A reference for everything you can do inside a conversational AI session—interacting with responses, response actions, managing history, and configuration.
Conversational AI features
This page covers every capability available in a conversational AI session—how to interact with responses, control output, and manage your conversation history.
Prerequisites
Before you begin, make sure you have:
- Set up conversational AI for your tenant. See Set up conversational AI.
Response modes
Conversational AI answers in one of three response modes. The mode selector sits in the chat input in Atlan chat, Copilot on asset pages, and the browser extension. Your choice is shared across those surfaces and carries across conversations until you change it, and a conversation keeps the mode it was started in. Slack and Microsoft Teams don't expose the selector.
- Quick: optimized for speed. Use it for quick lookups—what a table contains, who owns an asset, whether a dataset is certified.
- Adaptive: matches effort to the question. Adaptive answers on the fast path and moves to deeper reasoning when the question calls for it, so you don't have to judge each question yourself.
- Deep: the default. Thinks through every question in more steps before answering—slowest, but most thorough. Use it for multi-part questions, lineage and impact analysis, or anything that spans several assets.
All three modes ground their answers in the catalog metadata you have permission to see, and link to the supporting assets where there are any. Some answers—a how-to question about Atlan itself, or a question the catalog can't answer—have no asset to cite.
Response modes are enabled per tenant. If you don't see the mode selector in the chat input, contact Atlan support.
What applies in each mode
| Behavior | Quick | Deep | Adaptive |
|---|---|---|---|
| Custom instructions | Applied | Applied | Applied |
| Skills | Not applied | Applied | Considered on every question, and applied on either path when relevant |
If a question depends on a skill—a repeatable, task-specific behavior your team has defined—ask it in Deep or Adaptive mode. Adaptive looks for relevant skills before it decides how to answer, so a skill can be used whichever path the question takes. In Quick mode, conversational AI answers without that skill.
Think harder
When you're on Adaptive and an answer isn't deep enough, click Think harder below it. Conversational AI re-asks that same question on the advanced agent and adds the new answer as the next turn in the same conversation—the first answer is kept, so you can compare the two. This is a one-off re-run: the conversation stays on Adaptive.
Think harder appears only on the most recent completed answer in a conversation on Adaptive, and only when that answer came from the fast path. It doesn't appear on answers Adaptive already sent down the deeper path, on older turns, on answers that errored, or on answers you stopped.
Tool approval
When conversational AI needs to use a tool mid-answer—to write metadata back to Atlan, or to query a connected external app—it asks first. A card appears in the reasoning chain naming the tool and, for a connected app, that app's name and logo (for example, "Atlan wants to use Search from Glean").
- Allow (Atlan's own tools) or Allow for this chat (connected apps): approves the call. Approving a connected app once means the rest of that conversation can use it without asking again—a new conversation asks again.
- Deny: blocks the call for this request. The tool may ask again on a later turn.
Click Show request on the card to see the exact request before deciding.
Once you decide, the card collapses to a one-line result: allowed, denied, or failed with a reason.
Connect external apps
An admin can connect external MCP servers—such as Glean—from Admin → Integrations → MCP, so conversational AI can call their tools alongside Atlan's own. Once connected:
- Results from a connected app appear in the same reasoning chain as Atlan's own search steps, each tagged with the source app's logo.
- Links in a response that point to a connected app (Confluence, Jira, Slack, Notion, Google Drive, GitHub, Zoom, and others) render as a logo citation you can hover to preview before opening.
A banner on the chat home page surfaces which apps are connectable, or—once something is connected—reminds you that conversational AI can answer from your connected tools.
Interact with responses
These features are available inside the chat input and within each response in the chat panel.
-
@ mention an asset: Type
@followed by a table, column, dashboard, glossary term, or any catalog asset name to tag it directly in your question. This tells conversational AI exactly which asset you mean so it pulls the right metadata, lineage, and documentation rather than guessing. -
Citations: Every response includes source citations linking back to the actual assets in your Atlan catalog. Click any citation to open the original asset page and verify the information. To view citations, click the # icon on a response. Use the > icon to navigate between sources.
-
Click to know more: When conversational AI mentions a table, column, or other catalog asset in its response, the asset name is clickable. Click it to open the sidebar with full details about that asset.
-
Suggested questions: After each response, conversational AI suggests relevant follow-up questions based on your conversation. Click a suggestion to continue, or type your own. Context is preserved throughout the session.
Attach files and links
Attach a file or paste a public link in the chat input to give conversational AI something concrete to read alongside your question. An attachment stays available for the rest of the conversation, so you can refer back to it in later turns without re-attaching.
- Attach a file: Click + in the chat input and select Attach file, or drag a file directly onto the chat input. Supported formats: PDF, DOCX, TXT, MD, HTML, CSV, XLSX, PNG, JPG, GIF, WebP, and
.skillfiles. Maximum file size is 10 MB. - Attach a link: Paste a public URL directly into the chat input. Conversational AI reads the linked page's content for the conversation.
- Multiple attachments: Attach more than one file or link to a single message; conversational AI reads all of them together.
- Failed uploads: If a file fails to upload, the attachment shows the reason and a Retry option where retrying can help. You can't send a message while a file is still uploading or has a failed upload still attached—remove or retry it first.
Attachments are enabled per tenant. If you don't see the attach option in the chat input, contact Atlan support.
Response actions
These controls appear at the bottom of each response in the chat panel.
-
Copy: Click the copy icon at the bottom of any response to copy it to your clipboard. Useful for sharing answers in Slack, Jira, or a document.
-
Retry: Click the retry icon to regenerate a response. Use this when you want a different angle or have refined what you're looking for. You can also rephrase your question and ask again for a more targeted result.
-
Stop: Click the stop button in the typing box while a response is generating to stop it. Useful when the response is going in an unexpected direction or you want to ask a more focused follow-up.
-
Feedback: Use the thumbs up or thumbs down icons at the bottom of each response to share feedback. When possible, leave a short comment explaining what was helpful or what missed the mark.
Artifacts
Conversational AI can answer with a file instead of a chat message. Reports, documents, diagrams, and skills open in a preview panel beside your conversation.
-
Preview: Click View on an artifact card to open the artifact in the side panel.
-
Edit with AI: Describe the changes you want and conversational AI generates a new version in place. The earlier card is marked Revised so you can tell which version is current.
-
Download: Save the artifact to your machine in its original format.
-
Create skills: Ask conversational AI to create a skill. It drafts the skill in the panel for you to review and save, and the skill is available in your next conversation.
For supported formats and the full workflow, see Artifacts in conversational AI.
Manage your conversations
-
View past threads: All previous conversations are saved and accessible from the chat sidebar. Click the clock icon at the top of the chat panel to browse past threads and pick up where you left off.
-
Rename a chat: Hover over any past thread, click the three-dot menu, and select Edit chat name to give it a meaningful name so you can find it later.
-
Delete a chat: Hover over any past thread, click the three-dot menu, and select Delete to remove conversations you no longer need.
Scope your conversation
The scope selector sits in the chat input. It controls what catalog data conversational AI searches when it answers your questions. Pick one of:
- All assets: No filter applied. Conversational AI searches your entire catalog. This is the default.
- A persona: Conversational AI only considers assets that the persona can access. Answers can differ from what you see when you browse the catalog yourself, because they follow the persona's access rather than yours.
- Assets available to me: Conversational AI only considers the assets you can access.
When you change the scope, a toast confirms the switch. For a persona, the message reads "Persona switched to <name>".
What you can pick depends on your role:
- Admins see every enabled persona in the tenant, sorted alphabetically.
- Non-admins see only the personas assigned to them.
You can change the scope until you send your first message. After that, the selector is read-only for the rest of the conversation. A conversation keeps the scope it was started in, so to use a different scope, start a new conversation.
Configuration
- Custom instructions: An admin can set persistent preferences for how conversational AI responds across the organization—for example, to always prefer verified assets, to exclude an asset type, or to answer in a particular language. Custom instructions apply to every new conversation automatically. To configure them, see Customize conversational AI.