@ai-gui/plugin-resultset

Host-owned result tables for AIGUI: the numbers in an answer come from the query, not from the model retyping them.

Why it exists

A model that reads 4624290 out of a tool result and types it into a sentence has introduced a transcription step that nothing checks — and transcription is exactly where a digit goes missing.

@ai-gui/plugin-evidence proves which query ran. It does not prove the number in the prose came from it. This plugin closes that gap from the other side: the host emits the rows verbatim, and the prompt tells the model to point at the table instead of copying out of it.

Install

pnpm add @ai-gui/plugin-resultset

Usage

import { AIRenderer } from "@ai-gui/react"
import { resultset, resultsetPromptSpec, serializeResultsetFence } from "@ai-gui/plugin-resultset"

<AIRenderer text={answer} plugins={[resultset()]} />

1. Append resultsetPromptSpec() to the system prompt:

Never emit a ```resultset fence. The application appends result tables from the rows it really returned.
Do not retype figures from a table into your prose. Refer to the table by id, e.g. [[result:by_region]], and describe what it shows.

2. Append the fence server-side, from the rows you actually got back:

const fence = serializeResultsetFence({
  id: "by_region",
  label: "Revenue by region",
  columns: result.columns,
  rows: result.rows,
  source: "warehouse",
  truncated: result.truncated,
})
return answer + "\n\n" + fence

The fence

```resultset
{"id":"by_city","label":"Orders by city",
 "columns":["city","orders","amount"],
 "rows":[["Shanghai",2104,4624290],["Beijing",1460,3011880]],
 "source":"shop","truncated":false}
```

| field | | |---|---| | columns / rows | required; every row must be exactly as wide as the header | | id | what the prose points at — [[result:by_city]] | | label | caption, usually the question this answers | | source | where it ran | | truncated | more rows existed than were returned |

A row that does not match the header is rejected, not padded. A misaligned table puts a number under the wrong column, which is worse than no table.

null renders as a distinct, muted null rather than an empty cell — "we have no value" and "the value is empty" are different findings.

Options

| option | default | | |---|---|---| | maxRows | 200 | rows rendered per table (hard cap 500) | | showId | false | print the id beside the caption, so prose references are traceable |

Numbers

Numeric cells are right-aligned, tabular-nums, and grouped: 4,624,290, never 4.62429e+06. The same figure either way — but only one can be compared with the number below it at a glance, which is the entire job of a column of numbers.

Rows beyond the cap, and truncated, are both stated in the footer. A table silently cut to its first page is a wrong answer that looks complete.

Streaming

Complete-gated, like plugin-chart and plugin-evidence. A table that grows a row at a time reads, mid-stream, as a table that is already finished: a reader who looks away at the wrong moment sees four rows where there were nine, and nothing tells them so.

Styling

resultsetCss is a plain string of the defaults; the markup carries data-aigui-resultset* attributes so you can ignore it entirely. Colours use color-mix against currentColor, so the table follows the surrounding theme.

License

MIT