AI Chat

The AI chat experience is available as:

• a standalone chat UI (/ai-chat, /ai-chat/{kb_id})
• an embedded/compact tester inside the admin KB pages (when enabled)

This page explains entry points, sessions, streaming, cancellation, sources/citations, and feedback.

Entry Points

/ai-chat (chat home)

Common behaviors:

• shows a KB selector (often only “ready” KBs)
• may require auth (depends on client-side auth state and deployment policy)

/ai-chat/{kb_id} (KB-scoped chat)

Common behaviors:

• chat is locked to one KB
• used for targeted testing and sharing

Sessions

Sessions allow multi-turn conversations.

The session_id URL parameter

When present, the UI uses it to reload prior messages and continue the same conversation thread.

Operational rules:

• supply session_id to “resume” a conversation
• omit session_id to start a new conversation

Streaming Answers

Chat answers may stream progressively instead of waiting for one final block of text.

Typical UI behavior:

• a “thinking/loading” bubble appears immediately
• text content grows over time
• at the end, sources and message actions appear

Progress States

Some UIs show progress states such as:

• understanding the question
• searching the knowledge base
• selecting the best information
• preparing the response

These are UX hints and do not necessarily map 1:1 to internal steps.

Sources / Citations

If a KB allows it, answers can show sources used to support the response.

If the KB is configured to hide sources:

• the assistant may still use retrieval internally
• but the UI suppresses visible citations/snippets

Feedback

Messages can collect feedback such as:

• thumbs up / thumbs down
• optional reason (inaccurate, incomplete, irrelevant, too slow, lacked sources, other)
• optional free-form comment

Operational advice:

• use feedback patterns to prioritize KB fixes (extraction issues often show as “irrelevant”)
• validate “too slow” feedback against trace timing

Cancellation

While a response is in flight, the UI may offer “Cancel”.

Cancellation behavior depends on backend support but typically:

• stops streaming
• marks the message as cancelled/aborted
• leaves the session otherwise intact

Identity

Depending on deployment:

• admins may chat using authenticated identity
• public flows may chat using a lightweight username/guest id

Troubleshooting

The UI loads but messages never stream

Check:

• DevTools → Network for failing /api/... calls
• whether window.API_BASE_URL points to the correct API origin

Citations don’t appear

Check:

• KB config: sources hidden vs shown
• whether retrieval returned any chunks (use tracing)

Chat feels slow

Check:

• first query vs second query (warm-up effects)
• tracing timing breakdown (retrieval vs generation)