Manage and inspect the URLs, documents, and snippets used as session context
| Prop | Type | Default | Description |
|---|---|---|---|
references | ReferenceItem[] | required | Array of reference entries with id, title, optional type, optional location, optional excerpt, optional selected, and optional relevance |
onSelectReference | (referenceId: string) => void | undefined | Optional callback fired when a reference is selected |
onRemoveReference | (referenceId: string) => void | undefined | Optional callback fired when a reference is removed |
title | string | "References" | Optional heading for the references panel |
variant | "list" | "cards" | "list" | Display style for reference entries |
showRelevance | boolean | true | Whether to show relevance indicators for each reference |
References make the AI's working context explicit by listing the documents, URLs, and snippets currently informing a session. Instead of hiding source context behind the scenes, this pattern lets users inspect relevance, open sources, remove noisy references, and control what evidence the model can rely on.
Variants
- Reference list — A simple stacked list optimized for scanning titles, excerpts, and quick actions.
- Reference cards — Card-style references that surface richer metadata and work well in dashboards.
- Selectable context — References can be toggled or selected to narrow the active context for the next run.
Use Cases
- Reviewing which uploaded docs are influencing a generated answer
- Removing low-quality or outdated sources from session context
- Comparing relevance across internal and external references
- Auditing evidence used in regulated workflows
- Curating focused context before expensive generation steps
Best Practices
- Show concise excerpts so users can judge relevance without opening every source.
- Expose source type and location to reduce ambiguity.
- Allow quick removal to prevent stale context from polluting outputs.
- Use relevance indicators as guidance, not as absolute truth.
- Pair references with citation views for output-level traceability.