Answer types
Pick the narrowest type that holds the answer. A narrow type is cheaper, easier to check, and lets a decision model do the work.
Classify
Section titled “Classify”Puts the row in a known box. Any model answers these, each answer with a probability (0 to 1) that it is right; decision models return it, LLMs are asked for it.
| Type | Answer | Columns |
|---|---|---|
| Yes/no | true or false |
<column>__<model>, _proba |
| Pick one | one of the options | the option’s name, _proba |
| Pick any | any of the options | one true/false column per option: <column>_<option>__<model> |
| Scale | one of ordered levels | the level’s name, _proba |
Options have a name and, when the name alone could be read two ways, a description. A scale’s
options are its levels, listed lowest first (low, medium, high), and the answer is one of
them, not a number. Options must
cover every row: add not applicable or can’t tell when the question doesn’t apply to some rows,
so the model never has to guess.
Extract
Section titled “Extract”Pulls typed facts out of the row. LLMs only. Fields have a name, a type (text, number, boolean,
date, list, or nested items) and a description; a field is empty when the text doesn’t say. One field
makes one column; several make <column>_<field>__<model> each. Several facts from the same rows
belong in one extract task: one call answers them all.
Produces new text from the row: a summary, a translation, a rewrite. LLMs only. The column holds the text.
Models
Section titled “Models”Models come from OpenRouter, and are picked by name per task. Two kinds:
- Decision models (such as
typesafe/jev-1.14, the default) answer Classify questions fast and cheaply, and are often as good as an LLM on clear-cut questions. - LLMs (the default is
anthropic/claude-haiku-5.5) answer everything, and are needed for Extract and Write.
Run two or three on the same rows and compare before choosing.