AI Chat Interface

When to use

Use for conversational AI, assistants, support bots, or chat surfaces that generate or retrieve responses.

When not to use

Do not use chat as a universal interface when a form, table, or direct workflow better supports the user's task.

Inputs and assumptions

Identify user intent, model/provider, streaming behavior, available tools, source grounding, privacy rules, conversation persistence, and failure semantics.

Workflow

1. Set expectations about capabilities, limitations, and data handling.
2. Distinguish user messages, assistant responses, citations, and tool activity.
3. Show generation progress and allow cancellation; do not imply completion before the stream ends.
4. Make citations inspectable and connect each source to the claims it supports.
5. Provide retry/edit/regenerate controls with clear consequences.
6. Preserve conversation context and label failures without discarding user input.
7. Support keyboard navigation, announced new output, and long-response reading.

Decision rules

• Assistant output may be wrong; avoid trust signals that imply certainty.
• Show meaningful tool status without exposing secrets or unsafe internals.
• Make data retention and conversation deletion behavior clear.
• Do not silently drop partial output or user messages.

Example

During a tool call, show a pending status, then a completed or failed result; render the final answer as complete only after the operation finishes.

Common failures

• A spinner with no cancellation or progress context.
• Sources shown but not tied to claims.
• Streaming UI that duplicates or loses content on retry.

Validation

Test streaming, cancellation, retry, tool failure, source links, keyboard and screen-reader announcements, long responses, and persistence rules.