8. OpenAI Financial Advisor Q&A Embeddings - Python Tutorial 

hey Larry i have a question, wouldnt it be much faster and less GPU intensive (and maybe more accurate) if you instead of passing the whole video through whisper to transcribe, just get the text transcription directly? (Most videos already have transcriptions or are even auto generated using google's speech to text, which is pretty accurate) - Any reason you are preffering to go thru the whisper route vs just downloading the video's transcription directly? thanks!!

Thank you so much for this continent. I was wondering how to structure WAV files is there a module that I can use with open AI whisper that will allow me to have some sort of data frame, rather than just a string of words

Oh shit, you're here! How would you feel about having your voice cloned for science?

@@parttimelarry lol if this is how I find immortality let's do it ;) I definitely want to learn more about what you've created here

Thanks for watching, much more to come!

Indeed, Larry, have you test to make the part of "completions" with another model than davinci? I'm asking it just because price. You know.

Hi, how did it go?

@@megaedwin2363 still in progress!! Regarding this part of the project I’m wrapping everything up in a discord bot that will answer has a tutor in a online course… keep you posted on final results!!

@@marcysalty Great!!! Keep me posted

Did you give ChatGPT thumbs down feedback so that hopefully OpenAI will fix it?

@@NickWindham Oh absolutely. I would be surprised if anything changes. I asked - dispassionately and objectively - over and over for it to "just focus on these categories." To which it replied about perceptions and biases. At this point, I think it is just a Google Copy and Paste ... thing. I don't see it doing anything more than that. No actual analysis or connecting the dots.

Thank you! This is very kind!

I have made some videos on CCXT before, which supports many exchanges. It looks like there are some recent code merges for CCXT on Github that support the new Coinbase stuff. So it should just be a configuration option.

You don't necessarily need questions in advance, but you need to divide up your text in a logical way. If you check the OpenAI cookbook, they have an example using Wikipedia articles and asking questions about the Olympics. In this case, they use the headings + paragraphs to divide the text and finding the section that is most relevant to the question.

There are some great tools that handle common patterns like this that I need to cover - Langchain and Llama-Index.

Many of the timestamps in the video are not full questions. There are many timestamps with 1 or 2 word titles like "Market sell-off", "Tax Strategy", etc, so I thought it made sense to check a combination of the question + the answer in case a user question was answered but wasn't directly contained in a timestamped question.

@@parttimelarry Maybe a third useful way, would be to add 2 more columns at the CSV of contexts: one to store an ABSTRACT done by davinci of the column CONTEXT (the transcribed audio between 2 time marks), and another column with the EMBEDDING of that ABSTRACT 😁 It probably would be useful to "have more clear" which row (Q&A) are closer to a new user question. And using these 2 "abstract embeddings" the request to davinci model will be quite more short and so quite more cheap. It would be needed to test it, of course. Maybe you would lose too much useful information to build the answers... who know.

This can be done in a few ways - 1) with the RU-vid API, 2) with some screen scraping , or 3) By hand :). I can touch on this when I should how to process in batch.

@@parttimelarry right click inspect or ctrl shift i var scroll = setInterval(function(){ window.scrollBy(0, 1000)}, 1000); window.clearInterval(scroll); console.clear(); urls = $$('a'); urls.forEach(function(v,i,a){if (v.id=="video-title"){console.log('\t'+v.title+'\t'+v.href+'\t')}});

The code seems to break only when I try to find similarities from a df just by loading the csv file with embeddings. When I create the csv file with embeddings from a data frame that data frame seems to work with finding similarities.

Thanks for the feedback, this gives me an idea for a cost calculation video. The embeddings calls I use in the project are very cheap. Will do the full batch of podcasts and show my costs for a large project. Also planning to do some projects with open source alternatives to compare results. Cheers.

@@parttimelarry I think gpt3 is overkill for this task. One reason embedding are great is the cost. Summarizing text should be fine for Currie, or a cheaper model even. BTW, what would you do if the text was a book without distinct q&A? I guess you need to determine how to split the text. You could automate the process of splitting using another embedding model maybe. Make an embedding of each sentence and then split into contexts based on the similarity of sentences? Hmm, interesting

@@rshrott for books/documentation, i suppose that probably could be useful to treat paragraphs in the same way Larry has worked with the CONTEXTS in this video (sentences between 2 time marks), and the CHAPTER of those paragraphs be indexed in the same way Larry has indexed the YT URLS 😁 Indeed, if you think it well, "to know the question" is no so important. You only need to find "text the most related to the question the user do" and then ask to the model to use it as a context to return an answer to the question. Sincerely... i want so much to try all this by myself...!! super! The applications are endless... finally WE HAVE A SEMANTIC TEXT CALCULATOR !!! It is our last dream for those of us who have been in this AI since the 90's.

@@parttimelarry yes cost calculation become extremely interesting. It looks like one could extract the text, let it summarize with a cheaper model and then do the embedding. That way you also limit context as you will paste less text in the prompt. Thanks for the best video on embedding and specifically how to feed the context back to the model (and I watched half a dozen)

It's just a variable that I defined closer to the top of the notebook. I set it to text-davinci-003 (the OpenAI model) but you can change the value there to use a cheaper model if desired. COMPLETIONS_MODEL = "text-davinci-003"

@@parttimelarry d'oh! 😉 Excellent vid as always, super stoked for your new AI channel

Totally said model a few times in the last few videos where it wasn't appropriate. Noticed this later but hard to go back since it takes a long time to record and edit.