Skip to content

RAG Chatbot

Pick an agent to start a conversation, then use streaming replies, image attachments, source cards, feedback, and the debug timeline.

The RAG chatbot is where you talk to a published agent. The agent searches your organization's documents to find supporting evidence, then generates an answer from that evidence in real time. As you ask a question, the answer streams in token by token, and source cards appear below it showing which documents were referenced.

Every conversation is tied to a single published agent. Only agents that were set to published in the Agent Builder show up here, and the agent you pick determines which document groups, retrieval tools, and answer style are used.

Starting a conversation

The main chat screen is split into an agent list on the left and a recent conversations list on the right. A new conversation begins when you choose an agent, and it is created and added to the list the moment you send your first message.

  1. Pick an agent

    Click the agent you want to talk to in the left-hand list. Only published agents appear, and each entry shows a name and a short description.

  2. Send your first question

    Type a question in the input box at the bottom and send it. The new conversation is created the instant you send the first message, and a title is generated automatically from your opening question.

  3. Read the answer

    The answer streams in live. Once it finishes, source cards for the referenced documents appear below it, along with copy, regenerate, and feedback buttons.

  4. Keep the conversation going

    Send follow-up questions in the same view. Conversations are saved automatically, so you can reopen them anytime from the recent conversations list.

If the agent list is empty, no agents have been published yet. Create and publish an agent in the Agent Builder, and it will appear here.

Composing and attaching

The input box supports both text questions and image attachments. While an answer is being generated, the send button turns into a stop button so you can interrupt at any time.

Send a message

Type your question and press the send button or the Enter key. You cannot send a new message while a response is still being generated.

Attach images

Use the image button to attach a picture to your question. JPEG, PNG, GIF, and WebP formats are supported; attached images show as previews and can be removed before sending.

Stop generation

Press the stop button while an answer is streaming to halt generation. Whatever has been produced so far stays on screen.

Live streaming

Answers appear token by token instead of waiting for completion, and the view scrolls down automatically as new content arrives.

  • Send messageEnter
  • Insert a line breakShiftEnter

Sources and citations

When an agent answers from your documents, numbered citation markers appear inline in the text, and a set of source cards for the referenced documents appears below the answer. This lets you verify the evidence behind every reply.

Each [N] citation in the answer maps one-to-one to the numbered source card below.
ElementWhat it does

Inline citation number

A small numbered badge placed within the answer. Clicking it highlights the matching source card and scrolls the view to it.

Source card

Shows a referenced document as a card. Documents and images are grouped separately, and clicking a card opens a preview of the original.

Page badge

Tap a page number on a card to expand the actual text chunk used in retrieval, so you can see exactly which passage the answer relied on.

Graph context

When the agent uses the knowledge graph, a separate section below the answer shows a graph summary and related entities, with a link that jumps straight to the Knowledge Graph Explorer.

If a citation number looks greyed out, the matching source could not be found in the card list. The answer text is still valid; it simply means there is no card linked to that marker.

Acting on an answer

A row of action buttons sits below each answer. They appear on hover, and feedback helps improve answer quality over time.

Copy

Copies the full answer text to your clipboard. Your own messages have a copy button too, so you can grab the question you typed.

Regenerate

Generates a fresh answer to the same question. Use it when a reply is unsatisfactory or you want to see an alternative phrasing.

Like and dislike

Rate whether an answer was helpful. Choosing dislike opens a dialog where you can pick a reason such as inaccurate, irrelevant, incomplete, or outdated and add a free-form comment.

Edit your message

Questions you sent can be revised using the edit button that appears on hover. This is handy for refining a prompt to get a more accurate answer.

Debug timeline

The debug timeline walks through the steps an answer goes through before it reaches you. It shows the flow of the RAG pipeline, such as request received, configuration loaded, document retrieval, image search, and the LLM call, along with how long each step took, which helps you check that retrieval is behaving as intended.

The debug timeline appears in chat only when debug mode is turned on in your organization settings. An admin can enable debug mode there to make it available.

Frequently asked questions