janek's website

sam local invoke "Dalleanek" -e tests/jtest.json

There-is

https://github.com/janek/dalleanek - main and dev branches

deno deploy - strangely doesn’t show the dev deployment

Old

http://github.com/janek/dalle-lambda-telegram-bot (old repo)

serverless dashboard

has an “interact” tab

CloudWatch dashboard (logs)

Testing and debugging

open Home → Projects/dalleanek in Arc for all links

use serverless dashboard for test events

write a message to the live bot as a test

Deployment

serverless deploy

older

pip install -r requirements.txt -t vendored

cmd+shift+r → serverless invoke local --function post --path tests/event.json (~2-10sec)

(doesn’t work as of 3.05.23)

quick open websites cmd+shift+k lambda, cmd+shift+k logs

One-time config

get telegram token from Botfather

set telegram webhook with curl

get Dalle token from “tasks” thing in Network tab after generating at labs.openai.com

set a Lambda function URL: https://docs.aws.amazon.com/lambda/latest/dg/urls-configuration.html

3.05 log

trying to use openai’s official package

remembered how hard it is with packages for this
noted that you can just make a web request instead of using the lib. should be way easier, no?

Debugging lambda infinite loop on content filter failure

Debugging Lambda Infinite Loop

Status log

V1: make.com + lambda

getting images from DALL•E works

error handling is problematic

V1: loop completed: whitelist checking works, generating images and getting them in a message works

V2: lambda only

connected AWS in VSCode

set webhook in Telegram?

test if make one works
test if there’s a getWebhook

went a tiny bit too far with a tutorial that didn’t work, went back to a simpler tutorial that looks good

ok, lambda gets a message and responds

realized that one “life” of a function doesn’t have state and such - e.g. can’t deal well with someone just spamming. Any state would have to go through a DB, which can be fine, but keep it in mind

how to test locally?

not sure how to run it locally if it’s webhook-based, but can’t be too hard
but if it runs, does online have to be disabled, will there be 2 webhooks?
don’t need a full dev version to test, just define a test event locally and run it
serverless and passing event wokrs. Had to print an event into the bot to know what it looks like, but worked well. For some reason cannot save stuff to return, but

Technological debt

Dependencies are added in two different ways: vendored and layers (see )

python telegram bot library doesn’t have a pre-made ARN available
dalle2 library has to be vendored because I’ve made a lot of changes
unfortunately, PIL cannot be vendored (see SO). I tried adding to serverless config, but it didn’t work fast enough, so I used Klayers

Relevant lambda settings - perhaps they can be exported, but at the moment aren’t saved anywhere

IAM user creation
create function from serverless based on the tutorial linked in this article
throttling and concurrency
retries (warning: this only applies to async, so I think it doesn’t do anything here)
timeout
→ they can most likely be defined in serverless

TODO:

maintenance into the bot

whitelisting into the bot

use files (from Johannes)

/g as alias (from piotr)

gf for files?

error handling

authorization error (bad token)

test

any OpenAI error
low credits
don’t worry about erorr handling for now, also consider they need to be converted to json for lambda anyway

Perspectives and ides

this is for fun and don’t put in too much time (?)

should be against TOS, but is it really? discord as the only source

Grudzień wrote to OpenAI about GPT-3, it’s easy and good to write (from privaterelay or svt, write it first though)

write to the person that’s in charge of the artist program
wanting to be part of the community, and also wanting to be early

If the bot is really good (inpainting and stuff) there could be a place for it when API is eventually released

save to gallery is a big complaint at the moment
especially for variations, making a nice “board” image as an option could be very nice
https://dallery.gallery/dalle-openai-features-ideas-roadmap/ read again

Idea - stickers

Idea (Dave) - generate personalized telegram stickers

Docs drafts

/info and /start

Usage:

/info shows this message To generate images, type “g” followed by your prompt. For example: ”g a man frustrated that he cannot find wheels for his wood project, painting by vermeer”

Privacy:

This uses external services and is not fully private. OpenAI automatically shows me the last 50 generations whenever I log in (but not who generated them). Metadata is saved by a service that made it faster to build bot, but I don’t want to and won’t look at it.

ADR (Architectural Decision Record)

tempted to use Integromat/make.com, because it seems like it could be fast

also considering using code, as then I could share it on GitHub

after fiddling in network tab in console it seems like actually making requests and getting images could be a bit complicated

looked at Ming (1998)’s project again and found that it uses an npm package

code seems like the better solution now, since the npm package could save a lot of work

where to host it if it’s in code?

found https://github.com/python-telegram-bot/python-telegram-bot/wiki/Where-to-host-Telegram-Bots and it seems like there are good options for hosting. Run locally now and worry later

prob serverless (aws, google, vercel, cloudflare) or raspi

if it’s in a serverless function anyway, maybe integromat + cloud func would be fastest?

only write openai adjacent code for now, not telegram adjacent

setting up a cloud function for this turned out annoying

(seems like) google doesn’t actually have authentication, so it would have to be public
AWS works but I couldn’t get myself to do it without having a smooth known process

reconsidering code-only or code-mostly

raspi
hetzner
cbase?

after a moment, make.com + lambda seems like the easiest option after all. Just need to re-learn and document how to use Lambda

make.com worked out nicely, lambda - struggling a bit to get npm packages to work

actually doesn’t have to be in JS, so I can also rewrite it in python, which probably won’t have deps

V1 Stack: make.com + + https://github.com/ezzcodeezzlife/dalle2-in-python

make.com worked great (this time, finally) for telegram bot interaction and whitelist checking
is used to run code that calls the dalle API and returns image URLs

uses https://github.com/ezzcodeezzlife/dalle2-in-python - pretty simple, but saved a good amount of time vs reimplementing

make.com downloads images and sends them via telegram
make.com could maybe also be used for the part that’s on Lambda - but no big point in doing that now

Reached operations limit on Make

turns out that they count individual operations, not executions of a scenario
turns out they don’t have a monthly plan despite displaying all prices per mo
still happy that I went with Make, was very much worth it as a prototype

Considering Lambda as hosting and Python or JS

Lambda because already using it, and because serverless is so much easier (considering boulderbot)
Python because already using it and know it better and to make a quick decision