeting_time,
parting_time,
budget,
place,
user_request)
print(result)
print("---------------------------------------------------")
# postprocessing
result = postprocessing(result)
print(result)
print("---------------------------------------------------")
# split paragraph
parsed_result = result.replace("\n\n", "").replace("\n", "").split("ACTIVITY")[1:]
# remove verb from activity, translate en to ko, and json formatting
parsed_result = [json.loads("{" + remove_verb(p) + "}") for p in parsed_result]
for r in parsed_result:
print(r)
print("---------------------------------------------------")
# translation and then budget; str to int
for idx in range(len(parsed_result)):
parsed_result[idx]["activity_name"] = translator(parsed_result[idx]["activity_name"], task='en-ko')
parsed_result[idx]["description"] = translator_chatgpt(parsed_result[idx]["description"], task='en-ko')
try:
parsed_result[idx]["budget"] = int(parsed_result[idx]["budget"])
except:
parsed_result[idx]["budget"] = 0
return
