Source grounding display

Problem Context

• Grounding can involve web search, file search, vector-store retrieval, enterprise knowledge sources, user-uploaded documents, selected topic sources, or policy-controlled source access.
• Retrieved sources are not always used in the final answer, and cited claims may not cover every answer claim.
• A source can be stale, permission-limited, missing, not searched, low quality, outside the selected scope, or inconsistent with another source.
• Grounding display often appears inside chat, streaming answers, research workspaces, review queues, document editors, and support-agent answer panels.
• Users need to know source coverage before applying generated output to legal, policy, medical, financial, operational, or customer-facing decisions.

Selection Rules

• Choose source grounding display when users need answer-wide evidence coverage, source scope, retrieved-versus-used source status, or unsupported-claim visibility.
• Use citation display when the task is inspecting one claim's citation marker and source preview.
• Use chat interface when the task is managing conversation turns, history, composer behavior, or follow-up context.
• Use streaming response when the task is exposing partial generated output and response lifecycle while text arrives.
• Use related links when destinations are optional onward navigation and not evidence used by the answer.
• Show grounding scope before or beside the answer when source selection materially changes what the answer can prove.
• Show unsupported and partially supported claims when the answer contains claims that are not fully backed by retrieved evidence.
• Do not show a global grounded badge unless all claim-support rules, source scope rules, permissions, and freshness checks used by the product are satisfied.
• Do not replace grounding coverage with a single confidence, quality, or trust score.

Required States

• Default grounded state with source scope, searched sources, retrieved sources, used sources, and supported-claim count.
• Partially grounded state with at least one partially supported claim and explanation.
• Unsupported claim state with claim text, no supporting source found label, and review or regenerate action.
• Pending grounding state while retrieval, web search, file search, or source matching is still resolving.
• Not searched or out-of-scope state when a source corpus was excluded by settings, topic, policy, or user choice.
• Permission-limited source state where a source exists but cannot be opened or quoted by the current user.
• Stale, missing, retrieval failed, conflicting source, and source quality warning states.
• Mobile collapsed grounding panel state that preserves answer position and source coverage counts.

Interaction Contract

• The grounding display names the source scope that was available for this answer, not just sources the user might expect.
• Searched, retrieved, used, cited, and unused sources are visually and textually distinct.
• Each unsupported or partially supported claim can be located in the answer and connected to the relevant source status.
• Regeneration, source-scope changes, document updates, permission changes, and answer edits invalidate or re-run grounding status.
• Copy, apply, publish, and export actions can include grounding status or warn when unsupported claims remain.
• Permission-limited sources explain what is hidden without leaking protected excerpts.
• Keyboard and touch users can open grounding coverage, inspect source groups, jump to claims, and return to the answer without losing context.
• Grounding status changes are announced as meaningful status messages without moving focus unexpectedly.

Implementation Checklist

• Model source scope, retrieval calls, searched sources, retrieved sources, used sources, cited sources, unsupported claims, partial support, conflicts, permissions, freshness, and grounding status as separate structured data.
• Track the answer text or claim ranges that each grounding status applies to, and update them after regeneration or editing.
• Render source groups for searched, retrieved, used, unused, permission-limited, stale, missing, and conflicting sources.
• Expose unsupported and partially supported claims with jump-to-claim, review, regenerate, change source scope, and report issue actions.
• Show pending grounding during streaming and only mark answer-wide coverage after source matching and final text settle.
• Preserve selected claim, source group, panel state, focus, and scroll when users expand or collapse the grounding display.
• Test no sources, too many sources, stale sources, private sources, conflicting sources, no supporting source, changed source scope, regenerated answer, mobile layout, high zoom, keyboard, and screen reader paths.

Common Generated-UI Mistakes

• Showing a global Grounded badge when only some claims have evidence.
• Mixing searched, retrieved, used, and related sources into one undifferentiated list.
• Hiding unsupported claims because some citations exist elsewhere in the answer.
• Treating grounding as a confidence score instead of evidence coverage.
• Omitting source scope so users do not know which corpus was searched.
• Making permission-limited sources disappear instead of explaining the access gap.
• Showing source coverage only after users already copied or applied the answer.

Critique Questions

• Can users tell which source scope was searched for this answer?
• Can users distinguish searched, retrieved, used, cited, unused, missing, stale, and permission-limited sources?
• Which claims are unsupported or only partially supported, and can users jump to them?
• What invalidates grounding status after regeneration, editing, source updates, or permission changes?
• Does the UI avoid turning grounding into a single confidence score?
• Would citation display, chat interface, streaming response, related links, or a data visualization solve the actual task better?