Target audience: Beginner
Estimated reading time: 3'
This post describes a generic implementation of a client to the OpenAI ChatGPT REST web service. This implementation has been written and tested with Python 3.9. Comments in the source code are omitted for the sake of clarity.
Table of contents
ChatGPT API overview
Notes:
- Environments: Python 3.9.16, ChatGPT 3.5
- To enhance the readability of the algorithm implementations, we have omitted non-essential code elements like error checking, comments, exceptions, validation of class and method arguments, scoping qualifiers, and import statements.
ChatGPT API overview
Let's review the parameters for OpenAI Chat Completion REST API
HTTP Parameters
The connectivity parameters for the HTTP post are
- OPEN_AI_KEY=xxxxx
- URL= https://api.openai.com/v1/chat/completions
- CONTENT_TYPE=application/json
- AUTHORIZATION="Bearer ${OPEN_AI_KEY}"
Request body
The parameters of the POST content:
- model: Identifier of the model (i.e. gpt-3.5-turbo)
- messages: Text of the conversation
- user: Identifier for the user
- role: Role of the user {system|user|assistant}
- content: Content of the message or request
- name: Name of the author
- temperature: Hyper-parameter that controls the "creativity" of the language model by adjusting the distribution (softmax) for the prediction of the next word/token. The higher the value (> 0) the more diverse the prediction (default: 0)
- top_p: Sample the tokens with top_p probability. It is an alternative to temperature (default: 1)
- n: Number of solutions/predictions (default 1)
- max_tokens: Limit the number of tokens used in the response (default: Infinity)
- presence_penalty: Penalize new tokens which appear in the text so far if positive. A higher value favors most recent topics (default: 0)
- frequency_penalty: Penalize new tokens which appear in the text with higher frequency if the value is positive (default: 0)
Note: OpenAI models are non-deterministic, as identical requests may yield different answers. Setting temperature = 0 will make the outputs mostly deterministic.
Chat completion request
The first step is to implement the data structure for the content of the request as described in the previous section. The content of the request, ChatGPTRequest is a read-only and therefore implemented as a data class. The class members reflect the semantic of the OpenAI chat completion API.
- model: Identifier of the model (i.e. gpt-3.5-turbo, code-davinci-002 ...)
- role: Role of the user as system, user or assistant
- temperature: Hyper-parameter that adjusts the distribution for the prediction of the next token
- max_tokens: Limit the number of tokens used in the response
- top_p: Sample the tokens with p highest probability (default 1)
- n Number predictions/choices (default 1)
- presence_penalty: Penalize new tokens which appear in the text so far
- frequency_penalty: Penalize new tokens which appear in the text with higher frequency
from dataclasses import dataclass
@dataclass
class ChatGPTRequest:
model: str
role: str
temperature: float
max_tokens: int
top_p: int
n: int
presence_penalty: int
frequency_penalty: int
ChatGPT client
There are two constructors for the client of type ChatGPTClient
- Default constructor for a fully customizable request ChatGPTRequest as argument.
- Constructor build for simple request that required only model, role and temperature (evaluated with values 0 and 1) with all other parameters using the default values specified in the openAI API, openapi.ChatCompletion, documentation [ref 1]. We use the annotated type, InstanceType to specify the type of instance to create [ref 2]
from typing import Type, TypeVar, Callable
Instancetype = TypeVar('Instancetype', bound='ChatGPTClient')
class ChatGPTClient(object):
import constants
# static variable for the API key and the default maximum # number of tokens returned
openai.api_key = constants.openai_api_key
default_max_tokens = 1024
def __init__(self, chatGPTRequest: ChatGPTRequest):
self.chatGPTRequest = chatGPTRequest
@classmethod
def build(cls, model: str, role: str, temperature: float) -> Instancetype:
chatGPTRequest = ChatGPTRequest(
model,
role,
temperature,
ChatGPTClient.default_max_tokens,
1,
1,
0,
0)
return cls(chatGPTRequest)
Post
Let's start with a simple version of the invocation that returns only the answer without any explanation, status or usage. The only argument of the HTTP post is the user prompt. The response is extracted from the message of the first choice of the answer.
def post(self, prompt: str) -> str: import logging try: response = openai.ChatCompletion.create( model=self.chatGPTRequest.model, messages=[{'role': self.chatGPTRequest.user, 'content': prompt}], temperature=self.chatGPTRequest.temperature, max_tokens=self.chatGPTRequest.max_tokens ) return response['choices'][0].message.content
except openai.error.AuthenticationError as e:
logging.error(f'Failed as {str(e)}')
Some advanced client applications may require processing and evaluating some metadata included in the response. In this case, the JSON content of the Chat completion answer is objectified.
The key variables of the ChatCompletion response are
- id: Conversation identifier
- object: Payload of the response
- created: Creation date
- usage.prompt_tokens: Number of tokens used in the prompt
- usage.completion_tokens: Number of tokens used in the completion
- usage.total_tokens Total number of tokens
- choices
- choices.message.role Role used in the request
- choices.message.content Response content
- choices.finish_reason Description of the state of completion of the request
As with the request content, the difference components of the answer are implemented as data classes to reflect the structure of the JSON response from the ChatCompletion API.
The previous implementation of the post method is upgraded by adding a conversion of the JSON response into ChatGPTResponse.
Note: We are using the built-in json Python library [ref 3] to decode the ChatGPT response. The list of alternative libraries include Orjson [ref 4] and SimpleJson [ref 5]
@dataclass
class ChatGPTChoice:
messages: []
index: int
finish_reason: str
@dataclass
class ChatGPTUsage:
prompt_tokens: int
completion_tokens: int
total_tokens: int
@dataclass
class ChatGPTResponse:
id: str
object: str
created: int
model: int
choices: []
usage: ChatGPTUsage
def post_dev(self, prompt: str) -> ChatGPTResponse:
import json
import logging
try:
response = openai.ChatCompletion.create(
model=self.chatGPTRequest.model,
messages=[{'role': self.chatGPTRequest.user, 'content': prompt}],
temperature=self.chatGPTRequest.temperature,
max_tokens=self.chatGPTRequest.max_tokens
)
return json.loads(response)
except openai.error.AuthenticationError as e:
logging.error(f'Failed as {str(e)}')
Note: We are using the built-in json Python library [ref 3] to decode the ChatGPT response. The list of alternative libraries include Orjson [ref 4] and SimpleJson [ref 5]
References
[1] OpenAI API
[4] Orjson
---------------------------
Patrick Nicolas has over 25 years of experience in software and data engineering, architecture design and end-to-end deployment and support with extensive knowledge in machine learning.
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning" Packt Publishing ISBN 978-1-78712-238-3
He has been director of data engineering at Aideo Technologies since 2017 and he is the author of "Scala for Machine Learning" Packt Publishing ISBN 978-1-78712-238-3
