Overview - ScaleDown

Cookbooks are end-to-end recipes: real input, a ScaleDown call, and the response you get back, in Python. There are 16, one for each pairing of data type and model. Run them from scaledown.ai/cookbooks; you need an API key (see the quickstart).

The four models

Each model has its own job. Compress runs before your LLM; the other three each stand in for an LLM call.

Compress

Runs in front of your LLM. Trims retrieved or pasted context down to what the question needs, so the LLM call uses fewer tokens and returns faster. Endpoint: POST /compress/raw/.

Summarize

Replaces an LLM call. Turns long, messy text such as transcripts, threads and filings into a short abstractive summary. Endpoint: POST /summarization/abstractive.

Extract

Replaces an LLM call. Pulls the fields you define out of unstructured text and returns them as structured data, exactly as written in the source. Endpoint: POST /extract.

Classify

Replaces an LLM call. Scores text against labels you define with a yes/no rubric and returns a probability for each, plus a top_label for routing. Endpoint: POST /classify.

Every recipe has the same five parts

Part	What it tells you
The move	What the model does here, in one line
Model role	Whether it sits in front of your LLM or replaces a call
Why it fits	The constraint that makes this the right model
How it goes	The pipeline, step by step
Payoff	The result. Only two recipes quote measured benchmarks: Voice Summarize (QMSum) and Documents Compress (FinanceBench). The rest are described qualitatively.

Browse by data type

Voice & Calls

Call recaps, caller routing, CRM fields from sales calls.

Chats & Tickets

Thread digests, triage and action items.

Files & Docs

Briefs, document tagging and clause extraction.

Web Pages

Boilerplate stripping, relevance filtering and enrichment.

Browse all cookbooks: Filter by data type or model on scaledown.ai.

​The four models