Files & Docs - ScaleDown

Filings, contracts, reports and case files, including OCR output. Four recipes, one per ScaleDown model.

Compress

Sits in front of your LLM. Open the full Python recipe.


Why it fits	Filings, contracts, and reports are long and repetitive across versions; you are paying to process restatement.
How it goes	Document or post-OCR text in → Compress trims to what the task needs → the LLM runs on the trimmed context.
Payoff	On FinanceBench, Compress with GPT-5.4 delivered +8% accuracy with 97% fewer tokens vs. GPT-5.4 alone.

Replaces an LLM call. Open the full Python recipe.


Why it fits	The reader needs the gist of a dense document, not the whole thing.
How it goes	Document in → Summarize produces the brief → surface it to the user or downstream step.
Payoff	Document briefs without a frontier call per doc.

Replaces an LLM call. Open the full Python recipe.


Why it fits	In regulated work a wrong field is a real problem; Extract returns the values as written.
How it goes	Document in → tell Extract the fields → it returns them structured → into the system of record.
Payoff	Clean structured data off unstructured docs.

Replaces an LLM call. Open the full Python recipe.


Why it fits	At high document volume you want a cheap first pass that decides what each doc is before deeper work.
How it goes	Document in → Classify assigns the label → route to the right pipeline or reviewer.
Payoff	Fast triage at page-volume scale.