What it does
Strips low-value words from prompts before they hit an LLM, cutting tokens while keeping the meaning intact. Tunable 0-100 intensity dial, plus web UI, CLI, and Python API.
How we built it
Python core that drops words in tiers — articles first, then auxiliaries, prepositions, pronouns, all the way to filler phrases at max intensity. Negations and question words are hardcoded as never-remove, checked before any drop logic runs. Flask web app + CLI + importable module all share the same compression engine.
Challenges we ran into
Deletion has a ceiling — too aggressive and output turns into unreadable fragments. Bigger problem: common words aren't the same as unimportant words ("not" is common but flips meaning if cut), so naive frequency scoring gets negation wrong. We tried fancier auto-scoring (IDF, embeddings, surprisal, Shapley values) and each one broke in different sentences — one even ranked "kubernetes" below "the." Ended up back at a simple, deterministic preserve-list because it's the only thing that was reliably correct.
Accomplishments we're proud of
Zero model calls — fully deterministic, instant, no latency or cost added to the thing it's trying to save cost on. Negation/question words are protected by hardcoded rule, not a probabilistic guess.
What we learned
Token reduction and meaning preservation are different goals that trade off against each other. Simple + transparent beat clever + statistical in basically every test we ran.
What's next
Dedup repeated context in long documents/chats (bigger win than word deletion). POS tagging instead of fixed lists. Real tokenizer instead of word-count estimate. Fix the README vs. code level-mismatch bug.
Built with
Python, Flask, regex, HTML/CSS/JS, argparse
Log in or sign up for Devpost to join the conversation.