For my first contribution, I filed an issue to add a new feature to another project which is to add a new flag option to display the tokens used for the prompt and the completion generation.
Feat: chat completion token info flag option
#8
A flag option that gives the user a count of tokens sent and received. I think that it is an important feature that guides the user to stay within the token budget when making a chat completions request!
To do this, we would need to add another option flag which could be -t and --token-usage. When a user includes this flag to their command, it should display in clear detail how many tokens were used in the generation of the completion, and how many tokens were used in the prompt.
I chose to contribute to fadingNA's open source project, chat-minal, a CLI tool written in Python that allows you to leverage OpenAI to do various things, such as using it to generate a code review, file conversion, generating markdown from text, and summarizing text.
I have written code in Python before, but it is not my strongest skill. So contributing to this project provides a challenging but good learning experience for me.
The challenge is that I would have to read and understand someone else's code, and provide a proper solution in a way that it does not break the design of the code. Understanding the flow is crucial so that I can efficiently add the feature without having to make big changes in the code and keep the code consistent.
FEAT: Token usage flag
#9
Added the feature to include a --token_usage flag option for the user. This option gives the user the information of how many tokens were used for the prompt and generated completion.
The solution I came up with based on the code design is to check for the existence of the token_usage flag. I do not want the code to check any unnecessary if statements if the token_usage flag was not used, so I made two separate identical loop logic, with the difference of checking the the existence of usage_metadata inside chunk.
if token_usage:
for chunk in runnable.stream({"input_text": input_text}):
print(chunk.content, end="", flush=True)
answer.append(chunk.content)
if chunk.usage_metadata:
completion_tokens = chunk.usage_metadata.get('output_tokens')
prompt_tokens = chunk.usage_metadata.get('input_tokens')
else:
for chunk in runnable.stream({"input_text": input_text}):
print(chunk.content, end="", flush=True)
answer.append(chunk.content)At the end of the execution of get_completions() method, a check for the flag token_usage is added, which then displays the token usage details to stderr if the flag was used.
if token_usage:
logger.error(f"Tokens used for completion: <span class="pl-s1"><span class="pl-kos">{completion_tokens}</span>"</span>)
logger.error(f"Tokens used for prompt: <span class="pl-s1"><span class="pl-kos">{prompt_tokens}</span>"</span>)if token_usage:
for chunk in runnable.stream({"input_text": input_text}):
print(chunk.content, end="", flush=True)
answer.append(chunk.content)
if chunk.usage_metadata:
completion_tokens = chunk.usage_metadata.get('output_tokens')
prompt_tokens = chunk.usage_metadata.get('input_tokens')
else:
for chunk in runnable.stream({"input_text": input_text}):
print(chunk.content, end="", flush=True)
answer.append(chunk.content)
Originally, the code only had one for loop which retrieves the content from a stream and appends it to an array which forms the response of the completion.
My reasoning behind duplicating the for while adding the distinct if block is to prevent the code from repeatedly checking the if block even if the user is not using the newly added --token_usage flag. So instead, I check for the existence of the flag firstly, and then decide which for loop to execute.
Even though my pull request has been accepted by the project owner, I realized late that this way adds complexity to the code's maintainability. For example, if there are changes required in the for loop for processing the stream, that means modifying the code twice since there are two identical for loops.
What I think I could do as an improvement for it is to make it into a function so that any changes required can be done in one function only, keeping the maintainability of the code. This just proves that even if I wrote the code with optimization in mind, there are still other things that I can miss which is crucial to a project, which in this case, is maintainability.
My tool, genereadme, also received a contribution. I received a PR from Mounayer, which is to add the same feature to my project.
feat: added a new flag that displays the number of tokens sent in prompt and received in completion
#13
Closes #12.
This simply required the addition for another flag check --token-usage:
.option("--token-usage", "Show prompt and completion token usage")I've also made sure to keep your naming conventions/formatting style consistent, in the for loop that does the chat completion for each file processed, I have accumulated the total tokens sent and received:
promptTokens += response.usage.prompt_tokens;
completionTokens += response.usage.completion_tokens;which I then display at the end of program run-time if the --token-usage flag is provided as such:
if (program.opts().tokenUsage) {
console.error(`Prompt tokens: <span class="pl-s1"><span class="pl-kos">${promptTokens}</span>`</span>);
console.error(`Completion tokens: <span class="pl-s1"><span class="pl-kos">${completionTokens}</span>`</span>);
}genereadme examples/sum.js --token-usage
This should display something like:
You can try it out with multiple files too, i.e.:
genereadme examples/sum.js examples/createUser.js --token-usage
This time, instead of having to read someone else's code, someone had to read mine and contribute to it. It is nice knowing that someone is able to contribute to my project. To me, it means that they understood how my code works, so they were able to add the feature without breaking anything or adding any complexity to the code base.
With that being mentioned, reading code is also a skill that is not to be underestimated. My code is nowhere near perfect and I know there are still places I can improve on, so credit is also due to being able to read and understand code.
This specific pull request did not really require any back and forth changes as the code that was written by Mounayer is what I would have written myself.
以上是My first open source contribution的详细内容。更多信息请关注PHP中文网其他相关文章！
