AI Chat

AI Chat is the primary way to interact with Fabric's AI capabilities. It supports real-time streaming, document uploads for context, MCP tool execution, and multiple reasoning depths.

Chat Modes

Fabric offers two chat modes, each suited to different tasks:

Direct Chat

Best for quick conversations and straightforward tasks:

• Real-time streaming responses
• Direct MCP tool execution
• Document upload and RAG context
• Semantic memory integration
• Conversation persistence

Orchestrator Chat

Best for complex, multi-step tasks:

• Durable workflow-based execution
• Multi-agent task coordination
• Human-in-the-loop approval flows
• Task plan visualization and progress tracking
• Step-by-step reasoning with verification

Reasoning Depth

The depth of AI reasoning depends on the model and task type you have configured. Models with reasoning capabilities (such as o1 or DeepSeek R1) provide deeper analysis with step-by-step verification, while fast models (such as GPT-4o-mini or Llama via Groq) optimize for speed.

Configure model preferences per task type in Settings → AI Providers. See AI Providers for details.

Starting a Conversation

Document Upload

Upload documents to give the AI context about your work. The AI uses RAG (Retrieval-Augmented Generation) to reference your documents when responding.

Supported Formats

Maximum file size: 10 MB per file

How Document Context Works

1. Upload — Drag and drop files or click to upload
2. Processing — Files are automatically extracted, chunked, and embedded
3. Ready — Documents become available for semantic search
4. Retrieval — When you send a message, the AI retrieves the most relevant sections from your documents and includes them as context

Document Status

MCP Tool Integration

When MCP servers are configured, the AI can use their tools during conversations.

Tool Suggestions

As you type, the AI suggests relevant MCP tools based on your input:

• Suggestions appear after 5+ characters
• Shows tool name, source server, and description
• Click a suggestion to include it in your message

Tool Execution

During a conversation, the AI may call MCP tools to complete tasks:

• Tool call cards display the tool name, arguments, and results
• Some tools require confirmation before execution
• Results are incorporated into the AI's response

Conversation Management

History

Past conversations are listed under History: in the sidebar of the full page and, in the Fabric Agent drawer, behind the Conversation history button in its header. Both show the same list.

• Mine, Shared with your organization and Archived switch which conversations are listed.
• Search looks at titles. It does not search what was said inside a conversation.
• Load more brings the next page when the list is long. If a page cannot be loaded, the conversations already shown stay and Try again loads it.
• Pin a conversation to keep it at the top of your list. The button keeps its name and shows whether the conversation is pinned.
• Rename changes the title Fabric generated.
• Archive moves a conversation out of the way. Open Archived and choose Restore to bring it back. Delete removes it for good, and asks first.

Opening a conversation reopens it on the engine it ran on, Direct or Orchestrator. In the drawer, a conversation you reopen continues with the context of the page you are on now, not the page it began on, and a note above it says so. While a reply is still being written, the drawer keeps the current chat and does not open another. If a conversation is gone, the drawer says so; if it could not be loaded, it says that instead and offers Try again.

Sharing a conversation

A conversation is private to the person who started it. Choose Share with your organization from its menu to let everyone in your organization read it. The same menu item reads Sharing settings once a conversation is shared, and you can make it private again at any time.

Conversations others shared appear under Shared with your organization, marked Shared by and the name of the person who shared them. One you shared yourself reads Shared with your organization in your own list. A conversation someone else shared opens read-only: there is nothing to send and nothing to approve. Continue in my own chat starts a new conversation of your own that begins from it and runs with your tools and permissions, not theirs, on the engine your own chats use now rather than the one the owner used. Only the person who started a conversation can rename, archive, share or delete it.

Checkpoints

Create checkpoints to save conversation state:

• Create a checkpoint at any point in the conversation
• Restore to a previous checkpoint to try a different approach
• Delete checkpoints you no longer need
• Track token usage at each checkpoint

Automatic Titles

Fabric automatically generates descriptive titles for your conversations after the first exchange, making them easy to find later.

Token Usage

The chat interface tracks token usage:

• Input tokens — Tokens in your messages and context
• Output tokens — Tokens in AI responses
• Reasoning tokens — Additional tokens used for deep reasoning
• Cached tokens — Tokens served from cache (faster, lower cost)
• Context estimate — How much of the context window is used

Workflow Execution

The AI can suggest and execute workflows during a conversation:

1. The AI identifies a relevant workflow for your task
2. A confirmation dialog appears with the workflow name and description
3. Click Execute to run or Cancel to skip
4. Workflow results are incorporated into the conversation

Multi-Tenant Support

AI Chat works in both personal and organization contexts:

• Personal (/app/nexus) — Your private conversations
• Organization (/app/{org}/nexus) — Organization-scoped conversations
• Each context has separate conversation history
• AI model preferences are resolved per context (personal or org-level)
• Document uploads are isolated by context

---

Next Steps