Perplexity And Exa Issue

Thanks @Bob X @Questflow , I will surely check it soon. Really appreciate all the hard work you are putting here.

Attaching image of second point mentioned. I received the response but payment failed.

Hi @Bob X @Questflow , I respect the fact that it is “LLMs under the hood“ which are leading to high costs. But, I need to show you an analysis of how weird this can be.

The TARS agent of Questflow consumed 0.691 dollars as visible above. Since I will be making lot of assumptions, let’s say it consumed 0.60 dollars.

Assuming that GPT-5 is being used under the hood, the GPT-5 (high) has a cost of $1.25 per 1M input tokens. Since the output in the above session was comparatively smaller than input, let’s assume that the total cost occurred because of input. Which means that about 500,000 tokens were processed as input tokens. This is sonewhat to minimum of 300,00 words. To give you an idea of how big this number is :

1. Most modern novels range between 70,000-120,000 words.
2. A Clash of Kings by George R.R. Martin (326,000 words).
3. This is an average equivalent to 1.5 million characters, which is rougly 20k to 25k lines of HIGH QUALITY code (or maybe more), which is approximately 2 to 3 percent of Official React code.

A simple query of mine shouldn’t trigger this huge input amount. The math is not mathing. It would be kind of you if you may re-check the input ingestion process.

Also, my assumption in the beginning was that you are using GPT-5 (high) model, which is comparatively expensive than many other models in the market. Hence, if you are using any other model like Qwen2.5, which would be cheaper, and hence means that the input provided for this query was way larger than 500k tokens.

Two payment issues I just noticed :

1. Perplexity Agent was used but the payment to the Agent was not done. Check the 3 screenshots I have attached in the post. (I hope you are able to see all 3 screenshots)

2. I just checked and realised that payment to TARS agent was failed. @Bob X can you please check if the payment of 0.691 dollars to TARS was failed ? The money didn’t deduct from my wallet.

I hope that you solve the payment issues first, because it is high priority for you and the platform. After that, you may check what wrong happened during input ingestion.

Also, as you can observe that exa was called 3 times and perplexity was called 2 times. I believe this is too much tool calling. Maybe 1 exa call and 1 perplexity call was enough ?

It would be kind of you if you may please take a look into these issues. I am trying to provide you honest feedback given what I can understand. If I am wrong, kindly feel free to correct me. I believe that platform exist so that conversations are a two-way exchange. Hence I do not expect that I will always be right.

@Bob X I try to put as much observations I can, but it is difficult for me to put everything on feedback.questflow.ai , any chance I can DM you somewhere ?? Your twitter DMs are closed. I don’t know your telegram. (Sorry my telegram id name is a bit weird) … My telegram : https://t.me/AGicoming

Sorry for the inconvenience. The reason the Twitter agent wasn’t included is that you need to @twitter directly and specify the task you want it to handle (for example, “@twitter check trending tweets”).

As for the higher cost, that’s a valid concern. Please note the charge doesn’t come from Questflow itself. It’s mainly from the LLMs we use under the hood. We already provide 3 different LLM modes (Lite, Balance, Deep), and we’ll continue working on optimizing both cost and performance so you get the best value.

On it