Skip to main content

Canvas Panel

The canvas panel is a side panel that opens when you click a citation badge or the sources pill in the chat. It provides detailed views of citations, sources, tool calls, and the agent's reasoning trace.

Introduction

The canvas pairs with the chat to give you transparent access to everything behind a response. From it, you can drill into a single citation, browse all sources used for a message, inspect the exact tool calls and search queries the agent ran, and review the full reasoning trace from prompt to answer.

The canvas surfaces four views — Tool Call, Sources, Citation, and Run Summary — through a tab pill at the top. Tabs enable based on what is available for the current message, so you only see views that have content to show.

Opening the Canvas

You can open the canvas in three ways:

  • Click a citation badge (superscript number) in any assistant message — opens the Citation Detail view for that citation
  • Click the "N sources" pill below a message — opens the Sources List view
  • Click the compact summary line above a message — opens the Agent Run Summary view

The canvas layout adapts to your screen size. Press Escape to close it.

Canvas Toolbar

A floating pill at the top-center of the canvas provides tab buttons:

TabIconWhen Available
Tool CallWrenchAt least one tool call exists for the message
SourcesLibraryAt least one citation exists
CitationQuoteAt least one citation exists
Run SummaryActivityAlways available when canvas is open

Tabs are enabled or disabled based on the data available for the current message. The active tab is highlighted with the primary color.

Citation Detail

The citation detail view shows a single cited source in depth:

  • Citation number badge — the display index
  • Source title or document ID — links to the source document if a URL is available
  • Nexset name — which nexset this citation originated from
  • Relevance score meter — color-coded: green (≥80%), yellow (≥50%), grey (<50%)
  • Snippet display — JSON snippets appear in an expandable tree viewer; plain-text snippets appear as styled blockquotes
  • "View source" link — opens the source URL in a new tab (when available)
  • Metadata table — document ID, raw relevance score, citation ID, chunk count
  • Previous/Next buttons to step through citations
  • "All sources (N)" link to return to the Sources List
  • Position indicator: "3 / 7"

Sources List

The sources list shows all cited sources for the current message:

  • Each source displays: citation number, title, nexset name, relevance score badge, and a two-line snippet preview
  • JSON snippets render as formatted code; plain-text snippets are clamped to two lines
  • The currently selected citation is highlighted with a ring and glow
  • Click any source to open its Citation Detail view

Tool Call Detail

The tool call detail view lets you inspect exactly what the agent searched and what it found.

  • Display name and status chip — Running (with spinner), Complete, or Error
  • Copy trace button — copies the raw tool call result to clipboard

Tool Call Ledger

When multiple tool calls exist, a row of clickable tabs lets you switch between them.

Collapsible Sections

Five sections are available, with Summary, Query, and Results expanded by default:

SectionContent
SummaryHigh-level overview of the tool call
QueryThe search query or SQL that was sent to the nexset
ResultsParsed results from the tool call
SourcesCitations generated by this specific tool call
RawRaw JSON input and output

Agent Run Summary

The run summary shows the complete reasoning trace — every step the agent took to produce its answer:

  • Thinking phase — the model's chain-of-thought
  • Tool call cards — each search operation with status, elapsed time, query, and result count
  • Generating phase — when the model started composing its answer

This is the same content as the Agent Timeline in the chat, but displayed in the canvas for detailed review.