Creating Columns

2 min read

Draft branch. This page explains what a column is for. The tutorials under it show how to write the prompt, call a subagent, set review labels, and turn on the hallucination score.

Overview

Each column is one job on the row. The header name is what people see. The filter, renderer, and prompts are what the agent and the UI do with the cell.

Columns a questionnaire usually needs

Column

Job

Typical setup

Section

Groups rows by the part of the questionnaire they came from

ValueMatchFilter, editable, rerun_trigger: true

Question

Holds the question text

PartialMatchFilter, editable, rerun_trigger: true

Answers

The generated answer

System prompt plus assistant prompt. rerun_trigger: false so editing the answer does not immediately regenerate it.

References

The sources behind the answer

ReferenceFilter, not editable. This column has no tutorial of its own yet.

Hallucination Score

Whether the answer is supported by those sources

HallucinationFilter, not editable. See Configure Hallucination Check.

Review Status

Where the row is in review, and whether it is locked

ReviewStatusDropdown. See Custom Review Statuses.

Assignee

Who owns the row

CollaboratorDropdown and AssigneeFilter

Fields you set on every column

Field

What it does

name

Header text. Other config refers to this name, not to a column index, except ingestion_config which still uses indexes.

filter

How the column can be filtered. The component list is on Agentic Table Components.

cell_renderer

How the cell is drawn. Use SelectableCellRenderer when the person must pick one answer among several answer columns.

editable

Whether a person can type in the cell.

rerun_trigger

Editing this cell reruns the answer columns on the row. Turn it on for inputs such as Section and Question. Leave it off for outputs.

width

Width in pixels.

is_selected_by_default

Only matters for SelectableCellRenderer. The selected answer is the one that exports.

system_prompt and assistant_prompt

Instructions for an answer column. See Writing Prompts.

language_model

Model for this answer column. Leave it unset to use the table's model. Ignored when subagent_id is set.

reasoning_effort

none, low, medium, high, or xhigh. Only sent if the column's model supports that level; otherwise the model uses its own default. Ignored when subagent_id is set.

subagent_id

Set this instead of the prompt pair when another space does the work. See Setting Up Subagents.

search_config

Which knowledge the column may read. This still needs its own page.

Prompts can only read cells to their left, on the same row. Put Section and Question to the left of Answers.

Last updated