Seve - @Abse this isn't working on prod, it's b...
@Abse this isn't working on prod, it's been there a long time, I don't think we should ship broken features- we need to decide if we're going to remove this or fix it
39 Replies
so I've tested it for a while on local and the hallucination is very high I'm not sure we can fix it , I tried lots of message but nothing fixed the issue , wdyt ?
can you figure out how to get it fixed on prod
btw, we should never ship a broken feature to prod this long
like you just want it to work ?
yea i want to see the feature
yeah
and i don't want useless buttons on prod
just annoyed because probably lots of users are pressing this going "wtf"
i want to see this feature
I mean if you add an API key for openrouter I think it should work an we can check if we want to keep it or remove it
where does it need to be added
does an endpoint need to be implemented?
no one sec
VITE_OPENROUTER_API_KEY
this is the .env
fake-snippets-api\routes\api\autocomplete\create_autocomplete.ts
the end point does exist
on fake , not sure if you guys implemented it on the backend tho ?
+ we need to guard against spaming right ?VITE_OPENROUTER_API_KEY doesn't make sense
that would be insecure
it should just be OPENROUTER_API_KEY and should not be passed to the frontend
yeah it's not passed
it can never be passed
that should be removed
or renamed OPENROUTER_API_KEY
I mean the api key in .env and the end point in the backend
yeah I can rename it
yes
VITE_* always means that it's being passed to the frontendoh
there's no implementation on the backend for create_autocomplete rn
i also noticed that when i hit the bot my editor started to lag a lot
but idk maybe unrelated
could be a rendering issue
it does work after renaming
working locally doesn't matter at all to me
like the only way it helps users is on prod
you don't have backend access so it's not really your fault, but it shouldn't have shipped without working
ofc so how do we move it from fake to prod never worked on the back end I remember you talking to @DOPΣ about it IIRC
yeah for sure
yea i'm frusterated but it's clear that it's more or less a project management issue
that was a mistake
but it's very frusterating still haha
true it is
we needed a checklist issue and delegation for the different parts
ty
no problem
btw this prompt should work much much better if adapted https://docs.tscircuit.com/intro/quickstart-ChatGPT
Quickstart ChatGPT | tscircuit docs
Preview and iterate on tscircuit designs directly inside ChatGPT using a small CDN bundle and a simple repo convention.
i noticed you're loading the props readme, that's super slow and way way too verbose
we've experimentally tested the chatgpt prompt in benchmarking
ok let me do this
so we don't need the props anymore just this promt and do more testing ?
for prompts we have a repo that runs benchmarks https://github.com/tscircuit/tscircuit-prompts
GitHub
GitHub - tscircuit/tscircuit-prompts: Devvelopment of prompts that ...
Devvelopment of prompts that enable AI to write tscircuit via context - tscircuit/tscircuit-prompts
so we'll always use stuff from that repo and benchmark it
does it accept openrouter api key ?
let me try to use it
uhh good question
we might also want specific benchmarks for autocomplete, but it is good
note that gpt5 is much better than 4
we would need to add support for handling either OPENAI_API_KEY or OPENROUTER_API_KEY
I thought that endpoint is totally spammable so thought to wait to learn something like lsp
You know?
Or should we go with basic endpoint? With rate limiter and account credits integration
we don't need anything fancy, we have redis so we should be able to rate limit
Hmm, if nobody working I can give it a try on api prod. After Wednesday
i think i have an implementation soon via codex
but let's review weds
Yeah more options better, public codex ig if