mathieu/jarvis-server-v2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

improve gpt json parsing and history system

Browse Source
This commit is contained in:
Mathieu Broillet 2023-06-01 19:51:55 +02:00
parent ecf2595273
commit af4e7bb7f3
6 changed files with 146 additions and 75 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
src/resources/db-jarvis-commands-memory.sqlite
View File

Binary file not shown.

4
src/resources/gpt_prompt_v3.txt
View File

@ -1,2 +1,2 @@
You a sarcastic, introverted smart home assistant, dry humor.
Simplify complex sentence (ex: turn on tv in room), it should not be a question, add short response, return only JSON {"simplified_sentence", "response"}
You a sarcastic, introverted smart (home assistant) bot, dry humor.
Simplify complex sentence (ex. to : turn on tv in room), add short response, return only JSON {"simplified_sentence", "response"}

47
src/utils/chatgpt_utils.py
View File

@ -5,8 +5,8 @@ import time
import openai
from src.get_path_file import project_path
from src.database import db_utils
from src.get_path_file import project_path
chat_messages = {}
@ -24,11 +24,25 @@ def chatgpt_recognise(text, uuid):
# Add the user message to the chat history
chat_messages[uuid].append({"role": "user", "content": text})
history = chat_messages.get(uuid)
if len(chat_messages[uuid]) > 2:
history = []
prompt = chat_messages.get(uuid)[0]
last_messages = "\nFor context, last questions were: \n"
for message in chat_messages.get(uuid)[-6:]:
if message["role"] == "user":
last_messages += message["content"] + "\n"
prompt['content'] = prompt['content']+last_messages
history.append(prompt)
# Call ChatGPT API
start_time = time.time()
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=chat_messages.get(uuid),
# return first and last three messages from array
messages=history,
)
end_time = time.time()
logging.info("GPT-3 response in " + str(round(end_time - start_time, ndigits=2)) + " seconds")
@ -51,6 +65,15 @@ def chatgpt_recognise(text, uuid):
# Return the response
return query_json
elif 'response' in query_json:
response = query_json['response']
# Add the response to the chat history
chat_messages[uuid].append({"role": "assistant", "content": response})
# Return the response
return query_json
except Exception as e:
# If the response is not a JSON, it's probably a plaintext response
logging.error("Error while parsing ChatGPT response, probably not JSON: " + str(response.choices))
@ -110,16 +133,34 @@ def parse_gpt_json(input_string):
if not input_string.endswith('\"}'):
input_string = input_string.replace('}', '\"}')
# If at this point, no JSON seems to be found, we convert the output to JSON
if "response" not in input_string and "simplified_sentence" not in input_string:
input_string = input_string.replace("{", "").replace("}", "").replace("\"", "")
# If the input string contains "Simplifié:" or "Réponse:", we split the string
if "Simplifié:" in input_string or "Réponse:" in input_string:
simplified_sentence = input_string.split("Simplifié: ")[1].split("Réponse: ")[0]
response = input_string.split("Réponse:")[1]
input_string = '{"simplified_sentence": "' + simplified_sentence + '", "response": "' + response + '"}'
elif input_string.count(':') == 2:
simplified_sentence = input_string.split(" : ")[1].split(". ")[0]
response = input_string.split(" : ")[2]
input_string = '{"simplified_sentence": "' + simplified_sentence + '", "response": "' + response + '"}'
else:
input_string = '{"response": "' + input_string + '"}'
parsed_json = json.loads(input_string)
return parsed_json
except json.JSONDecodeError as e:
# Handle JSON decoding errors
print("Failed to parse JSON: " + str(e))
print("Input string: " + input_string)
# Return an error message
return None
def clear_chat(uuid):
"""Clear the chat history for a given uuid"""
logging.info("Cleared chat for uuid " + uuid)
chat_messages[uuid] = []
setup_messages(uuid)

2
src/utils/faster_whisper_utils.py
View File

@ -26,7 +26,7 @@ def faster_whisper_stt(audio_file):
logging.error("Model is not loaded")
load_model()
segments, info = model.transcribe(audio_file, beam_size=5)
segments, info = model.transcribe(audio_file, beam_size=5, language='fr')
print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
# combines all segments in one string

69
src/utils/whisper_utils.py.disabled
View File

@ -1,69 +0,0 @@
import json
import logging
import requests
from pywhispercpp.model import Model
from jarvis.utils import languages_utils
def load_model():
log_level = logging.getLogger().level
global model
model = Model('base', n_threads=8, suppress_non_speech_tokens=True, log_level=logging.ERROR)
logging.getLogger().setLevel(log_level)
def get_model():
return model
def whisper_cpp_stt(audio_file):
"""
Transcribe audio file using whisper-cpp, no additional server/service needed, runs on CPU.
:param audio_file:
:param model:
:return: text
"""
if model is None:
logging.error("Model is not loaded")
load_model()
segments = model.transcribe(audio_file, speed_up=False, translate=False)
# combines all segments in one string
text = ''
for segment in segments:
text += segment.text + ' '
return text
def whisper_asr_stt(audio_file):
"""
Transcribe audio file using whisper-asr (docker), a server is needed, runs on GPU.
See : https://github.com/ahmetoner/whisper-asr-webservice
:param audio_file:
:return: text
"""
headers = {
'accept': 'application/json',
# 'Content-Type': 'multipart/form-data',
}
params = {
'task': 'transcribe',
# TODO: add to config
'language': languages_utils.get_language(),
'output': 'json',
}
files = {
'audio_file': open(audio_file, 'rb'),
}
# TODO: add to config
response = requests.post('https://whisper.yourdomain.xyz/asr', params=params, headers=headers, files=files)
return json.loads(response.text)['text']

99
test_json.py Normal file
View File

@ -0,0 +1,99 @@
import json
import re
def parse_invalid_json(input_string):
try:
if "{" in input_string and "}" in input_string:
input_string = str(input_string).split("{", 1)[1].rsplit("}", 1)[0]
# Preprocess the input to fix common issues
input_string = input_string.replace('\n', '')
input_string = input_string.replace('\'', '\"')
input_string = input_string.replace('\"', '"')
input_string = input_string.replace('simplified_sentence:', '\"simplified_sentence\":')
input_string = input_string.replace('response:', '\"response\":')
input_string = input_string.replace(',}', '}')
# Fix missing commas
input_string = input_string.replace('}', '},')
input_string = input_string.replace(',"}', '}')
# Remove trailing commas
input_string = input_string.rstrip(',')
# Wrap the input in curly braces if necessary
if not input_string.startswith('{'):
input_string = '{' + input_string
if not input_string.endswith('}'):
input_string += '}'
# Replace matched patterns (didn"t by didn't) with single quotes
input_string = re.sub(r'(?<=[\wÀ-ÖØ-öø-ÿ])"(?=[\wÀ-ÖØ-öø-ÿ])', "'", input_string)
# Replace matched patterns with single quotes
input_string = input_string.replace('{\"Simplified Sentence\": ', '{\"simplified_sentence\": \"')
input_string = input_string.replace('"Simplified Sentence"', '\"simplified_sentence\"')
input_string = input_string.replace('"Simplified_sentence"', '\"simplified_sentence\"')
input_string = input_string.replace('"Response"', '"response"')
if "\"response\": \"" not in input_string:
if "response\": " in input_string:
input_string = input_string.replace('"response": ', '\", \"response\": \"')
else:
input_string = input_string.replace('"response":', '\", \"response\": \"')
# If missing a closing quote, add it
if not input_string.endswith('\"}'):
input_string = input_string.replace('}', '\"}')
# If at this point, no JSON seems to be found, we convert the output to JSON
if "response" not in input_string and "simplified_sentence" not in input_string:
input_string = input_string.replace("{", "").replace("}", "").replace("\"", "")
# If the input string contains "Simplifié:" or "Réponse:", we split the string
if "Simplifié:" in input_string or "Réponse:" in input_string:
simplified_sentence = input_string.split("Simplifié: ")[1].split("Réponse: ")[0]
response = input_string.split("Réponse:")[1]
input_string = '{"simplified_sentence": "' + simplified_sentence + '", "response": "' + response + '"}'
elif input_string.count(':') == 2:
simplified_sentence = input_string.split(" : ")[1].split(". ")[0]
response = input_string.split(" : ")[2]
input_string = '{"simplified_sentence": "' + simplified_sentence + '", "response": "' + response + '"}'
else:
input_string = '{"response": "' + input_string + '"}'
parsed_json = json.loads(input_string)
return parsed_json
except json.JSONDecodeError as e:
# Handle JSON decoding errors
print("Failed to parse JSON: " + str(e))
print("Input string: " + input_string)
# Return an error message
return None
if __name__ == '__main__':
inputs = [
'"simplified_sentence": "Turn on living room TV",\n"response": "Sure thing, turning on the living room TV."',
'simplified_sentence: "Turn on living room TV", response: "Sure thing, turning on the living room TV."',
'{"simplified_sentence": "Turn on living room TV", "response": "Sure thing, turning on the living room TV."}',
'{simplified_sentence: "Turn on living room TV", response: "Sure thing, turning on the living room TV."}',
'{simplified_sentence: "Turn on living room TV", "response": "Sure thing, turning on the living room TV."}',
'{"simplified_sentence": "Turn on living room TV", response: "Sure thing, turning on the living room TV."}',
'{simplified_sentence: "Turn on living room TV", response: "Sure thing, turning on the living room TV."',
'simplified_sentence: "Turn on living room TV", response: "Sure thing, turning on the living room TV."}',
'{"simplified_sentence": "Turn on living room TV", "response": "Sure thing, turning on the living room TV."',
'"simplified_sentence": "Turn on living room TV", "response": "Sure thing, turning on the living room TV."}',
'\"Turn on the TV in the living room\" - \"Sure thing, boss.\" \n\n{\"simplified_sentence\": \"Turn on the TV in the living room\", \"response\": \"Sure thing, boss.\"}',
'{"Simplified Sentence": She managed to turn off the living room couch."Response": Congratulations, I didn"t know a couch could be turned on in the first place.}',
'\"Simplified_sentence\": \"Turn off the couch in the living room.\",\n\"Response\": \"Sure, let me handle that for you.\"',
'{"simplified_sentence": "Éteindre la télé du salon.", "response": "D\'accord, j"éteins la télé du salon."}',
'Simplifié: Age de Robert Downey Jr. durant Iron Man1.Réponse: RObert Downey Jr avait 43 ans lors de las ortie u premier film Iron Man en 2008',
'Bien sur, voici une phrase simplifiée : Eteins la lumière du salon. Et voici la réponse : D\'accord, j\'éteins la lumière du salon.'
]
for input_string in inputs:
parsed = parse_invalid_json(input_string)
print(parsed)