Fix 0xacx/chatGPT-shell-cli#16

[mountaineerbr]

Add session option to follow up questions as per Issue #16 .

A new command, session and its toggle session!, is added. It starts a new chat session in which previous Q & A context is buffered and sent with new prompts.

We try to respect $MAX_TOKENS by removing the oldest text from $session_prompt.

[se7en-x230]

Perfect @mountaineerbr

Thanks for the fast response

[se7en-x230]

Can I set session as the default startup option?

[mountaineerbr]

Can I set session as the default startup option?

This patch really may require a little bit more testing. About setting session mode as startup option, as this is just a testing patch, there is no obvious way of doing that other than entering the session command, however you can start the script such as session_set=1 chatgpt.sh to get that set at startup. Cheers!

[se7en-x230]

Thanks that's great

that seams to work fine

[se7en-x230]

Also one more thing If image: defaults to the browser (which I prefer anyway) wouldn't it be better to skip the y/n question? What else would I want to do? Sure open the created image in the default browser!

[0xacx]

Nice work @mountaineerbr ! Thanks for the PR! Is this working perfectly? I used your approach for the Q&A and the prompt trimming in my PR, nice stuff! 👍

@se7en-x230 About the image, in my use case because I use iterm and imgcat the image is displayed in terminal by default, so I usually select no in the question unless I want to save the image. If you want to directly open the image in your browser after generating it, you can remove this code

        if [[ "$TERM_PROGRAM" == "iTerm.app" ]]; then
            curl -sS $image_url -o temp_image.png
            imgcat temp_image.png
            rm temp_image.png
        else
            echo "Would you like to open it? (Yes/No)"
            read answer
            if [ "$answer" == "Yes" ] || [ "$answer" == "yes" ] || [ "$answer" == "y" ] || [ "$answer" == "Y" ] || [ "$answer" == "ok" ]; then
                open "${image_url}"
            fi
        fi

and replace it with xdg-open "${image_url}"

[mountaineerbr]

Nice work @mountaineerbr !

Hey @0xacx . The patch is really just an idea of how such a feature can be implemented.

I also get an empty prompt sometimes, it seems it has something to do with temperature settings in those cases (probably just try to set temp to 0 [from script defaults 0.7]). The Q&A format is a suggestion because there is an example in the API docs in which they use Q&A.

[mountaineerbr]

It seems that extra newlines or spaces at the end of input text may make the reply empty. May want to remove them https://community.openai.com/t/empty-text-in-the-response-from-the-api-after-few-calls/2067/11

... or even add one ending space: https://community.openai.com/t/api-call-return-empty-but-works-in-playground/17734/4

[se7en-x230]

I have done some testing.

With the session_set=1 I have a conversation over about 20 min around 20 - 25 follow up question with long questions and spaces at the end of the question. No crashes at all.

Starting with -c option Random chashes.

Could be pure luck but should be tested further

[0xacx]

I just merged the changes that enable chat context. I used @mountaineerbr 's approach on building the chat context in memory but also added a prompt to guide the model and enabling with a single flag when starting the script. It seemed to work consistently but let me know if you run into errors. Thanks for the help! 🙏

[mountaineerbr]

Well that is very weyrd as a pre-promt may trigger English as defaults language and there may be some users that prefer to use another language

[0xacx]

In my tests the model replies to the language that you ask the question in. But even if that happens a few times I think it's a small price to pay for having a better chat experience overall, especially considering the majority of users use English. Thank you for the suggestions and help in the PR! 🙏