Suggestion for the best OpenAI API model to get the summary from teams meeting transcript

Community
1

Hi All,

I am new to this.

I need a recomendation which OpenAI API model will support me the best for the following:
Need to generate the summary from Teams meeting Transcript using the OpenAI API model.
Per day there may be of more than 10 meetings
Each meetings of size 7893 words or more than that(without the time stamp) rough about 46 KB.

Currently tested with “text-davinci-003” by supplying the chunks of the meeting transcript.

Thank you!

2

Hi @djeanthisivaradjou

Welcome to the OpenAI community.

Based on specs alone, gpt-3.5-turbo-16k should be sufficient for your requirements.

3

Hi Djean, curious to learn more so I can help. Feel free to message me.

4

Just to confirm , I am calling the OpenAI API in developer mode as given below using power automate flow
api.openai./v1/completions

{
“model”: “text-davinci-003”,
“prompt”: “@{items(‘Apply_to_each’)}”,
“max_tokens”: 1000,
“temperature”: 0,
“n”: 1,
“stream”: false,
“logprobs”: null,
“stop”: null
}

Please let me know if the above idea for the requirement of huge meetings more than 10 per will work.

Also any idea how will cost and give me the insight of this.

Thank you!!

5

I am calling the OpenAI API in developer mode as given below using power automate flow
api.openai./v1/completions

{
“model”: “text-davinci-003”,
“prompt”: “@{items(‘Apply_to_each’)}”,
“max_tokens”: 1000,
“temperature”: 0,
“n”: 1,
“stream”: false,
“logprobs”: null,
“stop”: null
}

Please let me know if the above idea for the requirement of huge meetings more than 10 per will work.

Also any idea how will cost and give me the insight of this.

Thank you!!

6

See the GPT Guide for how to get started with GPT, which uses a slightly different syntax than davinci completion model.

If you do run into length issues you may need to summarize the document in chunks based on length. Then you can summarize the summaries if needed.

1 Like
7

You’ll want to use the “chat completion” API.

You can give it two prompts; a system prompt that says “please summarize this meeting” and then the actual text as a user prompt. (Or each utterance/line of text as a separate user prompt, tagged with the user name)

Also, if the transcript is longer than the 16k context will allow, you can put in the first half of the text, and ask it to summarize so far, and then put in the summary plus the second half and ask it to update the summary. I do exactly this, and it works great! (With gpt-3.5-turbo-16k and chat completion)

4 Likes