💬│general

[warm cave]

they seem to add models when they are all over the news and twitter, and trim them down after the hype is gone. but if you have a different theory then i would be interested

Yeah, but you are looking for free, not many options, just giving one option

[cloud elk]

to be honest i believe the addition of labs to the site could be just for people who want to test the models. In my opinion it probably is not a marketing tactic.

also i think the reason they may be removing them is probably due to costs(?) but then again i am not entirely sure myself

[warm cave]

I just dont see a benifit to paying api usage if not for the sake of gaining new users

and limiting users to 50 opus per day while also giving free usage of different models to non paying users, haha

just a bit odd imo.

[cloud elk]

Opus is expensive to run. it's really not that hard to figure out.

[warm cave]

Well yeah its expensive, but if they are struggling to afford what they promised to users, why give away free money on the side, even if its smaller compared

[cloud elk]

huh.

[warm cave]

✨ Marketing money ✨

it cost money to give haiku at 200k context, and command r and dbrx is not particularly cheap. most sites give 7b as free models

[cloud elk]

Their company their decision i suppose

I will admit it though from the things you have mentioned yes it does seem kind of odd

[compact panther]

I ve had nothing but issues with Perpkexity Pro for the past few days. Hallucinations, restrictive quotas, all sorts.

I paid for pro today and I’m trying to get a refund, but even the info for accessing support channels given by Perplexity was inaccurate. I can’t even access the pro channels here, using the info perplexity gave me. Demoralising experience.

@frank rivet @signal hamlet
Can you help with this?!?

[warm cave]

Did you email support, I can give you the email if you need it. Also the pro channel can be visited if you join the discord using the link under your perplexity setting page (unless the bot is broken again)

📨 [e-mail]

[compact panther]

Yes, emailed support and the pro email address as well. I tried the discord link the way you mentioned. No luck!

Thanks for getting in touch so quickly.

[warm cave]

No problem, I'm juts a user but I've been here for a while, haha. @halcyon coral @serene sand @wraith crow it looks like the bot is broken again, he can't get into the pro role

[compact panther]

Thanks @warm cave

[cloud elk]

I don't know why but i find pinging people randomly a bit rude

just me tho

[wraith crow]

added you manually to pro #👑│pro-lounge you should have access now

[compact panther]

Thanks!

[warm cave]

The mods are here to help 🙂

[cloud elk]

yes i do understand haha

[wraith crow]

I will message someone from the team about the issue thanks for mentioning it. The bot has some issues lately..

[umbral blaze]

Is anyone else having the same problem as me? Perplexity is unable to read news from the Discover tab

It's like the chat can't see the news

[warm cave]

what do you mean? like when you ask it about a news topic it dose not know anything. can you share a thread?

[fast meteor]

or a screenshot

[warm cave]

Yeah, that too

[fast meteor]

mine seems to work fine

[umbral blaze]

Ask him to translate it into any language

He'll say there's no news

[fast meteor]

translate all above in espanol

you need to give it a little bit of contexts

[umbral blaze]

On mine it ain't working

It translates anything that has nothing to do with it

[warm cave]

You just need to be a little more explicit, let me try tho wiht the free version model

[warm cave]

can you expand the sources, it may be confusing the model, haiku (Default) is eaiser to confuse

Like here is a failed attempt

[fast meteor]

the default model does not link noticia to the article above...

[umbral blaze]

And this isn't just happening on the mobile version.

But it should probably be a problem with the free version

[warm cave]

I mean you can work around the free version to get it to work, but paid smarter models are going to mess up less

[fast meteor]

i found an easter egg? if you clicks on the icon left to the answer, it will spin

[south kindle]

I mentioned this in #👑│pro-lounge but part of the experience of working with the "bleeding edge" of tech is having to deal with getting "cut" aka experience and deal with problems.

[warm cave]

lol yes

[umbral blaze]

I don't think this is an option. Is it in the browser version?

[vapid onyx]

Secret fidget spinner

[warm cave]

also, what colors should i try that wont look like trash, haha

[fast meteor]

im frustrated that it only spins for 1 circle

[fast meteor]

yes to me

the more i clicked the faster it spins i wish

[south kindle]

muted colors. or pastel colors.

[warm cave]

Maybe we can make one with the brand kit, haha

too muted

[fast meteor]

ask gpt for it?

[warm cave]

GPT4o gave me that trash color, haha

[fast meteor]

and opus?

[warm cave]

Opus, put me back where i started

idk

@tame current @tribal harbor You guys having telepathically communication?

[south kindle]

The Yellow looks NICE actually.

[warped marsh]

perplexity is soo not good for coding

the context window it uses is sooo small

[warm cave]

You use Claude Pro instead?

[warped marsh]

the context they use for gpt 4o is smaller than the one openai uses in chatgpt plus

[warm cave]

last i checked chatGPT is also 32k, but maybe you are using "All" as your focus

GPT4o over api is 128k

[warped marsh]

does using all or writing change the context window of gpt 4o?

[warm cave]

no they are the same, but cramming 5-20 sources per message really drains the 32k fast, so writng mode is recommended for code

[warped marsh]

and also did they remove the pro tab?

I cant see the pro button

yeah it isnt available anymore

[south kindle]

That's CPLX. right click the web access icon

[warped marsh]

oh

[warped marsh]

ooooh I see

your extension is cool though

[south kindle]

when Pro is on, you'll see a blue outline

[warped marsh]

yeaah

[south kindle]

not mine! it's @harsh stag's

I'm just helping

[warm cave]

feline is the GOAT for making it

[warped marsh]

oh then @harsh stag your extension is super cool

also @south kindle can I like go in writing mode so disable web search for one prompt and then enable it again for the next prompt?

[harsh verge]

This is related to #💗│sharing message I've saved your page in a NotebookLM on Prompt engineering, thanks! 😉

[warm cave]

Thanks 🙏

[south kindle]

Web Access off turns on Writing mode, even for threads started that aren't started using writing "focus". and vice versa, you can turn web search back on for writing dedicated threads per query.

[warped marsh]

that is actually a feature I wanted for soo long

[warm cave]

same

[south kindle]

also sharing the new update where you can turn AI Profile On/Off and adjust it from the main search bar. You can also edit any Collection Prompt from this UI as well.

[warm cave]

There are so many new features i might have to make a new page

[warped marsh]

which one does perplexity use?

[fading moth]

I'd love to use CoT but delete the CoT after each prompt to prevent confusion and token use.

[south kindle]

I would assume the 8 or 32K.

[warped marsh]

shouldn't it be 128k though?

[fading moth]

It's 32k

[warm cave]

i think its 32k, def not 128k

[south kindle]

I have CoT built into other prompts so I can "call" it by talking to the Collection directly using the collection prompt. but it does get lengthy pretty quickly.

[warped marsh]

does it use 128k context length if I ask any coding related question?

[warm cave]

I wish, i think the only way to get 128k is with API or Poe

but poe sucks imo

well usage limits wise

[fading moth]

32k is hard limit for Perplexity, it's throttled by Perplexity not GPT

[warped marsh]

wdym?

like its perplexity that is limitting the model?

[fading moth]

They would go bankrupt if they maxed context windows

[warm cave]

well Perplexity will cut off anything after 32k even for models like claude that have 200k default

to save on api cost

[warped marsh]

thats sad

so if I use it for more of a coding related tasks then pplx isnt for me, right?

[south kindle]

Not for one shot stuff

[warm cave]

I mean it depends, chatGPT is not soo different imo when in web off mode

[south kindle]

you'll have to amend some stuff together. likely still "doable".

[warm cave]

claude is better context wise but supper short context window

eehh, said it wrong

claude pro is better context wise, but super tight usage limits, i will get 15 ish per 5h

the main benifit with chatGPT is teh continue button for longer code, but some some that matters and others it dose not

[warped marsh]

Does phind use 128k context window gpt4o model?

[warm cave]

nope, 32k all around

last i checked

[warped marsh]

even phind?

[south kindle]

I know typically with ChatGPT, onces it reaches the token limit for a reply, it presents a "continue generating" option, where the previous code/replie is just continued instead of a new message starting over. With PPLX, once you hit that wall, you can type "continue" but the output will be in a new message.

[warped marsh]

dont they cater coders so they have 128k context length?

[warm cave]

32k

[warped marsh]

oh thats sad again

[south kindle]

They do to the extent that people come on board and use the API i would suspect. the main "use" for PPLX is the answer engine.

[warm cave]

Yeah, i hit a lot oh dead ends looking for a good long context option

[warped marsh]

wdym?

[warped marsh]

yeah but like the problem is not when the output token reaches its end but when the input token reaches its end and the model starts to forget stuff

[fading moth]

Perplexity has nothing to do with Phind

[south kindle]

they cater to programmers in the sense that the API/Labs exists to build.

[warm cave]

Have you tried free Gemini 1.5 pro in the developer console?

[warped marsh]

nope?

[warm cave]

Gemini is not as good, but 1M tokens of context can be very good, but in the free version you dont get the data privacy stuff

[warped marsh]

so what about phind's main web interface?

[south kindle]

I haven't played with their stuff in a while. I know coding was a focus for them though.

[warped marsh]

yeah so exactly my point that if coding was their focus then shouldnt they be using the 128k context window and not 32k?

[warm cave]

Yeah, they should

[south kindle]

I'm not saying I disagree at all. I'd rather have more context as well. We all do. ha.

[warped marsh]

Is it free to use? or does it have daily limit?

[warm cave]

https://aistudio.google.com/app/prompts/new_chat It has rate limits but i have not hit them before

[warped marsh]

what about you.com? Do they use a bigger context window?

[fading moth]

It's a matter of cost not want

[south kindle]

I don't beleive so. but this would be a fantastic Page to have made up for LLM services/Apps and their available models and the stats of those models.

[warm cave]

You.com dose have a bigger window, but last time i used them they would have models down for hours every day, but maybe they got better

[warped marsh]

does openai use the 128k model in their web interface then or the only way to get the 128k context window is through their api?

[south kindle]

i believe API/playground only.

[warm cave]

Last i checked it was 32k, that was a few days after the GPT4o launch, i will check again

[warped marsh]

oh they do?

they also have code interpreter, right?

[south kindle]

You know what these other platforms don't have?

[warped marsh]

what?

[south kindle]

QOL features brought to you by CPLX.

lol. sorry.

carry on.

[fading moth]

Unlimited Opus

[warped marsh]

haha

[warped marsh]

thats truee

[warm cave]

Yeah, but its limited, like its good at solving math, but limited libraries and has no web access so it cant run many types of code. what do you want to use it for?

[warm cave]

Please, can i have some?

[south kindle]

Unless other platforms match the use of Collections (for organization but moreso dedicated collection prompts), i find it hard to fully utilize them daily.

[warped marsh]

uhm I mean I use perplexity to mainly assist me with coding so like a tutor where I ask it to teach me a topic or when I have a project in mind and then I ask gpt 4o in pplx to guide me through that project so that I can learn the stack and all and do the project

[fading moth]

Only 50

[south kindle]

https://tenor.com/view/tinytim-tim-poor-child-boy-gif-7331148

[warped marsh]

I havent used pplx for searching stuff online in a while cause I think it has gotten inaccurate lately? So I just search myself

[warm cave]

So non stand out as a clear winner, all are 32k except for Claude which is 200k but you can never use due to usage limits

[fading moth]

There's a certain level of irony that Perplexity's big feature of pulling sources also takes up a lot of tokens that fill up that 32k window pretty fast

[warped marsh]

so claude is 200k but with 50 usage limit?

[fading moth]

Pffft

I wish 200k

[warped marsh]

oh no?

[warm cave]

no, i mean claude pro has 200k but tight usage limts

[fading moth]

32k like the rest

[warped marsh]

sad AGAIN

are we even surprised

[warm cave]

sad all day

[warped marsh]

and also have you guys also found pplx to be inaccrate at searching lately?

[warm cave]

I wonder if API would be affordable enough for you?

[south kindle]

You almost need a dual system where one holds the entire context, while the one you interact with only grabs what it needs from it.

[fading moth]

I rarely run the search feature in order to save tokens

[warm cave]

use msty with GPT4o 128k

[warped marsh]

I dont think so

[warm cave]

yeah, prob too pricey

[warped marsh]

besides I dont search a lot with pplx anyways

I think google for me is better at searching stuff lol

[south kindle]

it depends on the Task, but with CPLX being able to turn web access on and off as needed helps with this for ongoing threads.

[warm cave]

you can use Gemini 1M free api and then switch to GPT4o api only when it gets stuck?

[warped marsh]

after subscribing to pplx still think google is better at searching online 🤣

[warped marsh]

I mean yeah I can find workarounds when I really do need the context window

[south kindle]

Depends on what your looking for. if it's a broad topic where you'd end up having to look through 10 pages to find the 5 sources you needed, PPLX does that but automatically.

[warped marsh]

it isnt all the time though

[fading moth]

Yeah but if it pulls 20 sources @_@

[warped marsh]

uhm and what type of query would that be?

[warped marsh]

trueee you know earlier it used to pull like 100 sources I remember

[warped marsh]

which website is this?

[warm cave]

its an app called msty

[south kindle]

anything research related. example: we took in a stray cat that had a litter of 6 kittens. I used PPLX to pull from multiple sources to make up a complete "timeline" of milestones to watch out for. plus anything else week to week to adjust to or pay attention to.

[warped marsh]

ooh is it paid?

[warped marsh]

I see

[warm cave]

it uses api keys

[warped marsh]

I normally do searches like

[warm cave]

so a lot of those api's are free right now

[warped marsh]

How do I find if I have a feature in my credit card

and pplx gave me vague information when googling gave me direct answer

[warm cave]

the gemini, groq, pplx (5$ free), and some others

idk, lots of options, hard to pick a good long term tool

[south kindle]

For a more recent "Event" where I wanted to pull updates from all sources :
https://www.perplexity.ai/search/Starship-Test-4-l5qs.WoNSkOmJIXJU85gMA
This saved me from having to individually read each article.
I rely on the <scratchpad-think> prompt/framework to verify what is being pulled from each source.

[warped marsh]

what is scratchpad-think prompt?

[warm cave]

@south kindle You got this one

[warped marsh]

lmfao

[south kindle]

#1223058316662538331 message > Prompt the PPLX community helped me put together to force reasoning and cohesion for longer tasks. Essentially the model dumps it's thought process into it's reply and follows a thinking structure before reaching its "final" output/result. there are a bunch of versions floating around but the one I'm linking too is the most generic/useful one that will fit into AI Profile or into a Collection. (https://www.perplexity.ai/collections/scratchpadthink-wBPEohuUQH6tz5qMlH4F7g)

[warped marsh]

like for example where pplx was inaccurate was literally this prompt. It got lucky at answering it correctly because the context window of gpt 4 and gpt 4o hasnt changed drasically I think but the sources of the middle answer that talks about the 32k window.. if you click them then they lead to the sources that was talking about the gpt 4 model and not the gpt 4o model

[south kindle]

It was featured in the #⭐│starred channel back in April as well.

[south kindle]

hit us with the query you used so we can toy around with it.

[warped marsh]

and how does it help?

[warped marsh]

here : https://www.perplexity.ai/search/What-is-the-21RKWtnqTg6a37M8k.sIow

[south kindle]

By forcing the model to explain it's process, it will include those aspects in it's replies. we've just found it to be very helpful, regardless of model, however it was built for use with Claude 3, based on their own documentation related to reasoning and logic.

[warped marsh]

A more appropriate place for this message would be #💗│sharing , no?

[warm cave]

bro i just got agentic search

just now

[warped marsh]

I have it too

its good I think

[warped marsh]

should I add it for gpt 4o as well ?

does it make the model accurate?

[warm cave]

thats with anti-h

[warped marsh]

btw heres the link where gpt4o got the answer or at least the sources wrong

[warm cave]

(Anti hallucination prompt)

[warped marsh]

what does that mean?

[south kindle]

Also, my favorite part of this framework, is it forces the model to produce "thought-provoking" questions, that almost ALWAYS are vastly better than the autogenerated follow ups. These alone add a TON to the overall use of the tools. plus the questions it presents steer the output as well.

[warped marsh]

I see

[warped marsh]

this makes the model more accurate?

[warm cave]

yes

at least from my testing

[warped marsh]

did the model got the answer right for you in this collection?

[warm cave]

yes

[warped marsh]

ask the follow up like this

and then for me it messed up with the sources

[south kindle]

throwing here so I can reference it. i'll mess with it after dinner.

[warped marsh]

the first answer was right for me too

[south kindle]

also it's more helpful when sending screenshots here to expand the Pro Search area, especially with the new Pro Search. this helps understand the output and how the model reached it.

[warped marsh]

makes sense

but I have sent the link to the thread itself

https://www.perplexity.ai/search/What-is-the-21RKWtnqTg6a37M8k.sIow

[south kindle]

new search may throw a lot of wrenches into standard workflows we use. we are still figuing it out.

[warped marsh]

did it mess up the source?

mine gave the same answer but messed up the source where for 32k tokens line it refereced two websites that talks about gpt 4 and not gpt 4o

[warm cave]

its a little bit of a weird answer, its talking about chatGPT and API at the same time

[warped marsh]

it messed up for you too the 32k tokens is being taken from the 5th source which talks about gpt 4

and not gpt 4-o

and same goes for 128k tokens line of the answer

[warm cave]

there is a problem, my Anti-H prompt was not desinged to work with the agentic setup

[south kindle]

without scratchpad. trying with now.

[warped marsh]

its not the answer though.. its the sources

[south kindle]

Sorry, i'm doing multiple things at once

[warped marsh]

ight ight

[south kindle]

that's what I get i guesss

[warm cave]

hmm, agent search dose not know that GPT4o is a thing, so it corrects it to GPT-4 as if i made a typo

[south kindle]

I understand what were are looking at now. whooosh.

[warped marsh]

you do?

wasuup?

[warped marsh]

🤷‍♂️

in my search it even searched for gpt4o and not gpt 4 but still messed up the sources

[warm cave]

Basically GPT4o dose not know that GPT4o is a thing that exists, it was trained when GPT4 was the best model, so it assumes you made a typo, let me change the question

[south kindle]

What is the cost of GPT 4o's api and in which context length is it available? <only source pages that reference the new model called GPT-4o. this is not GPT-4, GPT-4 turbo, or GPT-4.5>

[warped marsh]

okay yeah

so this is what I am saying

[warm cave]

basicly its stupid, haha. and we have to tell it

[warped marsh]

we figured out a solution to it by doing more work after the model got it wrong and then after spending 10 mins researching where we could have just googled or bing if anyone does that

[south kindle]

This is an issue with LLMs far beyond PPLX though. it's fair criticism, but this has been the territory for a year now with slow and slight improvements over time.

[warm cave]

I have an old ani h prompt that would fix this

[south kindle]

we are hardened to find the workaround, so it's naturally for us to get a bad output and instead of complaining directly, we'll see it as an issue to solve. that's not how everyone is though.

[warm cave]

basicly it tells the AI that the user never makes typos, and that if they said something they want you to find that, not somting close to it

[warped marsh]

but if the topic is new to you then you wont even realize if the answer was wrong in the first place

[warped marsh]

thats better

[south kindle]

Not for people that don't double check the things they read and take everything at face value. but that again is fair criticism that exists for all AI platforms currently. it's why so many models advertise this message.

That being said, when your platform is built on "exploring curiosity", ideally you are exploring accurate information.

this example is sort of "cherry-picked", because searching for anything with a version or "model" will cause the system to pull "related" content when it can't find enough to make a complete reply/output.

[warped marsh]

opus is smart enough to not mess up the sources and models though

[south kindle]

not always, but that's a area where having to specify exactly what you are looking for is very helpful at steering the output.

[warped marsh]

yeah but opus was smart enough to not do the same mistake

[south kindle]

We also have to factor in the model being used for Pro Search, and what it relays back to the main model for it's output. OPUS may just be working better to retain that text for some reason? Or rather, the model running Pro Search could also just be OPUS, so again it plays nicely.

[warm cave]

GPT4o gets tunnle vision, which dose not help

[south kindle]

Regardless OPUS 50

[warm cave]

Opus dose not

as much

at least

[warped marsh]

😢

[warm cave]

Wondering how sonnet dose

[warped marsh]

lets see

it messed up too but not as much as 4o

[warm cave]

One of these times i will have the Ultimate anti-H prompt

[warped marsh]

🤣

[south kindle]

It'll just be like the DAN prompts. you'll have to evolve it constantly as models change and new models based on different training data are released. I applaud your efforts though.

[warm cave]

I will try

[warped marsh]

yeaahh crazy efforts

impressive

[warm cave]

lol

[south kindle]

I think they made it to Dan 10.0 before they gave up or rather I stopped paying attention. ha.

[warm cave]

Dumb GPT4o

[south kindle]

whooops

[warm cave]

lets rewrite with GPT4t

[warped marsh]

🤣

I think models are the smartest version of themselves when you use them without the web search

[south kindle]

Something to try. export the entire thread to markdown and add it as a file for a new GPT-4T or Opus thread. see if it can "find" the mistakes and correct them. (export is part of CPLX, top right of the page. it will copy the entire page to your clipboard)

[warped marsh]

and so do your searches through google like you used to and ask the models questions that doesnt require information that it needs to look up the web for

[warped marsh]

and also the export button doesnt really work for me

I am using the 0.0.0.8 version

[warm cave]

Opus is a lot smarter than GPT4o and GPT4T, but it seems to not get the other option

[south kindle]

trying to influence the "steps" Pro Search uses.

[warped marsh]

oh this seems like a good prompt

[south kindle]

it can be improved. I just wrote it out as I was thinking of it to see if it would work. ha.

If I can figure out a generic prompt just for enhanced Pro Search, we could load that into a Collection and test it out without having to directly input the Pro Steps.

[warm cave]

@south kindle https://www.perplexity.ai/search/Plan-me-a-IQGF0vuqSjy66chq8Bp0yA#0

did not tell me what plane tickets to get

i will try again

[south kindle]

it at least presented the steps

BUT then didn't do them...

[sleek vortex]

funny since mine doesnt do that

and i use a smaller model with some good prompting

lol

[warm cave]

Haha, i was going to ping you bc of that

[sleek vortex]

lol right now just made my aws account

[warm cave]

nice

[sleek vortex]

lambda has a 6mb payload limit

is what i have just realised

[warm cave]

what is all of this embedded stuff

[warm cave]

Oh, then i guess aws is the answer

[sleek vortex]

send thread link

[sleek vortex]

aws is my choice if you saw the convo in the server lol

price works out pretty favourably

[warm cave]

https://www.perplexity.ai/search/Plan-me-a-zuz_rko1Rx.dtej6n9QIog#0

[sleek vortex]

product cards..?

weird

[warm cave]

oh, must have missed that, last i remember azure was also an option

[sleek vortex]

yeah but azure container apps is

eh

aws has best serverless

@agile jay had a good idea

i can just not have rate limits on paid plans

since it's serverless

or set them really high so youre more likely to be bound by your account balence

that would instantly blow firecrawl's 5 per min out the water lmao

[sleek vortex]

i might still have to use azure for phi3

which is the only reason i was considering it

but then again i can just make that crosscloud in theory

[warm cave]

Yeah, true

[south kindle]

trying again

[sleek vortex]

the questions r useless like 90% of the time

it was actully decent when they had the multiple choice options

but now that thats gone its often so useless

[south kindle]

and now it crashed 3 times at this question is presenting this instead

[sleek vortex]

LMAO

[south kindle]

which was the right thing. the step was to review the content so far and create thought-provoking queries for it to then use to search and further improve the reply.

[sleek vortex]

so the biggest website ive scraped was like 128kb

[warm cave]

no matter what i try, it wont tell me plane ticket details: https://www.perplexity.ai/search/Plan-me-a-.WikdPp_SjyQ0v4jL9_Zbg#0

[warm cave]

Was that the hourse one?

[south kindle]

probably related to web crawling limitations.

i know expedia gets them via API (i believe)

kayak also

[sleek vortex]

the horse one specifically the /yesterday

[sleek vortex]

hah

mine pulled average plane ticket details from some good site

it didnt pull live though

[sleek vortex]

(when i did it a while back idk)

rn ive only been working on getting my scraping to mvp lol

[warm cave]

Yeah, funny how capable yours is, and you only started a little ago

[sleek vortex]

pplx added agent pro search after what 2 years

[warm cave]

thats a fat number of sources

[sleek vortex]

still isnt 100% good

[warm cave]

Yeah, but it will be amazing soon

[sleek vortex]

pplx?

or me?

[warm cave]

You are making progress fast

[warm cave]

you, haha

its really good and will be polished after some work

[sleek vortex]

hm

this might be the old python scraper though

[warm cave]

Finally: https://www.perplexity.ai/search/Plan-me-a-H66hf48ARB.yoRq816KO5w#0

[sleek vortex]

go one is slightly better

. do not stop trying until you get the exact plain ticket details. make all decisions for me

this is so funny

[warm cave]

lol, took so much effort

[sleek vortex]

what did i even ask in mine just

plan me a trip to japan in may 2024

not sure if its because if i pre-prompted that query?

no clue

[warm cave]

idk if its hallucinated tho

[sleek vortex]

my friend asked it for like

which questions in some exam's past papers had a question type

it goes and checks like 100 pdfs supposedly

gives full results

i double check it

100% hallcuinated

[warm cave]

lol

😂

[sleek vortex]

i need some links

what on earth

[warm cave]

lol, watches

pokemon

flight lessons for one

there is a lot going on there

[south kindle]

just told it to use <scratchpad-think> for it's Pro Search framework. worth a shot?

[sleek vortex]

send thread links

running [...document.querySelectorAll("button > a:has(.rounded-full)")].map(x => x.href) and getting all these random sources as i need an estimate for average page scraped page in kb

[south kindle]

scan sharing

there's likes 1000 in there

[sleek vortex]

true yeah

[warm cave]

i wish somone that lives ther could tell me if this is BS or not: https://www.perplexity.ai/search/Plan-me-a-dFjEdP5yTs.SQ1hn4HpIEQ#0

[sleek vortex]

would give me a good idea if it works or not too

lots of sites requiring the real browser scraping so might have to get that running too

[warm cave]

like is this suff all good or hallucination

[sleek vortex]

why is the hotel 700 per night

wtf?

everything seems ok except that hotel price…

what

what 7* hotel has it found that costs as much as the flight per night

[warm cave]

hmmmmm.........

[sleek vortex]

either way why would you stay there

like do you want to go broke?

[warm cave]

it thinks i am rich

i will tell it im not rich rq

[sleek vortex]

but then it reccomended a bunch of things on low yen budget

but just the hotel is 700$

kinda funny to me

[sleek vortex]

wtf there’s a map?

or is that some image

[sleek vortex]

looks like a perplexity map though

[sleek vortex]

damn

more features for me to put on the roadmap 😞

[south kindle]

part of their Local Search

[south kindle]

same thing allows you to search your own local area

to see a map (i believe)

[warm cave]

its all yelp

[sleek vortex]

so pulls from yelp api

wow

[warm cave]

think so

[sleek vortex]

wonder if scraping gmaps is possible

their api isn’t that cheap so uh

[south kindle]

#1217939485199958120 message
#📢│announcements message

[sleek vortex]

right before the opus 600 gaslighting event

but ohh

[warm cave]

what...? OpUs 600 was never a thing

[sleek vortex]

i think it was for like a day

no clue

[south kindle]

Opus Unlimited existed for like a week or two at the start.

[sleek vortex]

yeah

[south kindle]

which is part of the uproar right? they "took" it away.

[south kindle]

aka realized how expensive it was. I know when WorldSim was botted by 4chan, there bill for Opus was in the thousands. they had to shut down and relaunch with a credits model. and added in haiku and sonnet to make it cheaper.

[agile jay]

They also made an announcement when they said they are working on getting it back to unlimited...

I'm guessing they'll drag it out until Claude 4...

[sleek vortex]

time to populate this with my todos lol

going to eventually forget things if i keep it all in my head

[south kindle]

that only works if you build in a "review" step, otherwise you get a list that never ends. or atleast that's what my ADHD brain requires.

[warm cave]

Even tho it cant handle a full trip to Japan in one shot, i am excited to have it, hopefully its great for general search

[sleek vortex]

elaborate?

[south kindle]

If you look into the GTD method (it's way old but still useful) it's a way to organize tasks and information so things don't get lost. based on GTD (getting things done) the system only works if you have a review step to go thru the backlog.

[warm cave]

They should make it so the agentic search can see your user profile

But i guess that is where prompt library will shine

like instead of saying i live at x and x. it would just know

[modest token]

hmm... I need to decide which HTTP 404 cat I prefer...
https://httpcats.com/404
https://http.cat/status/404

[agile jay]

Cat not found... But there's clearly a cat in the image...

[south kindle]

Thread: https://www.perplexity.ai/search/Fully-review-the-0RmThWSGTFC4i4kRiQStdg
Page: https://www.perplexity.ai/page/The-GTD-Method-XNa.pCFWT0eNitfO40bqrw

[livid mantle]

scratchpad

[fossil sentinel]

North Korea sends another wave of trash balloons into South Korea https://www.cnn.com/2024/06/09/asia/north-korea-balloons-response-intl-hnk/index.html

😭

The pettiness

[devout bronze]

let's just hope south korea doesn't escalate this to a military responce

[vagrant warren]

The appropriate escalation after trash balloons is, of course, compost-and-cat-litter balloons

[south kindle]

The landfill is full Mighty Leader what shall we do??

[agile jay]

Depends. It might be a plan for a Trojan horse.

[south kindle]

if I was evil, that's a hell of a chemical weapon. load some sars in there or something.

[agile jay]

Yep, and they get used to it, so it gets ignored...

[bitter summit]

Anyone help which my api credits are not being added, I bought the subscription yesterday

[sinful anchor]

What model do you use to translate and summarize PDF reports in @tranquil orchid ? I’m ususing Arabic documents to summarize into English. A few pics, some tables but mostly text

[wide mesa]

Actually a rare case of AI image generation making something that looks really cool

[mystic nacelle]

how to get chatgpt api

[fast meteor]

mind to share the prompt & images in #1194788138124587128 ? pretty dope indeed

[spice lance]

what's the logic for perplexity of whether it uses the original question as context when you ask followup question? sometimes when you ask followup question it just takes that question and searches for that only which is useless

if i would want that i would create a new thread clearly

[elder kite]

Consider how the prompt sent to the LLM contains not only the stuff you have written, but what the LLM has written to you, as well as all the context that has been used in answering the query.

So, if you were to ask something like "What are the opening hours?" after querying information about a cinema, the LLM might be confused with all the hidden context. Adding a minimal amount of context to your follow up queries usually help grounding the LLM responses. "What are the opening hours of the cinema"

I see many having issues with threads losing context. I have been able to mostly keep threads going, as long as I give small hints here and there. I made the mistake earlier today of asking a follow up of something like "translate the answer to finnish" and perplexity went and "googled" the translation for "the answer". I deleted the response and rephrased it as something like "translate your answer of x into finnish"

[south kindle]

Pro Search?

[spice lance]

on

[south kindle]

Do you know if you have the newer pro search that does multiple steps/searches?

[foggy pecan]

is it me or has the quality of answers gone down again?

[spice lance]

what?

[fluid mauve]

You could argue that these trash ballons are WMD's I mean they are delivering trash, trash generates activity by rodents, rodents carry diseases, humans could get bitten by afore mentioned rodent thereby transfering a disease to a human. Yep thats definitly a biologocal weapon /trash carrying baloon. I can just hear the minds in some militart bunker guaging a 'proportionate response"

[elder kite]

There seems to be changes to the question suggestions in perplexity. I had only typed "in" and perplexity started suggesting queries based on the past couple queries I have made (I was looking into interesting games)

[granite grail]

probably due to context window. The previous answer and search results occupied all of the context window so when you make a follow up question, it doesn't remember what your previous prompt was

Would be cool if someone from perplexity could confirm this though

[elder kite]

Not a dev, but I have a strong suspicion that the previous prompt is in the context window, but the context window also contains all the query results from RAG, which is confusing the context. So, when you for example say "it", referring to the previous answer or question, the model could think "it" is something it saw in a result somewhere

My argument/proof for this is that you can easily refer to the earlier discussion by giving just a bit of hints. "With the calculated hiking speed from your previous answer, when should I leave from the camp site if I want to hit the next camp site before sunset". By that point my thread history has the hiking speed calculation and the time of the sunset

The first question failed as I did only something like "When should I leave the camp site"

[granite grail]

it is in the context window, you did raise a good theory that it might just be it not being able to distinguish what you're referring to with the new prompt.

But in my experience, when it returns fewer results, it usually knows what I'm referring too, making a more "natural" convo. It's only when it returns 15-20 results that it fails.

Nowadays I just try to be clear as possible always, to prevent these kind of situations

[haughty sleet]

Have you noticed perplexity says yes to litterally everything if you use a tone where you request it to do something?

like "can you use react to make scratch games?" and it said yes

[restive osprey]

tried another custom gpt. this ones kinda funny

i continued. it didnt disapoint.

[granite grail]

I love it, link?

[restive osprey]

there was a slight issue with the first attempt. Ill copy over all the instructions and one file i have so far, to a new gpt. I think after its set, it wont allow a share. So gotta make a copy. If im not mistaken? First time i tried to set it to Share. But because the file i used had "kill killed, alot of that, open ai flagged it from processing

but, i think i can do it now, because i deleted all the kill words out the file i was using

[restive osprey]

ill dm you tell me if it works

[modest token]

I ended up using the bottom one for my documentation. Sure there's cats in the image for a 404, but it works for my InvalidCatsException gag I have on the 404 page

[sleek vortex]

time to do some testing

🙄

[restive osprey]

If anyone has gpt paid, and wants to check it out if it works https://chatgpt.com/g/g-bLk7udOo7-hk-47

[sleek vortex]

omg did you say cat

[modest token]

I did yes. I use it on my 404 page.

[grave coyote]

I think one of my complaints right now on perplexity is that it's search plan seems limited to about 4 elements. If I ask it a complex question I often have multiple elements I am interested in and frequently a more complex search plan I think would result in better coverage. It feels like there should be an option for longer form answers and more sources for when you really want a more indepth answer.

[tame current]

Is there a way to give paid version a "free trial"?

[warm cave]

They said free trials was removed due to abuse of the platform, right now the only thing I know of is $10 off with a user referral code

You do get 5 uses of pro search per day tho for free, that means better sources just not a better model

[tame current]

Hmm, but won't Pro be able to empower the search with advanced models like GPT-4 Turbo, Claude?

[warm cave]

“Pro toggle” is different from the “Pro” models. When you pay for Pro on perplexity you get access to both. I will send an image of my model page so you can see the option, the pro toggle makes it so it has better web sources

So basically with a paid account you can use any of these models for all searches with pro toggled on or off. When you toggle pro on you also get better web sources given to the AI

I know, having both be called pro is confusing and I wish they kept the toggle called copilot for search

[tame current]

Ahh so does "Pro toggle" simply feed the basis AI model the most relevant information it needs to produce an accurate answer?

[warm cave]

So if you have pro toggled off it’s the top 5ish web searches fed in. But Pro just got an upgrade, before it pro used the top 20 sources, but now it dose multi step searches, where it can use information from one search to inform the next. This feature has been slowly rolling out over the last few weeks so I don’t know which you will have. I will show an image of the new pro real quick

[harsh stag]

yo what's up with all those steps 😭

[warm cave]

https://www.perplexity.ai/search/Find-out-the-o4J2zXiaQXWZAkNu6RuVuA

[warm cave]

I asked it a question that requires a lot, lol

[harsh stag]

33 sources 👀

[warm cave]

Trying to push it to the limits

[warm cave]

Yep

I got 42 on one

[harsh stag]

i haven't got access to the new pro yet, I got it for 1 message and then poof

[warm cave]

Oh, I guess it’s not fully rolled out yet

Any question you want me to try?

[lunar dew]

I got the same trouble!
If anyone got info about the "noindex" Situation

Is looks like only the article from the Staff are indexed

Thank by advance

[placid coyote]

I am pretty sure I just saw omni writing the example code correctly (some numbers in square brackets), but after it finished, pplx removed it 😑

damn, it does look like some search sources processing. it either removes it, or transforms [1, 2, 3] into [1][2][3]. probably depending on if it is a valid search result (here are none)? I guess I'll have to head to chatGPT for this one 😞

[whole quartz]

Who watching apple wwdd?

[sleek vortex]

what in the level 12 agent was that example

[opaque stump]

This is going to be when LLMs go mainstream

Your boomer mother will be using this to generate MAGA memes in no time

[sleek vortex]

15 pro NOO

bruh

well saw that coming but

[opaque stump]

The multiple models integration is interesting.

[sleek vortex]

very interesting event for sure

apple focusing way more on the consumer than any other product

realistically i think the llm is the most boring part here

the core of this is their intent or whatever "semantic" model

[warm cave]

Fancy name for Rag I think

But maybe there rag is really good

[sleek vortex]

their rag is a seperate model

also they had that onscreen analysis model which was oss

probably based on that

apple silicion servers

is interesting

[warm cave]

Yeah, I guess we will have to see what it’s like

[cloud elk]

hello fellow know-it-alls

lol

[south kindle]

https://tenor.com/view/you-know-it-wink-the-office-michael-scott-steve-carell-gif-17547389

[jovial kayak]

I actually predicted the name for new mac being sequoia yesterday lol

Stupid reasoning but correct outcome

[south kindle]

It’s a Toyota SUV?

https://tenor.com/view/nwmsrocks-northwest-motorsport-pnw-pacific-northwest-toyota-gif-20681791

[south kindle]

FYI this section of your Page didn't "condense" into a Title.

PPLX will get there own Page out though for the discovery feed, so not sure if this will be duplicated work. or they'll just steal yours? lol.

[jovial kayak]

Yeah I left it like that since it was being weird

[south kindle]

i find when you can't adjust a part of a section, it's bugged and you need to nuke it and start the section over. not a huge deal.

not great though.

[tame current]

Will ios chat gpt use 3.5 and 4 omni or 4 turbo?? Also are there limits to it?

[south kindle]

https://www.reddit.com/r/ChatGPT/s/KrhcqUpEuq

[tame current]

Thanks

[south kindle]

ChatGPT allows some GPT-4o uses for free users but it’s limited before converting back to 3.5. That being said this isn’t related to PPLX at all.

[tame current]

Yeah so gpt 4o

Ofc

[south kindle]

Oh the Apple stuff.

That Reddit thread details it. Free 4o.

[tame current]

I mean how do you guys find 4o

I prefer turbo

[haughty burrow]

Just a quick question. I got my Rabbit R1 and I'm trying to see if my Perplexity Pro account is connected to it from the start? Or do I need to connect something?

[south kindle]

Do you have Pro working on the normal web UI? if we can confirm the same account is logged in to both, you should be able to verify if your Pro account is active on the Rabbit. (i don't have one personally so this is just what seems most logical)

[warped marsh]

oh you invested in a scam.. sad

[final holly]

I wouldn't say it's totally a scam, but it's definitely a huge case of false advertising.

[warped marsh]

no it is a scam

did you not see the coffeezilla video?

It is a scam.. 100%

[warm cave]

2 Videos, haha

[sleek vortex]

install android on it and call it a day

the actual device is pretty much gonna be useless in a few months i’d bet

[south kindle]

Ideally we aren't shaming people that have Pro access though. they can still utilize PPLX Pro regardless of how it works on the Rabbit.

[fading moth]

I'm unfamiliar with this Rabbit thing, can anyone explain what it is and why it's a scam?

[sleek vortex]

yeah

true

well i don’t wish to shame anyone

i also believed the rabbit was real for a bit

[sleek vortex]

well TLDR the owner is also involved with a history in some crypto scam

based on people from inside the company “LAM” doesn’t exist

and the current “LAM” they showed off isn’t actually AI nor is it secure at all

[fading moth]

Sheesh

And it looked like such a neat little thing

[agile jay]

Rip, it was pretty obvious they were just an Openai wrapper.

It's rather difficult to make a device that competes with a phone.

Can you think of a few at the top of your head?

[sleek vortex]

well there’s things like the
palm phone
light phone

honestly man if they just removed the scroll wheel and sold a rabbit r1 cute android phone like that would have brought it

it honestly would have been the perfect like detach yourself device

palm phone but not insanely expensive

[agile jay]

Probably the best option would be the Samsung Flip line.

[sleek vortex]

well that’s a whole phone then

the point would be a like phone replacement which is purposefully not as functional as my iphone or something

idk

[agile jay]

Yep, but you have to unfold it.

What makes you less likely to use it.

And you can use a few apps using the small screen, so you don't fall into notification hell.

[sleek vortex]

well if it’s low power enough it wouldn’t be running things like tiktok already

then again it just being an android phone would mean a lot

really isn’t a lot i can do to make myself an anti phone with my main phone

[worthy mesa]

Nvidia NIMS vs Microsoft Copilot+ PC vs Google Gemini Advanced vs Apple Intelligence… What kind of Battle Royale is this ? 😅

[cobalt widget]

😆

[south kindle]

I think Rabbit used some kind of new device paradigm as a marketing gimmick to get preorders. He somehow pulled one over on Teenage Engineering, a company literally built on integrity and authenticity thru innovative design. But the "goal" of had merit, but was never going to replace a persons actual phone, that's become part of their personality.

Thinking of unique ways smart devices integrate with others, if anyone has played around with android auto or apple carplay, your "phone" essentially gets an external monitor on the infotainment system. At current, some data is passed from the car to the phone, and that is getting pushed to possibly the phone taking over all screens on the dash. there will be pushback, but tech/progress wise, that's where the tech is being pushed.
Or thinking of a smart watch being an extension of your phone, displaying notifications along with personal health data. And a watch with those sensors to feed the data into the Phone for analysis.

Something needs to remain static as the powerful brain behind the personal intelligence system (which exists today as just having a smart phone.)
I could see a future where we (society) move on from a physical "phone" and instead just a central intelligence unit, that's basically a modern day "phone" itself, with extended "peripherals" ie: a watch, glasses, ear piece for phone calls, etc. At some point we will stop calling things "Phones" as in "device you call someone on". that's such a small aspect of modern smart phones.

[valid yoke]

Smartwatch

[south kindle]

Smartwatch requires a phone for most things. Garmin gets a pass here for being their own system.

At least the "smart" aspects of it.

or to report health data back to a centra device/app.

due to limited on device storage.

dumped this chat into a thread. having Pro Search do multi-step searching on different aspects.

[valid yoke]

Not my watch

My Apple Watch works completely independently from my phone

[south kindle]

it does, but it's not designed to. lol.

[valid yoke]

That’s not exactly true. My dad has an Apple Watch also and it’s not even paired to a phone.

Apple introduced that a few years ago

[south kindle]

Your using fancy color screen watches, not smart watches

They are smart obviously

[valid yoke]

wth?

[south kindle]

but my vision of a smartwatch being useful is getting phone notifications so I can avoid pulling my phone out of my pocket.

[valid yoke]

It’s a computer. So is a smartphone. So is a tablet. And a laptop.

They’re all just computers, they only differ in terms of whether they have cellular antennas

[south kindle]

My point is the devices are built and created to be used with a phone.

[valid yoke]

Heck even my tv is a computer

[valid yoke]

No, you’re thinking old

[south kindle]

i'm literaly not

[valid yoke]

The first smart phones from Apple and Google required a laptop to set them up lol

They were designed to be used with a laptop, but they involved

So have watches

[south kindle]

i'm 34, i was here the entire time son.

[valid yoke]

I’m older than you and definitely not your son unless you went back in time

[south kindle]

either way, your missing my point entirely.

[valid yoke]

No, I think we just disagree

[south kindle]

and i'm missing yours. or i'm looking past it because it disengenious to my point.

[valid yoke]

My point is that everything is a computer and people need to stop calling phones phones

[south kindle]

we agree on that

[valid yoke]

They are cellular-capable computers

A tablet is also a cellular capable computer. Well some of them.

[south kindle]

i'm talking about a future where devices are interconnected with your data.

[valid yoke]

I’m saying that isn’t future it’s right now

My watch and my laptop already sync to the cloud

That’s why I don’t need a phone

[south kindle]

in that sense, the phone is part of the central intelligence personal system. Rabbit failed. i'm looking at what they set out to do and envision what the future of connected devices could be

[valid yoke]

Honestly, so long as Apple and Google are at the helm, you’re not gonna see a post-phone future

It’s not profitable for them

[south kindle]

sorry for the miscommunication. I've only used Pebble/Fitbit watches over the years, but am an iOS person currently.

[valid yoke]

Yeah, I’ve had a couple pebbles. They were great.

[hard veldt]

Anyone not able to get the login email? I’m using iCloud.com🙏

[valid yoke]

I had no trouble with my iCloud

[south kindle]

only went to fitbit for the sake they bought out pebble and the watchface people moved over. so i kept using the same UIs.

[valid yoke]

Yeah, but then fitbit got bought out by Google so

[hard veldt]

Thanks. Maybe my problem

[valid yoke]

We are back to the original duopoly

[south kindle]

my fitbit only holds like a week of health data before it starts losing it. I look at that stuff daily and have tracked my health stats since like 2017. i get annoyed when there is gaps. currently, the only aspect of the wifi chip on the Fitbit Sense is for firmware/software updates.

[valid yoke]

If you’re an iOS person, then switch to Apple Watch it seems obvious

[south kindle]

never felt the need. Fitbit sense does everything I need. also i only recently as of a year ago got a iphone 13 mini. and was using the iphone in tandem with the Pixel 4a. the sense isn't getting replaced until it stops getting updates or the battery goes. I don't just upgrade my hardware for the sake of it pretty much ever.

[lapis spindle]

how do i generate image on perplexity

[restive osprey]

whats the best way to use the gemini pro in google environment ?

like with all their tooks. i signed up for the free two months [ plan to cancel ]

just trying it out .

got it to work on my google drive pdfs and text files

1 mill tokens

[safe karma]

do you have to be perplexity pro to get access to pages? i cleared my browser data and i dont seem to be able to check it out

[restive osprey]

was thinking, could i use it with maps, calanders etc?

[restive osprey]

was thinking the same thing. i asked this few days ago didnt get answer

[safe karma]

maybe the rollout is in stages. ill check back in a couple days i guess

[warm cave]

I think the beta is closed right now, I got in a little while back and if I remember correctly they said they had enough testers for now

But also that was a bit ago, maybe they are rolling it out now

[safe karma]

i saw a twtiter post from them saying they are rolling it out on may 30th. ill just chill and check back

[south kindle]

Currently only Pro users get access to Page generation. and Page generation itself only works on Desktop/Web.

[dark ether]

did you know that perplexity doesn't work well when sourced from academic?

[restive osprey]

have u used it?

[placid coyote]

hello, so, is it just me, or are all threads (history) broken? anything I access from the library results in homepage with the red notification ._. refreshing didnt help, it happened previously and I got logged out, but this seems to be a different cause

[foggy pecan]

Hi is there a way by which when i upload images in succession I only get result for the image i uploaded currently instead of the one's i uploaded a couple of queries before

generally when i upload multiple screenshot one after another i get the result for the previous one's as well which is kinda annoying and not needed

and dont know if it was reported before the prompt that we get after sending in a query is busted it doesnt allowing pasting or writing more that 15 words as of now.

[atomic talon]

Hey just want to see, if I'm creating images right way with perplexity. (Feels a bit weird, and images are often not fitting the description)
I enter my prompt in the normal search modal > hit stop as soon as it is possible > then I use the create image dialogue on the side.
Is this right? Happy about ideas! 🙂

[grave coyote]

I actually reported this in the bugs thread quite some time ago and it was ignored. My problem was that it was pulling sources like Wikipedia and various blogs not academic papers. When it did select papers it seemed to only have access to the abstracts. I could download papers and then upload them to perplexity and get better answers but it really isn't great. I can't recommend Perplexity for academic research in its current state. I am having better results with Google's NotebookLM. Of course there you must gather and upload your sources on your own, answers are shorter and more focused and not as nicely written, but they are accurate and can be limited to reliable sources.

Another problem with the Pages feature is after getting a good thread from uploading quality sources, covert that to a Page. It makes rather short summaries and looses much of the detail. Checking each section it is always set to concise. There seems to be no way to change this default. When you go to each section and edit it to be detailed, or if you add a new section, it rewrites it entirely choosing new sources and ignores the original thread. With Pages there is no way to upload a document or otherwise add your own source, making it extremely limited.

I love the concept of Pages but the lack of editing ability, the inability to control sources, and the lack of social features (comments, feedback, discovery) and no easy way to export it to another system really make it hard to recommend.

Overall Perplexity is a very interesting tool for searching and summarizing general web content, but it is not really suited well toward more serious research, even though they try to advertise themselves as such with the "focus" modes, which is disappointing. I feel there is a lot of room for a competitor to swoop in and do this better.

[placid coyote]

soo, I guess pplx doesn't like when LLM is instructed to respond with entirely empty text 😆

[mystic nacelle]

skibidi sigma pomni digital fortnite chamba free gigachad rizz omaiga floo xxxtentacion smurf cat lionel ronaldo junior mewing tercero chikibay xocas insano

[opaque stump]

what's the link to check Opus Limits again?

they should just put it in the UI unless I am missing it

[harsh verge]

https://www.perplexity.ai/p/api/v1/user/settings

[livid aurora]

They removed one day

Where can I find Addons for Perplexity

[opaque stump]

ty

[tame current]

how do i get it to write romantic messages for me

with gpt 4 omni

[orchid elm]

(moved to quick-questions)

[tame current]

dayum!

[limber shore]

What is the best way to use chatgpt or Perplexity in WEAR OS without the use of API?

[fading moth]

[shut plaza]

why wearOS

[fading moth]

Because Apple

Iunno, I personally hate digital watches in general

I don't think you can use LLMs without an API key.

Kinda like... How do I drive a car without starting it

[fading moth]

[tribal spoke]

Receiving this error from Safari iPad, curious if anyone else is seeing it or if it’s a layer 8 issue (me) - Application error: a client-side exception has occurred (see the browser console for more information).

[spare smelt]

perplexity does not yet interpret images with the pro ?

[pallid bough]

Is perplexity AI file uploads/image interpretation down?

[pallid bough]

It does, I'm assuming it's down for the time?

[barren gyro]

Yes

[jovial kayak]

Dang I think it's not just me

Was about to ask if anyone else had problems just now with the ai's not being able to view links or files

About 15 minutes ago I sent it a link to a pdf as normal and asked it to give me the answers so I could check when I was done (science final tomorrow), but it got the first question wrong so I said it did and it sent it again with an even worse answer. I went to a new thread and this time in pro mode to see if it would do a nicer job even if it's slower by like 3 seconds, and when I tried uploading the pdf as an actual file, it didn't work. I then sent it as a normal pdf link like how I did before with a different pdf, and it just didn't work; kept saying that it couldn't view external files or links. I tried the old pdf and it stopped as well.

[agile jay]

What was the prompt?

[south kindle]

It's related to the new Pro Search A/B testing. If your attached files show up like the 1st image, it's not actually reviewing the file at all. Confirmed by @signal hamlet that they have a ticket started for the issue. So it's being worked on/looked into.

[agile jay]

So just making stuff up

[south kindle]

Just had it tell me a cyberpunk city skyline was a handwritten note, and then it detailed the entire made up note.

[sleek vortex]

override the flags😄

[jovial kayak]

Sorry didn't see this, but since pdf's and links didn't work I tried just screenshotting the practice test I wanted it to check for me, and this is what it said when I gave it direct images as its sources:

And here is it on the non-pro search:

[agile jay]

Can you link the sources, so I can try it out?

[jovial kayak]

I'll send one cause it has my name on it for one of the pages

https://pplx-res.cloudinary.com/image/upload/v1718145936/user_uploads/oflavjjzg/image.jpg

[sleek vortex]

Yeah because pplx doesn’t just use one model

and they didn’t execute it 100% well - a bit of a mess

[agile jay]

Yep, it fails in perplexity, but works on ChatGPT.

[jovial kayak]

I am using gpt-4omni though

[agile jay]

Yep, but that doesn't mean anything. That's just the language model.

Not what is being used to process images etc.

[jovial kayak]

Oh thats weird

By the way, for the average user, does language model really matter? There's significant differences between the image generation models but is there really a big difference between the language models when using them? Like I've tried just eyeballing speed and really there is not an actual difference whether you use the giant 32k sonar thing vs claude sonnet which is said to be the fastest one.

[agile jay]

Sadly can't share chatgpt messages which include an image, yet.

[agile jay]

Claude Sonnet and Sonar, should be around the same size.

[jovial kayak]

Wait what does that mean

[agile jay]

Infact llama 3 70B, which is the sonar model, should be slightly better.

[jovial kayak]

So you think I should start asking questions in that like sonar 32k llama thingy

[agile jay]

Are you a pro user?

[jovial kayak]

Yeah

[agile jay]

And do you only use perplexity for search?

[jovial kayak]

Oh I just realized that the discord wasn't pro only

[agile jay]

Or in writing mode too?

[jovial kayak]

I never noticed the discord icon until after it like recommended me it when I got pro

Um I don't use it just for search. I use it as an AI to get answers from whether it be search or by giving it my own sources which hasn't been working recently as you know, and also to help write.

[tranquil orchid]

Hey @jovial kayak!

If you find the original message helpful, please consider reacting to it with the :star: emoji. If the post is appreciated by the community and receives 5 stars, it will go to the https://discord.com/channels/1047197230748151888/1082806833938436228 channel and the post author will get the <@&1082034222778302614> role on Perplexity.

[agile jay]

Overall, the best model is probably GPT4o

But depends on the task you are doing.

[jovial kayak]

I really don't understand the differences lol

If I just ask it anything whether it be answering a math question, science problem, or writing a prompt for english, or even just searching for something, what's the best one to use?

[agile jay]

GPT4o

• good prompts, lol

I can help you check the answers to the questions on the image you provided. Let's go through them one by one:

31. One cause of muscle soreness is:

• a. alcoholic fermentation.
• b. glycolysis.
• c. lactic acid fermentation.
• d. the Krebs cycle.

Correct answer: c. lactic acid fermentation.

32. The air bubbles and spongy texture of bread are due to which process?

• a. lactic acid fermentation.
• b. glycolysis.
• c. alcoholic fermentation.
• d. the Krebs cycle.

Correct answer: c. alcoholic fermentation.

33. Milk is converted to yogurt under certain conditions when the microorganisms in the milk produce acid. Which of these processes would you expect to be key in the production of yogurt?

• a. the Krebs cycle.
• b. photosynthesis.
• c. alcoholic fermentation.
• d. lactic acid fermentation.

Correct answer: d. lactic acid fermentation.

34. Cellular respiration is called an aerobic process because it requires:

• a. light.
• b. exercise.
• c. oxygen.
• d. glucose.

Correct answer: c. oxygen.

35. Breathing heavily after running a race is your body’s way of:

• a. making more citric acid.
• b. repaying an oxygen debt.
• c. restarting glycolysis.
• d. recharging the electron transport chain.

Correct answer: b. repaying an oxygen debt.

36. When the body needs to exercise for longer than 90 seconds, it generates ATP by carrying out:

• a. lactic acid fermentation.
• b. alcoholic fermentation.
• c. cellular respiration.
• d. glycolysis.

Correct answer: c. cellular respiration.

37. All of the following are sources of energy during exercise EXCEPT:

• a. stored ATP.
• b. alcoholic fermentation.
• c. lactic acid fermentation.
• d. cellular respiration.

Correct answer: b. alcoholic fermentation.

38. Which process does NOT release energy from glucose?

• a. glycolysis.
• b. photosynthesis.
• c. fermentation.
• d. cellular respiration.

Correct answer: b. photosynthesis.

39. How are cellular respiration and photosynthesis almost opposite processes?

• a. Photosynthesis releases energy, and cellular respiration stores energy.
• b. Photosynthesis removes carbon dioxide from the atmosphere, and cellular respiration puts it back.
• c. Photosynthesis removes oxygen from the atmosphere, and cellular respiration puts it back.
• d. all of the above.

Correct answer: b. Photosynthesis removes carbon dioxide from the atmosphere, and cellular respiration puts it back.

40. Photosynthesis is to chloroplasts as cellular respiration is to:

• a. chloroplasts.
• b. cytoplasm.
• c. mitochondria.
• d. nuclei.

Correct answer: c. mitochondria.

Based on these correct answers, the answers provided in the image are all correct. Great job!

@jovial kayak Here's GPT4o response to the image, in ChatGPT

[jovial kayak]

How did you do that magic

Also so glad I sent you that page because I was really hoping you weren't gonna send the answer in general only for there to be like 6 incorrect's

[jovial kayak]

Oh is that the actual gpt-4o

[agile jay]

Free users now have access to 5 GPT4o messages a day, I believe.

[agile jay]

Yep

[jovial kayak]

Maybe I'll try a different language model then

Get my money's worth lol and just cycle through all of them

Is the default one good?

[agile jay]

Which one?

[jovial kayak]

Perplexity

[agile jay]

Nope, pretty sure it's claude haiku

[jovial kayak]

Oh wait if you got gpt can you do the cool facetime thing and send a screenshot

[agile jay]

Fine for simple search, bad for anything complicated.

[jovial kayak]

I doubt they will now because of what you said earlier about how they're just models but it'd be so cool if perplexity had that feature

[agile jay]

Perplexity should be able to support the same features through the API, just the devs are pretty slow at implementing stuff.

[agile jay]

The Voice chat feature?

[jovial kayak]

Yeah

I call it ai facetime

[agile jay]

Yep, but not the latest one.

The first version has been available to all users since January.

And earlier for plus users.

[jovial kayak]

Also talking about image generation, I really like the outcomes of stable diffusion but I notice it likes to just ignore my prompt completely and only implement like one thing, here's what I mean:

Image:

Here is a prompt based on your description:
"Write a story about a mysterious room where a character discovers a hidden trap. The room has sea blue walls and a white door with a shiny gold doorknob. A string contraption stretches horizontally between the door frame and the entrance, waiting to ensnare an unsuspecting visitor. What happens when the character enters the room and triggers the trap?"

And actually in my like question, I also include the fact that I want the scene to be viewed at as if the camera was on the ground

[agile jay]

Yep, just strange that the bottom right of the door is further back, lol.

[jovial kayak]

I wish you could also just upload an image to the AI and then ask it to put something in there

[jovial kayak]

I try with photoshop generative fill and then regret my life choices afterwards

[agile jay]

This one probably better fits the prompt you wrote.

[jovial kayak]

Now that's what I call stringy

Did you make that with the chatgpt thing

[agile jay]

Yep

[jovial kayak]

I might get it honestly

[agile jay]

The previous one looked like this:

Where it looks like an actual trap.

Perplexity has dalle 3 too, you can try changing the image model.

[jovial kayak]

Perplexity is very confusing as you can tell by my hundreds of questions, and while it does have the language models for chat gpt and other things like claude, I really like the simplicity of chatgpt's interface and prompting, as well as the fact that they are constantly coming out with revolutionary features that will take forever or maybe never to implement to other ai sites.

Also that you can just enable a plugin and simply ask to generate an image and it will

[agile jay]

But the ChatGPT version has inpainting.

[jovial kayak]

I know about dall-e 3, I just switched to sdxl to test it after like months of banishing it from my settings

[agile jay]

Did you want any changes to the previous image?

[sleek vortex]

me always reading this server to see other people's issues with perplexity lol

[agile jay]

I can try inpainting it.

Yep, the problem with perplexity is that their team is really only focused on search.

The rest of the features are not really implemented well.

[jovial kayak]

Sorry didn't see, my mom offered to get me a milkshake from baskin robbins lol

Is inpainting like uploading an image made by the AI and then asking to add something in it?

[agile jay]

Highlight an area and give it a prompt to edit that specific area.

[jovial kayak]

Oh shoot give me the website I wanna try

[agile jay]

https://help.openai.com/en/articles/9055440-editing-your-images-with-dall-e

https://chatgpt.com/

[jovial kayak]

Oh wait I need the chatgpt+ right

[agile jay]

Might not anymore

The features have supposedly been released to the free tier too.

Just they get a lot less usage.

[jovial kayak]

Ok well I gotta do a bunch of studying since I got my biology/chemistry exam tomorrow

Thanks for all the help earlier

[tranquil orchid]

Hey @jovial kayak!

If you find the original message helpful, please consider reacting to it with the :star: emoji. If the post is appreciated by the community and receives 5 stars, it will go to the https://discord.com/channels/1047197230748151888/1082806833938436228 channel and the post author will get the <@&1082034222778302614> role on Perplexity.

[agile jay]

Yep, better study

SAT's or something?

[jovial kayak]

No just finals

I'm a freshman so don't gotta do those yet

[agile jay]

Oh, so 17ish

Are they actually mostly multiple choice?

[jovial kayak]

Wait actually question about that, is it the end of junior yet when you take the SAT so then at the start of senior year you use that and then like the essay and other stuff to send or is it the start of senior year

No I'm 14

Freshman in hs

[agile jay]

Oh, Thought highschool started later.

So when is middleschool?

[jovial kayak]

Wait I'm confused are you talking about like final exams?

[agile jay]

Nah, it's just here it's different age wise.
primary is around: 6 - 11
high school is around: 11 - 16
and collage is around: 16 -18

And uni is depends on what you're taking.

[jovial kayak]

Oh are you in England

[agile jay]

Yep, in the UK

[jovial kayak]

Yeah so here it's like 13-18. For me 13-17 since my birthday is in December so half the year I'll be on the younger side you know, but for my high school we get in start of september-ish and then out middle of June so this is my finals week which is basically just the big exams going over every topic either the entire year or for some classes just the last unit or from midterms to now. Basically just like 5-6 tests in one.

Didn't do so well today for my first two, math I got a bad score on cause it was the WHOLE year, but for my history class I got a 97

[agile jay]

Rip, at least you have a lot of time left, since you're only 14. Just don't go slacking off and trying to cram last minute, lol.

[jovial kayak]

I think what I do which I should change but kinda too late now, is not lighlty studying and then week of or even less than that, start cramming. Instead I really should spend more nights actually going through everything but there's a problem with that. My math teacher was putting so much pressure on us to study the entire year's worth of notes, but then at the same time would give us 16 problem homeworks due the next day.

Also I'm kind of notorious to my teachers now for asking questions the night of the test and then like 6 before it starts in the morning, so I wanted to give my bio teacher a little farewell note before I leave class.

[pallid bough]

image analysis is still down?

[sharp lodge]

mine too

[fast moon]

Hope to increase the GPT4 model, but the effects of GPT4 Turbo and GPT-4O are not satisfactory.

[spare smelt]

it's possible not to have a question

[hidden knot]

Hey, I am experiencing significant delays with perplexity when I have attachments added and ask it to answer questions using the attachments as sources. This was not a problem for me before, but the last month or two it is a problem I keep on having. Anyone know why? its taking at least 30-40 seconds before responding. Note: I am a paying user.

[limpid warren]

Does anyone know if Perplexity still has a policy of inviting others for a $10 discount?

[south kindle]

#💬│general message

[harsh verge]

This problem has been solved for me this morning (Brussels time). I can use the attached files again.

[pine marsh]

Hey everyone, is it possible to reach an admin there ? 🙂

[halcyon coral]

Hey, what's it about?

[pine marsh]

Partnership 🙂

[blazing mulch]

Great to met everyone

[charred saffron]

Hi all, has anyone used PPLX to create a content calendar?

[spice lance]

why tf cant u say eat near me and it gives you near you when you specify your exact location in settings on iphone

im in iceland gives me boston like it's nothing

[sleek vortex]

the downfalls of pplx releasing half baked features yet again

[proper sage]

Hi, I want to create a GPT in a Perplexity AI collection. But I am limited by the number of lines or characters. What is the limit of lines or characters allowed to create content in the Perplexity AI Collection? When i ask to Perplexity AI i didn't have answer

[harsh stag]

the limit is exactly 2k characters

[sacred mauve]

Are there business owners want to increase their revenue. Just DM me. I'll help them for free 🤫

[random ocean]

Where can one input Scratch pad prompt in Perplexity App?

[south kindle]

@random ocean > Either copy paste, AI Profile, or a Collection. Let me know if you need any help or have further questions! scratchpad is mostly my baby but the community here did help revise and improve it a great deal to make it what it is today. Personally I have it set as a Collection Prompt, and just start threads within that Collection directly to invoke the framework/prompt. I also have a version loaded into my AI Profile for when I either don't use my dedicated collection, or if I'm trying to do something else specifically but still want the model to use the structure to breakdown my query.

[jovial kayak]

What is this supposed to mean?

Was asking what this image was since my friend sent it and then asked to find it on google

[south kindle]

Try the query, "review this screenshot" instead.

Asking for a direct link to the image is likely too specific for PPLX.

[jovial kayak]

So do I resend the screenshot and then ask for it to find something else online that looks like it?

Yeah so I just did what you said basically and I guess perplexity doesn’t feel like using any other sources if you give it something, kinda stupid ngl

[south kindle]

I'm trying it as well. it will review the image, but I don't see a way to get Pro Search to actually source similar images and provide direct links. You can either start a new thread or try in the same thread, shouldn't really matter.

[jovial kayak]

Yeah no matter what language model, pro or not, perplexity never wants to look for other sources if you give it an image or file as an attachment

[south kindle]

Well given the task, I'm not sure how else to present the image. hmmm.

Just as a "back-up", I also test file/photo uploads to free ChatGPT with 4o regularly. may be worth a shot for the heck of it, since Non pro users for ChatGPT get limited access to file uploads/GPTS/4o/etc.

[hallow linden]

https://www.forbes.com/sites/randalllane/2024/06/11/why-perplexitys-cynical-theft-represents-everything-that-could-go-wrong-with-ai/

lol if I were Forbes I’d be pretty pissed too

That being said, other publications do this too

Summarization and attribution is not a new phenomenon in the news industry

[south kindle]

It really isn't. people are just up in arms about ideas and implementation being mirrored on different platforms it seems. I don't fully see where they "stole" anything aside from a good idea.

I'll have to dig into it more. I'm also confused on their statement about the podcast. was it the content of the podcast that was directly from Forbes, or a summary style podcast itself?

[hallow linden]

Probably this one https://podcasts.apple.com/us/podcast/discover-daily-by-perplexity/id1732181427?i=1000658127961

It should be noted that there’s no attribution in the podcast transcript

[south kindle]

I would assume that is just them being lazy not adding citations to the transcript or description of the podcast. we can see the Page does cite the sources, even if it doesn't call them out specifically.

The discover daily podcast seems like it would get messy if every story ends with a minute of sources, however, maybe that's what's needed to avoid headaches like this?

[austere kestrel]

yeah it's the fact it was an investigation i think, like months-long work.. if it was just aggregating a bunch of sources about something broad and widly-covered (e.g. Gaza) I think it's eaiser to justify not having direct attribution

[south kindle]

That's fair. I mean, i would assume they dumped the forbes article text directly into Pages. or rather built a thread from their article as they do for all updates like this.

But how do you avoid doing it in a way that doesn't upset the people that created the sourced articles and content?

How many other original source creators had their article lifted for the Disovery feed / Discovery Daily? I'd assume more than just this situation.

Not in a "bad" way, but in a "that's just how PPLX is meant to function" way.

[fast meteor]

hmmmm

hard to image that his photos are not in the training data

https://www.perplexity.ai/search/who-is-he-nKR6DwraTAmY0HGdYCuHhg
hmm

so is he zuck?
no
so is he musk?
yes
...

[south kindle]

I was able to get "close" by telling it the initials were E.M.

Based on the image of a smiling man with short dark hair wearing a suit, and the query providing the initials "E.M.", the individual depicted appears to be Elon Musk, the CEO of Tesla and SpaceX. His facial features, hairstyle with a slightly receding hairline, and professional attire are consistent with Elon Musk's well-known public appearance. While the image alone does not explicitly identify him, the context of the initials representing his first and last name increases confidence that this is meant to be the famous tech entrepreneur Elon Musk.

https://www.perplexity.ai/search/who-is-this-UpstadEaSiSkXzfG8iSgjA

[fast meteor]

telling it the initials are cheating

[south kindle]

well obviously, but I don't think it has a way to actually compare an attached photo to photos online. I also don't know that any LLMs are trained on images specifically.

[stark torrent]

Hum since it use gpt 4o and other ai, perplexity must be the best ai out there ?

[fast meteor]

reverse image search, google custom api/yandex/tinyeye provides such a service

[south kindle]

Well until we get reverse image search baked into an LLM

[fast meteor]

depends on the use cases obviously

[south kindle]

i know the tech exists. ha.

[fast meteor]

anyways, i tried to mimic rabbit eye and came across this. i had to use ocr as a not so optimum walkaround

[random ocean]

Any way to increase font size or interface style etc. of the IOS app. I find font size to be small particularly on an IPad

I have to use perplexity on safari and use the window manually for increasing the text size,

[fast meteor]

#🍏│feedback-ios message

have a look 👀

[opaque stump]

Loving this new search

Need it in iOS now pls

[hardy willow]

Will Perplexity make OpenAI's Chat and Vision available when they're released

[fading moth]

Is anyone else having the issue with Perplexity's website sending the view of the chat half-way up the page after each answer?

[storm jetty]

Is there anyone I can talk to about support questions? I have an enterprise plan and emailed 2 weeks ago only to be dumped in the ticket system and never get a response

[fast meteor]

as i understand: paid users get priority support here: https://www.perplexity.ai/settings/account
or contact Alex

[storm jetty]

i already did that and sent an email to the priority Enterprise support which is separate

i didnt wanna bother them but yea i will dm alex

[sand hatch]

hello @everyone, nice to meet you all.
I am looking for a skilled AI developer to complete our AI mobile app. If you have a good experience in development RAG system, dm me and discuss about that

[hallow linden]

honestly the discover podcast kinda sucks. I get why people listen to it, but it's always just surface level

might as well just read the headlines

[empty abyss]

Hello everyone, I am searching for a speech-to-text dataset that includes MP3 files and original transcripts with diarization. Any recommendations would be greatly appreciated.

Additionally, if you have experience with any speech-to-text solutions that you found to be particularly effective, please let me know. I have already used AWS Transcribe, OpenAI Whisper, and Deepgram Nova-2. I'm interested in finding the best available solution for speech-to-text tasks. Thank you!

[sleek vortex]

perplexity in my control center with ios 18 lol

can’t afford apple intelligence so

[low stream]

I need some help with the API,
currently using "llama-3-sonar-large-32k-online" to query the following (which is giving me different results and error to if I use perplexity directly which is almost always right:

"Determine GCM (kg) for {{vehicle_details}}:

Check direct GCM specification
If unavailable, check similar variant's GCM
If still unavailable:
Find GVM and max towing capacity
If GVM unavailable, estimate from kerb weight + max payload
Calculate: GCM = GVM + max towing capacity
Verify calculated GCM across multiple sources

Return GCM value"

[austere kestrel]

huh that's interesting.. for a moment I wondered whether OpenAI had expunged Musk from their training data (one too many law suits from a bitter estranged founder..)
jks - but weird how the older gpt-4 plays along

[sleek vortex]

Perplexity API has much less capability

Do each task as a seperate request to the smaller model

[austere kestrel]

but then I noticed Opus does the same thing

[low stream]

okay, thanks how does that impact token and price?

[sleek vortex]

let me quickly check the documentation for you

[austere kestrel]

Seems it's basically refusals from Opus and GPT4o to identify people. weirdly though Sonnet doesn't seem to be trained to refuse (or just doesn't adhere)

[sleek vortex]

ok weirdly they’re priced exactly the same

[low stream]

okay, so it must be based on text quantity which would be the same either way

[austere kestrel]

that's way too complicated for the API models imo.. they can't do multi-step querying like with the web product

[sleek vortex]

well if you went with large or small and broke this query down

$5 per 1000 requests
so you’d end up paying quite an amount per request

[austere kestrel]

it retrieves search results based on the prompt

[sleek vortex]

uncovr api soonTM

:D

[austere kestrel]

it doesn't breakdown the query and iterate through it bit by bit (to my knowlege)

[sleek vortex]

yeah no

it injects first 5 sources into the prompt

that’s it

i was able to build a better api MVP in a week

[low stream]

okay so if i broke it up I would improve the quality of the output?

[sleek vortex]

perplexity api is so half baked

[sleek vortex]

it should at the expense of cost

[low stream]

hmm okay

[austere kestrel]

yes

[sleek vortex]

you could try and make this a non llm request to some extent

wonder if there’s an api for that data

or some page with a reliable source

[sleek vortex]

unless your workflow involves lots of prompts like this

[low stream]

it's exists all seperately which is why I use this

[sleek vortex]

if it’s one specific task then it could be worth abstracting out the LLM

[sleek vortex]

ah,hm

[low stream]

are there any alternatives that could do this?

[austere kestrel]

there

's a browser plugin called HARPA - you can create workflows, with the output from step used in the next steps. It's just what I use and I quite it like it, would say worth checking out - but there are no doubt various other options out there

you can kinda use the perplexity API with it, and also other connections. like google search / serp etc

[sleek vortex]

yeah or wait a month for my product lol

[sleek vortex]

yeah this is the best solution now though requires extra work

[austere kestrel]

yeah i think it's quite powerful but it still requires a bit of time to work out how it works properly.. any workflows that are somewhat complicated take a bit of time / trial and error to get to work reliably.. agree for sure

[low stream]

what's your product?

[sleek vortex]

i’m working on a sort of API like perplexity’s but bringing full agentic responses to the API and more

well first i’m building a foundation with a great scraping and search api

releases are still a while away - slowly working on it daily

[low stream]

sounds great, do you have a website I can monitor for release?

also, can you tell me if this is a resonable query to now expect, it's a lot more simpler in execution:
Determine kerb weight or curb weight (kg) for {{vehicle_details}}:

1. Check direct kerb weight or curb weight (kg) specification
2. If unavailable, check similar variant's kerb weight or curb weight (kg)
3. Verify kerb weight or curb weight (kg) across multiple sources

Return kerb weight or curb weight (kg)

[sleek vortex]

If you’re asking perplexity API, you have to think of it like a google search

Since the LLM doesn’t choose what sources it gets

The query is simple enough but it won’t actually be able to execute it in steps, ittl likely just hallucinate an answer

[low stream]

hmm okay, so what would you change in this consideration?

[sleek vortex]

hmm

I think with these tools: i would maybe do 2 queries to perplexity api trying to find the weight- basically encapsulating step 1 and 2

Then if you want to use an LLM to formulate the comparison/cross check for step 3 then maybe something cheap like llama3 on groq to formulate a final answer

If you need production scalability then maybe claude haiku (also cheap)

though it isn’t really an ideal solution, to say the least

[low stream]

thanks a lot for the advice. Just to clarify with you, would you seperate step 1 and 2 (I will omit step 3 for the time being) or could the model utilise the two together as it's quite simple
"1. Check direct kerb weight or curb weight (kg) specification
2. If unavailable, check similar variant's kerb weight or curb weight (kg)"

[umbral pollen]

is this accurate?

[south kindle]

effectively yeah, but the new Agentic Pro search does multi-steps/searches to reach a conclusion. but if you use search on ChatGPT or Copilot, it's doing the same thing, sources the latest top results to a search. and then uses the results to produce a reply that adheres to the User Query.

[austere kestrel]

there isn't anything special about RAG, whether for documents or web results. But i think Perplexity does it best (for now). There's Bing/Copilot and Gemini and ChatGPT, but none perform as well as the perplexity web product imo.