devops/encoach_backend
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial updates to most recent openai api version.

Browse Source
This commit is contained in:
Cristiano Ferreira
2024-05-19 14:37:50 +01:00
parent 070e8808b1
commit e568aff4e4
4 changed files with 162 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
helper/constants.py
View File

@@ -7,6 +7,8 @@ GRADING_TEMPERATURE = 0.1
TIPS_TEMPERATURE = 0.2
GEN_QUESTION_TEMPERATURE = 0.7
GPT_3_5_TURBO = "gpt-3.5-turbo"
GPT_4_TURBO = "gpt-4-turbo"
GPT_4_O = "gpt-4o"
GPT_3_5_TURBO_16K = "gpt-3.5-turbo-16k"
GPT_3_5_TURBO_INSTRUCT = "gpt-3.5-turbo-instruct"
GPT_4_PREVIEW = "gpt-4-turbo-preview"

187
helper/openai_interface.py
View File

@@ -1,15 +1,14 @@
import json
import os
import re
import openai
from openai import OpenAI
from dotenv import load_dotenv
from helper.constants import GPT_3_5_TURBO_INSTRUCT, BLACKLISTED_WORDS
from helper.constants import BLACKLISTED_WORDS, GPT_3_5_TURBO
from helper.token_counter import count_tokens
load_dotenv()
openai.api_key = os.getenv("OPENAI_API_KEY")
client = OpenAI(api_key=os.getenv('OPENAI_API_KEY'))
MAX_TOKENS = 4097
TOP_P = 0.9
@@ -50,105 +49,20 @@ tools = [{
}]
###
def process_response(input_string, quotation_check_field):
if '{' in input_string:
try:
# Find the index of the first occurrence of '{'
index = input_string.index('{')
# Extract everything after the first '{' (inclusive)
result = input_string[index:]
if re.search(r"'" + quotation_check_field + "':\s*'(.*?)'", result, re.DOTALL | re.MULTILINE) or \
re.search(r"'" + quotation_check_field + "':\s*\[([^\]]+)]", result, re.DOTALL | re.MULTILINE):
json_obj = json.loads(parse_string(result))
return json_obj
else:
if "title" in result:
parsed_string = result.replace("\n\n", "\n")
parsed_string = parsed_string.replace("\n", "**paragraph**")
else:
parsed_string = result.replace("\n\n", " ")
parsed_string = parsed_string.replace("\n", " ")
parsed_string = re.sub(r',\s*]', ']', parsed_string)
parsed_string = re.sub(r',\s*}', '}', parsed_string)
if (parsed_string.find('[') == -1) and (parsed_string.find(']') == -1):
parsed_string = parse_string_2(parsed_string)
return json.loads(parsed_string)
return json.loads(parsed_string)
except Exception as e:
print(f"Invalid JSON string! Exception: {e}")
print(f"String: {input_string}")
print(f"Exception: {e}")
else:
return input_string
def parse_string(to_parse: str):
parsed_string = to_parse.replace("\"", "\\\"")
pattern = r"(?<!\w)'|'(?!\w)"
parsed_string = re.sub(pattern, '"', parsed_string)
parsed_string = parsed_string.replace("\\\"", "'")
parsed_string = parsed_string.replace("\n\n", " ")
parsed_string = re.sub(r',\s*]', ']', parsed_string)
parsed_string = re.sub(r',\s*}', '}', parsed_string)
return parsed_string
def parse_string_2(to_parse: str):
keys_and_values_str = to_parse.replace("{", "").replace("}", "")
split_pattern = r'(?<="),|(?<="):'
keys_and_values = re.split(split_pattern, keys_and_values_str)
keys = []
values = []
for idx, x in enumerate(keys_and_values):
if (idx % 2) == 0:
keys.append(x)
else:
values.append(x)
parsed_values = []
for value in values:
parsed_values.append(("\"" + value.replace("\"", "").strip() + "\""))
for ind, parsed_value in enumerate(parsed_values):
to_parse = to_parse.replace(values[ind], parsed_values[ind])
to_parse = to_parse.replace(":", ": ")
return to_parse
def remove_special_chars_and_escapes(input_string):
parsed_string = input_string.replace("\\\"", "'")
parsed_string = parsed_string.replace("\n\n", " ")
# Define a regular expression pattern to match special characters and escapes
pattern = r'(\\[nrt])|[^a-zA-Z0-9\s]'
# Use re.sub() to replace the matched patterns with an empty string
cleaned_string = re.sub(pattern, '', parsed_string)
return cleaned_string
def check_fields(obj, fields):
return all(field in obj for field in fields)
def make_openai_call(model, messages, token_count, fields_to_check, temperature):
global try_count
result = openai.ChatCompletion.create(
result = client.chat.completions.create(
model=model,
max_tokens=int(MAX_TOKENS - token_count - 300),
temperature=float(temperature),
top_p=float(TOP_P),
frequency_penalty=float(FREQUENCY_PENALTY),
messages=messages
)["choices"][0]["message"]["content"]
messages=messages,
response_format={"type": "json_object"}
)
result = result.choices[0].message.content
if has_blacklisted_words(result) and try_count < TRY_LIMIT:
try_count = try_count + 1
return make_openai_call(model, messages, token_count, fields_to_check, temperature)
@@ -156,57 +70,22 @@ def make_openai_call(model, messages, token_count, fields_to_check, temperature)
return ""
if fields_to_check is None:
return result.replace("\n\n", " ").strip()
return json.loads(result)
processed_response = process_response(result, fields_to_check[0])
if check_fields(processed_response, fields_to_check) is False and try_count < TRY_LIMIT:
if check_fields(result, fields_to_check) is False and try_count < TRY_LIMIT:
try_count = try_count + 1
return make_openai_call(model, messages, token_count, fields_to_check, temperature)
elif try_count >= TRY_LIMIT:
try_count = 0
return result
return json.loads(result)
else:
try_count = 0
return processed_response
return json.loads(result)
def make_openai_instruct_call(model, message: str, token_count, fields_to_check, temperature):
global try_count
response = openai.Completion.create(
model=model,
prompt=message,
max_tokens=int(4097 - token_count - 300),
temperature=0.7
)["choices"][0]["text"]
if has_blacklisted_words(response) and try_count < TRY_LIMIT:
try_count = try_count + 1
return make_openai_instruct_call(model, message, token_count, fields_to_check, temperature)
elif has_blacklisted_words(response) and try_count >= TRY_LIMIT:
try_count = 0
return ""
if fields_to_check is None:
try_count = 0
return response.replace("\n\n", " ").strip()
response = remove_special_characters_from_beginning(response)
if response[0] != "{" and response[0] != '"':
response = "{\"" + response
if not response.endswith("}"):
response = response + "}"
try:
processed_response = process_response(response, fields_to_check[0])
reparagraphed_response = replace_expression_in_object(processed_response, "**paragraph**", "\n")
if check_fields(reparagraphed_response, fields_to_check) is False and try_count < TRY_LIMIT:
try_count = try_count + 1
return make_openai_instruct_call(model, message, token_count, fields_to_check, temperature)
else:
try_count = 0
return reparagraphed_response
except Exception as e:
return make_openai_instruct_call(model, message, token_count, fields_to_check, temperature)
return ""
# GRADING SUMMARY
@@ -254,7 +133,7 @@ def calculate_section_grade_summary(section):
messages[2:2] = [{"role": "user",
"content": "This section is s designed to assess the English language proficiency of individuals who want to study or work in English-speaking countries. The speaking section evaluates a candidate's ability to communicate effectively in spoken English."}]
res = openai.ChatCompletion.create(
res = client.chat.completions.create(
model="gpt-3.5-turbo",
max_tokens=chat_config['max_tokens'],
temperature=chat_config['temperature'],
@@ -298,20 +177,32 @@ def parse_bullet_points(bullet_points_str, grade):
def get_fixed_text(text):
message = ('Fix the errors in the given text and put it in a JSON. Do not complete the answer, only replace what '
'is wrong. Sample JSON: {"fixed_text": "fixed test with no '
'misspelling errors"}] \n The text: "' + text + '"')
token_count = count_tokens(message)["n_tokens"]
response = make_openai_instruct_call(GPT_3_5_TURBO_INSTRUCT, message, token_count, ["fixed_text"], 0.2)
messages = [
{"role": "system", "content": ('You are a helpful assistant designed to output JSON on this format: '
'{"fixed_text": "fixed test with no misspelling errors"}')
},
{"role": "user", "content": (
'Fix the errors in the given text and put it in a JSON. Do not complete the answer, only replace what '
'is wrong. \n The text: "' + text + '"')
}
]
token_count = count_total_tokens(messages)
response = make_openai_call(GPT_3_5_TURBO, messages, token_count, ["fixed_text"], 0.2)
return response["fixed_text"]
def get_speaking_corrections(text):
message = ('Fix the errors in the provided transcription and put it in a JSON. Do not complete the answer, only '
'replace what is wrong. Sample JSON: {"fixed_text": "fixed '
'transcription with no misspelling errors"}] \n The text: "' + text + '"')
token_count = count_tokens(message)["n_tokens"]
response = make_openai_instruct_call(GPT_3_5_TURBO_INSTRUCT, message, token_count, ["fixed_text"], 0.2)
messages = [
{"role": "system", "content": ('You are a helpful assistant designed to output JSON on this format: '
'{"fixed_text": "fixed transcription with no misspelling errors"}')
},
{"role": "user", "content": (
'Fix the errors in the provided transcription and put it in a JSON. Do not complete the answer, only '
'replace what is wrong. \n The text: "' + text + '"')
}
]
token_count = count_total_tokens(messages)
response = make_openai_call(GPT_3_5_TURBO, messages, token_count, ["fixed_text"], 0.2)
return response["fixed_text"]
@@ -340,3 +231,9 @@ def replace_expression_in_object(obj, expression, replacement):
elif isinstance(obj[key], dict):
obj[key] = replace_expression_in_object(obj[key], expression, replacement)
return obj
def count_total_tokens(messages):
total_tokens = 0
for message in messages:
total_tokens += count_tokens(message["content"])["n_tokens"]
return total_tokens