How a 7800+ token article can shrink into 500+ token context? Screenshot

i have copy paste content from 2 chapter of the books.

38K characters

I uploaded into this website
( which it was based on latest GPT3.5 turbo, as it claimed )

Currently GPT3.5 Turbo (chatgpt) does not accept any fine-tuning nor embedding.
Only can be customized via Prompt Context.

How this website can achieve this? ( context I provided is 38k char long )
I have ask multiple question regarding the content, and its reply is accurate.
The Library upload process is take only less than 3 second ( so, its not fine tuning thing )

from the front-page of this site :

It has a free-tier, go check it out.

I really wish some1 can give me some insight how they can achieve this with OpenAI API.
( they based on openai, mentioned in pricing page )

  • my guess, summarized the uploaded article? and use it as context?

my another guest :

in the pricing page, the 1st tier $39 included with 500k openai token and other services.
based on ChatGPT3.5 turbo pricing, $39 could net you 19.5M token. ( but sure, $39 paid to ingestAI not fully converted to OpenAI token )

I guess the actual token spent on OpenAI API is much more than 606?? ( from the screenshot )

But the IngestAI only deduct 606 token from my account. ( since the token price is about 78x higher compared what OpenAI offered )

My guess… if it is using ChatGPT

  1. It first summarizes batches of N tokens
  2. It then summarizes the summaries

Where does it say it’s using GPT3.5 Turbo? I thought they were finetuning using older GPT 3. At least this existed for a few weeks and it’s older than ChatGPT API (GPT 3.5 Turbo)

please read their discord 1st.
they annouced they using 3.5Turbo

and, it only took 2~3 second for my document to be working.

it even mentioned aware of Chat History.
( so the library, chat history are compiled into prompt each time. )