AI Assistant
The AI Assistant lives inside AI Chat — one unified workspace for human conversations, workspace channels, and AI sessions. Ask questions, generate content, call your tools, ground answers in your own documents, and switch between text, voice, and avatar — all from a single surface that understands your organization.
The assistant lives right inside AI Chat — one home for every conversation — and an embeddable assistant dock appears throughout the app, right next to the feature you're working on.
What the AI can do
Answer questions
Ask anything — general knowledge, code, or questions grounded in your own organization documents and past conversations.
Call your tools
The model decides when to search the web, generate an image, query a knowledge folder, or run an integration action, then reasons over the result — a full tool-calling loop.
Write & edit
Draft emails, proposals, or code, then refine for tone, clarity, or length. Hand off long-form writing to a Canvas beside the chat.
Read your files
Upload a document or attach a knowledge doc, and the AI reads and reasons over its extracted text in context.
Run deep research
Open Deep Research beside the chat for a planned, multi-step web investigation with cited, structured output.
Use your knowledge
Retrieval-augmented answers pull the most relevant chunks from your knowledge folders and cite them inline.
Starting an AI chat
- Open AI Chat
Click "AI Chat" in the navigation rail (the chat icon). The workspace lists your recent and archived sessions, with a "New chat" button at the top — much like a familiar chat history.
- Start a new session
Click "New chat" to create a fresh session. Each session keeps its own history, model choice, and custom instructions.
- Pick your model
Use the model selector at the top of the conversation. Recent models surface first; the model used for each AI reply is recorded on the message.
- Chat and collaborate
Ask your question. Attach files, type # to add context, toggle a tool from the tool strip, or open Canvas or Deep Research in a side panel — all without leaving the session.
Choose your model
YOffice is model-agnostic. Administrators enable providers and seed the model list per organization, and you pick a model per conversation from whatever is available. The selector is data-driven — new models appear as soon as an admin adds them, with no app update required.
| Provider | Example models | Notes |
|---|---|---|
| AnthropicDefault | Claude Sonnet 4.5 | Nuanced writing, analysis, long documents, strong tool use. |
| OpenAI | GPT-4o mini | Fast, cost-efficient general reasoning and iterative work. |
| Gemini 2.0 Flash | Low-latency responses and large context windows. | |
| Groq | Hosted open models | Very high token throughput for latency-sensitive flows. |
| Ollama Cloud | gpt-oss:120b, gemma3:4b, llama3.3:70b | Managed open-weight models; larger options like qwen3 / deepseek-v3.1 can be enabled. |
| Ollama (local)Self-hosted | llama3.2, llava (vision) | Runs on your own infrastructure — data never leaves your network. |
Beyond your per-conversation choice, admins can route specific task types (summarisation, embeddings, classification, and more) to different models, so the right model handles each job automatically.
Tool-calling loop
The assistant doesn't just answer from training data — it can act. When a question needs more than the model knows, it calls a tool, reads the result, and continues reasoning, repeating as needed before replying. Each tool call is shown as a card (or in a side panel) so you can see exactly what the AI did.
The tool strip beneath the composer exposes (subject to availability):
- Web Search — live results via Brave Search.
- Image Generation — create images from a prompt.
- Data Analysis — analyze CSV or tabular data and generate charts or statistics.
- Canvas — open or write to a document beside the chat.
- Deep Research — launch a multi-step research task.
- Knowledge — scope which knowledge folders are searched.
Add context with #
Type # in the composer to open the context picker and attach live context to your message. Depending on the surface you can add an agent, a saved prompt, a knowledge folder, an integration or custom integration, a memory, a workflow, or a skill. Use @ to mention people or the AI. Mentions are stored as structured references, not just text, so the AI resolves them precisely.
Attach files
Drag a file into the composer or attach a document already in your Knowledge Base. YOffice extracts text from a broad range of formats — PDF, Office and OpenDocument files, spreadsheets, and many plain-text formats — before sending it to the model. Knowledge Base page for the full supported-format list.
Image understanding depends on the selected model (for example a vision model such as Ollama's llava, or a multimodal cloud model). Document files are read by extracting their text, so text-based sources give the most reliable results.
Research & Canvas beside the chat
Two companion surfaces open next to your conversation (a side-by-side split on desktop, an overlay on mobile):
- Canvas — ask the AI to draft a document and it opens in a Canvas panel you can edit together..
- Research Mode — kick off a planned, cited web investigation..
Text, voice, and avatar
Each session has a per-message input mode that's remembered for the conversation:
| Mode | What it does |
|---|---|
| Text | The standard typed chat experience. |
| Voice | Speak instead of type. Voice routes through the same chat engine, so it has full parity with text — #-context, @-mentions, web search, files, and memory all work. Unified Voice + Chat. |
| Avatar | An animated assistant tile (powered by Simli) speaks responses aloud with a talking-head video. |
Productivity touches
- Message search across all your chats.
- Per-message translation into another language.
- Undo send within 30 seconds of sending.
- Spoken-vs-typed badge on each message.
- Custom instructions per chat, plus a memory sheet the AI can draw on.
Skills, Agents, and Prompts
Three reusable building blocks shape how the assistant behaves. They are distinct — don't confuse them:
| Concept | What it is |
|---|---|
| Skills | Lightweight reusable instruction snippets you toggle on for a chat. A skill has no model, no knowledge, and no tools — it only steers behaviour. Skills. |
| Agents | Persistent AI personas that do bundle a model, system prompt, knowledge folders, tools, and sub-agents. Agents. |
| Prompts | Reusable prompt templates with {{variables}} you insert via # or save from a message. Prompt Library. |
Select a self-hosted Ollama model and inference runs entirely on your own infrastructure — nothing is sent to a cloud provider. Cloud models (Anthropic, OpenAI, Google, Groq, Ollama Cloud) process data through their respective providers under your organization's agreements. Security & Privacy for full details.
Context is tenant-scoped
Everything the assistant sees is scoped to your organization and your access. Knowledge is only retrieved from folders you can read, and the AI can never reach data in another tenant.