jschuur / chatgpt-repl Goto Github PK

Code and text completion.

What would the value of this be? Would this fit in well enough with the current structure?

Use streamed-chatgpt-api.

See also: #46.

• stream responses as received
• move response summary down below the response
• add jddev273/streamed-chatgpt-api to stack in readme
• option/command to not stream?

Play back response with text to speech.

Use https://github.com/Marak/say.js?

Check out https://beta.elevenlabs.io/speech-synthesis (amazing audio, pretty expensive though).

• command to play back last one
• option to ply back all responses (not on by default)

https://github.com/privatenumber/cleye

Change settings ($temperature=1.7) or issue actions (!reset or !retry) for the prompt to affect the current session without exiting it.

https://circleci.com/blog/testing-command-line-applications/
https://javascript.plainenglish.io/how-to-test-a-node-js-command-line-tool-2735ea7dc041
https://medium.com/@zorrodg/integration-tests-on-node-js-cli-part-1-why-and-how-fa5b1ba552fe\
https://www.youtube.com/watch?v=ngbL6x4ma3Y

https://platform.openai.com/docs/guides/chat/response-format

• save conversation to config store
• .conversation command asks if you want to create a new one or switch
  • creates a new one
  • will list previous ones with an inquirer auto-complete prompt if you selected pick
  • alternatively .conversation new or .conversation pick does either right away
• conversations command to list conversations
• option/command to keep conversations ephemeral
• some kind of quick way to toggle between recent ones or by number?

See also: #30.

For performance reasons, changing the system mid-session clears the conversation history. But this means you can't use the .retry command to rerun a prompt after you used .system (up arrow still works though).

Ideally, save the last prompt separate from the conversation history, so the retry command behaves consistently.

https://platform.openai.com/docs/guides/speech-to-text

Upload local audio file and use for transcription (or other prompt)

Seems to exit the chat loop if I try and use an ora spinner.

Have raised an isssue about this on ora repo.

Set variables and then use them, in a prompt.

Use case: reusing text in similar or related prompts.

• load variables from a default file or specify a file to use
• prompt then can interpolate them via curly braces with handlebars
• command to list current variables\
• set variable based on all the contents in a file
• save response to a variable?
• conversation specific variables?
• specific system variables like '_prefix' or '_suffix' that are always added to the prompt

Could use cli-highlight.

Would need to know the language for each code block (or assume a default). Could parse it out of the response text or use a (setable) default?

From REPL save interactive conversation to a file

• provide file path in command line option or via command (off by default)
• optionally append or create file
• 'raw' formatting option, where it doesn't include prompt or response prefixes ('You', 'ChatGPT').

The try/catch doesn't seem to show all non-401 status code errors. It hits the catch block, but ignore this part, leading to a confusing blank response when there's an issue with the openai.createChatCompletion payload.

Just doing console.error(error) at the top of the catch block shows the error though.

PR is merged, but currently still in development branch.

Take a prompt from the command line and output only the result with no additional UI/text (suitable to pipe into a file/another script/tool).

Should be able to hit up arrow when starting a new session and get the last text entered.

• save the readline history to a file and load it back for the next session
• CLI option to disable this
• disable if leading space
• prompt command to enable/disable this
• set max history buffer size
• rename .historylength (.conversationlength?) to avoid confusion
• custom history file location

• .apikey [string] lets you set API during a session.
• .apikey shows the last 6 digits of the key for security reasons
  • update .settings so it doesn't show the full API key

Initially, .models command just shows current internal list of compatible models (KNOWN_MODELS).

Related: #21 will let you update models from the API.

Make a list models API request when a key is added to verify that it's valid.

This is the REPL prompt that currently says 'You:'.

Template with variables for settings, usage, time etc.

Also allow this for the startup message?

Minor issue: When exiting ChatGPT REPL by just hitting the enter key, the last prompt field shows 'undefined'. This is likely an issue with clack.

Ideas after #29 implemented:

• archive a conversation
• conversation specific history length
• tab completion when selecting a conversation
• remove messages within a conversation?

Is this a 'response in this language' feature or 'translate this exact text to this language'.

May make more sense as an interactive translation feature.

Would just add 'answer in {language}' or 'translate to {language}' to the prompt.

Requires #20

• .usage generates sparkline chart
• specify recent time period

• add commas to token numbers
• always show at least 2 digits after the decimal points
• align the currency periods even when it's more than 1 digit before the dollar amount
• show session usage too

Turn this:

Session cost: $0.0008 (377 tokens)
Total cost:   $10.1 (57619 tokens)

Into:

Session cost: $ 0.0008 (377 tokens)
Total cost:   $10.10   (57,619 tokens)

• on first use, update the list of models via the apI
• store this lis in config
• validate against it when switching models via .model command
• also periodically (once a week?) update this list
• .models update command forces an update
• when you try and switch to an unknown model, remind about .models update` (but still allow it?)

Better set the style of the response by changing the default from 'You are a helpful assistant'.

Take in a new string as a CLI option or a mid-conversation command (#3).

Tab complete from list of known models

During a Chat Completion API request, hit Ctrl-C to cancel it.

• apply to streamed requests
• apply to async requests

Log usage into a SQLite database for per day/mode/API key. Can then show a usage sparkline.

Good excuse to use Drizzle (after migrating to TypeScript).

Ideas:

• any code snippet in the response gets saved to a file
• .gpt-output Subdirectory of local folder but default but support override
• handle file overwriting (date stamped or incrementing version number subdir?)
• detect file extension if response provides it in the markup
• one file per code snippet or all in one file? Pick one and have option to change default
• not on by default, but add .savecode command to do this from the REPL if not on
• specify default file extension to use (.savecode js)

• allow new commands to be defined via a shorthand alias: .alias gpt4 '.model gpt4'.
• list all aliases
• read aliases from a local (or global) .chatgpt-repl.json config file.

• look for settings in ~/.chatgptrepl.
• merge in potential local .chatgptrepl
• config file settings should still have lower priority than CLI options or environment variables

Streamed responses don't include usage info in the response. Would have to calculate this via [tiktoken])(https://github.com/dqbd/tiktoken) e.g., but that would be a big dependency for the project.

Alternatively, count chunks manually for rough estimate (more info).

• estimate usage
• mark usage as estimated in usage history (assumes #20)
• update Readme section when implemented

This needs more research. Ideally, you just run a .continue command and it picks up where you left off, or asks you how much you want to (temporarily) raise the max token amount by.

• .run [filename] uses text from a local file to run a prompt
• .load [filename] adds the contents of the file to the prompt, but you can still edit it

Inspired by https://github.com/BuilderIO/ai-shell

Do I need to shut down readline entirely during this time?

Doesn't quite seem as easy as chatgpt-repl -m text-davinci-002 e.g..

Check if other models have a different API/parameter set.

GPT-4's pricing is different from GPT-3, so this will need consideration in the usage tracking.

Total usage tracking should track usage per model too.

Hitting something like shift-enter should insert a newline into the prompt, rather than it sending off the response.

This might only be possible in GNU readline and not the Node.js one.

The history file should encode the newline in this case.

When it outputs a list (but not just for lists?)

I usually use commander, but this time around, add better command line option handling of default values, the help and the version commands by using cac.

If a response contains a code segment, use a commend to copy just that

Use a regexp to extract based on triple backticks?

• '.copycode' copies all async for async requests
• review how this might be more complicated for streamed responses (possibly create new issue)
• Create separate issue for advanced ideas: .copycode 1 copies the first block (use parse-numeric-range)

Related: #32

• show month to date
  • compare to previous month (with % difference)
• show usage today
  • compare to previous day

Requires #20.

In true REPL style, navigate the previous user inputs via up/down arrow. Use cliffy?

Response is 'unknown'.

https://twitter.com/sandbags/status/1632358924664856576