Condenses

Condenses is a context compressor for LLM and AI workflows · shorter context, the same intent

[Condenses @condensesai]

@DrowsyGN @EasyA_Kickstart we will lock🫡

[Condenses @condensesai]

Someone asked if the answer changes after compression

the model still sees the context, just packed into soft tokens it reads natively

shorter input, the same output

check it on your own prompt in the playground

https://t.co/TSXiuN9mpO

[Condenses @condensesai]

Long context is where the token bill climbs
RAG pulls pages, agents carry history, chats keep scrolling
Condenses folds that context into soft tokens so you fit more into the same window
more retrieved context, the same budget

[Condenses @condensesai]

Condenses sits between your prompt and the model

one call to /compress and the long context goes in as soft tokens

the model reads the short version and answers the same

drop it into the pipeline you already run
https://t.co/JDLhXnLxaS

[Condenses @condensesai]

Token efficiency you can measure

other tools tell you the savings, Condenses shows them on your own prompt in the playground

soft tokens in, the same answer out, the count visible every call

run it on your workload and read the number yourself
https://t.co/JDLhXnM50q

[Condenses @condensesai]

Someone asked how compression keeps the answer correct

the method keeps the key tokens and folds the rest into soft tokens the model reads natively

shorter input, the same intent, context kept intact

docs are open if you want the detail
https://t.co/9FjZQlC8Jh

[Condenses @condensesai]

Proof is the playground

paste a long prompt, watch it compress into soft tokens in real time

the model answers from the shorter context and you save up to 40% on tokens

runs on Llama 3.1 8B, your prompts, your numbers
https://t.co/TSXiuN8OAg @EasyA_Kickstart

[Condenses @condensesai]

Context, like bandwidth, should be cheap to move

every long prompt carries tokens you already paid to send twice

Condenses turns that overhead into soft tokens and gives the budget back to your model

direction is simple, make context affordable enough to use freely

[Condenses @condensesai]

Condenses is a context compressor for LLM workflows

it shrinks long prompts into soft tokens and keeps the meaning intact

up to 40% fewer tokens per call

try it on your own prompts in the playground
https://t.co/JDLhXnLxaS