We want your feedback: Help us improve OpenAI's documentation and resources

Hi everyone,

As you know, here at OpenAI, we’re committed to making our platform and APIs easy to use for developers of all skill levels. One of the ways we want to do that is by providing comprehensive documentation and resources for our users.

We’re interested in hearing about any specific topics or use cases that you think should be covered in more detail, as well as any general feedback on the structure and organization of our documentation.

Additionally, we’re also looking for suggestions on other types of resources that would be helpful for developers and users working with OpenAI, such as tutorials, code samples, or videos.

If you have any thoughts or suggestions, please leave a comment below. We appreciate your feedback, and we’re excited to work together with the community to make our documentation and resources even better.

Thanks in advance for your input!


(1) fine tuning best practices or case studies - or bad practices, (2) end to end examples for building openai API into streamlit or gradio app. I’ve built one for gradio & streamlit and wished there was some guide beforehand.


Long chats management. As a discussion becomes longer, the number of tokens exceeds the limit. What would be a good strategy for this?


More on frequency_penalty and presence_penalty … I’ve seen people using 2 as a setting for the former because the docs aren’t clear that you need a value from 0 to 1 for the API… at least I think that’s the case?

1 Like

I am facing mainly two issues:
(1) Same sentences are repeated multiple times. What to do to avoid these?
(2) Sometimes, we ask Codex to generate in a particular language and it continues to generate in multiple programming languages. How to avoid this?

Please incorporate tips to avoid these problems in your documents. It will be of great help.

To add support for those that use Ruby, I made an OpenAI quickstart public template that has been cloned quite a few times already.

Might be worth making it the official quickstart for those that use Ruby/Sinatra: GitHub - aliyeysides/openai-quickstart-ruby: unofficial openai quickstart using ruby

I have been using codex for code generation and pretty happy with the result.
but would like suggest the following

  1. is it possible to capture user’s feedback by going beyond a binary thumb up/down vote?
    In the playground, should give user an option to submit valid code or comment,
    those feedback would provide valuable data to fine-tune your LLM model for future improvement.
    As an example, when I generate SQL code, the statement is often embedded in Python,
    experienced developer would pick up what is needed, yet this would be confusing for beginner.

  2. is it possible to add a post-processor in your pipeline to remove duplicates?
    I also encounter duplicated response, which is very annoying.
    This post-processor could pick up embedded content from the raw response (see above suggestion)

  3. is it possible to add a search feature in the 30-day history
    I log request/response into a local DB, but the 30-day history seems to be too short when using playground, it would be good to have a keyword search in case the history is very long

Is your app for gradio shareable? I have built a streamlit app here - GitHub - wgong/gpt3sql: use GPT3 to generate SQL from text you can check it out, Thanks

I just moved to streamlit as well :sweat_smile: here’s mine: GPT-VC

1 Like

How about we start with the basics - make it so that people in your support team actually respond to anything.

1 Like

Please, enable davinci-002 and 003 for fine tuning!

Would love to see more documentation on how to build solutions (Q&A, chats) that can constrain responses based on a restricted set of documents - if this is possible. Or pointers to strategies, approaches how to march down this path. And documentation on what is NOT feasible or reasonable.


Document the exceptions that the OpenAI library raises and the reasons for raising each exception.

1 Like

Greetings. Nice products. Suggestion: on /image (and perhaps other endpoints), have an explicit code for mesasge too large. Perhaps 413 Payload Too Large would work here? That way our code could try to automatically trim prompts or warn users without having to parse message text for string at the end.

The conversations are a great resource of information and it would be significantly more convenient to keep these chats organized if there were a system in place for this. Folders and subfolders, tags, a searchbar to look up a spcific chat, etc. :smiley: Hope you guys consider including this feature.



Creating a premium subscription membership $10 a month if you have 100 avatars to choose from such as Albert Einstein Mark Twain Queen of England Steve Jobs Sigmund Freud…it’ll be worth it can generate billions of dollars per year

here are the main pet peeves I have with OpenAI’s chatgpt:

  1. no reference list. You need chatgpt to give annotated references for books/magazines/etc for the sources that resemble the question that has been asked
  2. no empirical/skeptical mindset. You need to build in a scientific mindset to chatgpt, with the ability for it to cross-check references on the answers that it gives, and build in - at a very low level - the scientific method into the way that it views the world.
  3. you need to use mathematical and scientific software to generate thousands of datapoints for questions that are directly related to how the world functions to better tune its scientific understanding.
  4. you need to be able to turn chatgpt around, so that a teacher/professor can give it questions and its goal is to be an interface that asks students followup questions to record their understanding of the material.

#1 and #2 and #3 are crucial for reliability of the answers given and ground the results that chatgpt gives back to reality, and #4 is necessary to prevent it from being used as tool to spread ignorance by widespread plagiarism. In essence, it turns the plagiarism issue on its head, providing educators a way to use AI as a tool to foster new knowledge rather than as a tool for students to game the system.

1 Like

The AI should be able to send emails with attachment or links. Or should even send a text message to the registered phone number