> ## Documentation Index
> Fetch the complete documentation index at: https://amd-gaia.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Chatting with GAIA

> The message box, what you see while GAIA works, and what you can do with an answer

Type a question, send it, and watch GAIA work through it. This page covers everything in the chat itself.

![A chat with a finished answer, the resource strip above it and the session panel on the right](https://mintcdn.com/amd-fe836e11/VjLPaDKuSjUBLb6M/images/agent-ui/answer.png?fit=max\&auto=format\&n=VjLPaDKuSjUBLb6M\&q=85\&s=14a7d065437bd9c825c2b108df79cc36)

## The message box

* **Message box:** **+** attaches files (or drag them in), the shield picks the
  permission mode, the model chip picks the model, and Send turns into Stop
  while GAIA is working. In a chat, the folder chip beside the model names the
  folder GAIA works in; click it to see and manage the chat's folders in the
  session panel.

`Enter` sends and `Shift+Enter` starts a new line. The rest are under [Shortcuts](https://amd-gaia.ai/docs/guides/agent-ui/documents.md#shortcuts).

## While GAIA works

* **While GAIA works** the reply builds up as a timeline. Short notes GAIA
  writes before a tool call ("I'll check the forecast first") appear in full,
  each step shows as one line with its own timer, and live reasoning scrolls
  by under **Thinking**. A line at the end says what the model is doing between
  steps. Picking up a skill gets its own highlighted line.
* **Once the answer is in**, the steps between two notes fold into one line
  ("Read web page, read file") you can open for the detail, and reasoning
  folds to "Thought for 8s".

## The line under an answer

* **The line under an answer** says how long it took and where it ran
  ("Answered in 40 s on your GPU", or "via Fireworks AI" for a cloud model).
  Hover or click it for the rest: when the first word came, how many steps GAIA
  took, and the exact numbers, including the tokens the whole turn read and
  wrote across every model call. "Steps" is the agent's own count, the same
  figure the TUI and evals report.
  Files the answer names by full path,
  such as a saved deliverable, open in your file explorer when clicked. So does
  a file the answer names only by file name, in code formatting, when it is one
  of the files that turn wrote; two written files with the same name stay plain.

## Type while GAIA works

You can keep typing while a reply runs. Press Enter and the message waits in a
queue above the message box, then goes out as its own turn as soon as the reply
ends, in the order you typed. To add a detail to the work already running, press
**Add to this reply** on a queued message: GAIA reads it at its next step
without starting over, though it may not answer it separately. If the reply
ends before reading it, it goes back in the queue. If the reply can't take it,
the message says so and waits to go out when the reply ends. Press Stop and anything still
queued goes back into the message box.

A stopped answer keeps what it wrote so far. **Continue** asks GAIA to carry on
from there, and **Try again** asks the question afresh.

Each chat keeps its own queue. Switch to another chat and the queued messages
still go out when their chat's reply ends, and the chat list shows how many are
waiting. They survive a reload of the app too. A message that can't be sent
stays in the queue marked **Not sent**, with the reason and **Send again**, and
the messages after it wait so the order holds. Deleting a chat deletes its
queued messages; the delete button says how many first.

## Edit a message or try again

To fix a sent message, hover it and choose **Edit message**, or press the up
arrow in an empty message box to open your last one. Change the text and press
Enter: GAIA answers the edited message, and the turns after it are replaced.
Escape leaves the message as it was. **Try again** on the latest answer asks
the same question again with the same model.

## Compare models on your question

To see how another model answers, use **Retry with…** on the latest answer and
pick a model downloaded on this PC. On a chat answered by a cloud service, the
menu lists that service's other models first, then the models on this PC; a
cloud model answers without loading anything here, and nothing goes to another
service. Both answers appear side by side with the time each took and, for a
model on this PC, its speed. **Keep this answer** on either one, and the chat
continues from it. The other model answers from what the first reply already
gathered and doesn't run tools again. It loads after the first answer, since
the model server holds one model at a time, so the reported time includes
loading it. Your next message loads the chat's own model again.

## Cards for recipes, workouts, plans and calculators

Ask for a recipe, a workout, how to fix something, a packing list, a trip plan,
a tip split or a loan payment, and the answer arrives as a card you can use
rather than a wall of text. The **Try** row on a new chat fills in an example
of each, and the first card you get carries a short tip on what it can do.

| Card | What you can do with it |
| - | - |
| **Recipe** | Change the servings and every amount rescales; switch between US and metric units; tap a time in a step to start a timer |
| **Workout** | Tick off equipment; count sets and run hold and rest timers |
| **How-to** | Tick off materials and tools; timers on any step with a time in it |
| **Checklist** | Tick items off; your ticks stay when you come back to the chat |
| **Trip plan** | A day-by-day timeline with a packing list to tick off |
| **Tip or loan calculator** | Change any number and the tip, per-person share or monthly payment updates at once |

**Start cooking**, **Start workout** and **Start** open the steps full screen,
one at a time, with a large timer for the step. Use the arrows or the ← and →
keys to move, and **Esc** or the **X** to leave. A running timer keeps going
when you leave. Every card can be copied as plain text, and recipes and guides
can be printed.

> *"Give me a recipe for banana bread"*
>
> *"Split a \$142 dinner four ways with a 20% tip"*

The model fills in the card and the app does the math, so amounts and payments
are exact even on a small local model.

## Images and audio

An image the agent generates shows in the answer; click it to see it full size.
Images you attach show as small tiles in the message; click one to see it full
size and use the arrows to step through the others in that message.
Music and sound effects it generates show a player with the file name and
length (see [Local Music and Sound Effects](https://amd-gaia.ai/docs/guides/audio-generation.md)). Both
still show when you reopen the chat. Only files GAIA saved in its own
generation folders are shown this way. Speaking to GAIA through a microphone
is not part of this; voice input is tracked in
[#702](https://github.com/amd/gaia/issues/702).

Recordings and video links can be transcribed in a chat too; see [Transcription & Recording Analysis](https://amd-gaia.ai/docs/guides/transcription.md#in-the-agent-ui).

## Read aloud

The speaker button on a reply reads it in a voice that runs on
this PC through Lemonade; press it again to stop. Code blocks and links are
skipped, so you hear the words. **Settings → Read aloud** picks the voice
(American or British English, female or male, with **Preview**), the speed
(0.5× to 2×), and whether every reply is read when it finishes. The voices
speak English, so a reply in another language is read with an English accent.
The first reading after a restart takes longer while Lemonade loads the voice.
On a PC without the voice model, the first press downloads it once (\~350 MB)
and says so under the reply; reading starts when it finishes. (The agent's own
`text_to_speech` tool never downloads it in the middle of a chat.)
The [terminal app](https://amd-gaia.ai/docs/guides/tui.md) uses the same voice and speed, and so does GAIA when you ask it to save speech to a file.

## Private chat

* **Private chat:** the eye button in a chat's header. An open eye means the
  chat is remembered as usual, and a click makes it private. While the chat is
  private the button is filled and shows a crossed-out eye, the same eye that
  marks the chat in the sidebar, and a click turns private off. A private chat
  isn't saved to memory and can't change or erase what is already there. To
  keep a new chat's first message out too, open the **Remember across chats: on** chip next to
  the message box and choose **Private chat** before you send it.

## Several chats at once

**One chat at a time on the model.** A second chat sent while another is
working waits its turn and says so ("Waiting for the GPU — 1st in line", also
on its row in the sidebar), then runs when the first finishes. **Stop** takes a
waiting chat out of the line. A chat waiting on you (a permission prompt or a
question) steps out of the line until you answer, then goes ahead of chats that
arrived meanwhile. Chats on a cloud model don't wait for the line at all.

## Long chats

### What GAIA keeps in view

In-process chat agents retain file reads, tool calls and tool results for later
questions, including after the server restarts. Complete older turns leave the
model's context only when its estimated history budget fills; the transcript and
stored evidence remain on your machine. A single turn too big for that budget
stays in view shortened: long tool results become a one-line note naming the
tool, and its longest messages are cut with a note saying how much. When more older turns leave the model's
view — a long chat, or a switch to a smaller model — the chat shows one quiet
line before the reply — "Older messages are now out of the model's view (N
turns)" — and the TUI shows the same line in its transcript. Ask GAIA to recall
anything from before it. The model is told too: a short note where its view now
starts says earlier turns exist and are no longer shown, so it says it can no
longer see something from them instead of saying you never mentioned it. A GAIA agent relayed from the Agent UI gets the text of
earlier turns under the same budget. A single message too long for that budget
reaches the model cut short, with a note saying how many characters were cut,
and the chat says so the same way. Earlier conversations saved before this
feature retain text only.
If the current prompt leaves no history budget, GAIA warns and sends no older
turns; saved evidence remains intact. Switching to a model without native tool
calling replays recorded calls and results as text.

### See how full the context is

The meter beside the model chip shows how much of the model's context window the
next message carries. Click it for the breakdown: system prompt, tools,
documents, conversation, and the room kept free for the reply. As the window
fills, a one-line note above the message box says the oldest turns will soon
drop out of the model's view; **Details** names them by their question, and you
can dismiss the note until another turn reaches the edge. They stay saved, and
you can ask GAIA to recall them.

For more, choose **Details: context & performance** in the meter's breakdown.
It opens the **Context & performance** section of the [session panel](https://amd-gaia.ai/docs/guides/agent-ui/documents.md#the-session-panel),
which shows what fills the window (system prompt, tool definitions,
skills, memory, documents, messages and free space), counted with the
model's own tokenizer when it is loaded and marked **approx.** when not. Tools
that load only when a turn needs them are listed apart. Below that, the panel
shows what is live in the window and, once GAIA has made room, how many older
tool results it moved out of the prompt. They stay readable, and GAIA reads them
again when needed. When the window fills up, the oldest turns may drop out of
the model's view; they stay saved, and you can ask GAIA to recall them.

**Under the hood** shows the last turn or the whole chat: where the model
runs, time to the first word, how fast it read the prompt and wrote the
answer, tokens in and out, how many were reused from the prompt cache, time
in the model against time in GAIA (tools, waiting on you, GAIA's own work),
tool calls, model loads, model memory and cost. Hover a number for the exact
value. Anything that was not measured says "unknown". The panel reads nothing
until you open it.

## Long responses

A chat reply has no time limit while the agent is making progress: model output, a tool starting or finishing, any step. A running tool, or a question or permission prompt waiting on you, also counts, since each has its own time limit. A model that stops sending is handled by GAIA's model-stall check, which ends and retries the request.

GAIA stops a reply only after 75 minutes with no progress at all, and the reply says so: "This turn ran for 75 minutes without progress, so GAIA stopped it." Anything finished before then is kept; send "continue" to pick up where it left off. Change the window with `GAIA_STREAM_TIMEOUT`, set before GAIA starts:

```bash theme={null}
export GAIA_STREAM_TIMEOUT=7200   # seconds without progress; whole number from 1 to 86400
gaia chat --ui
```

An invalid value (not a whole number, or out of range) stops the Agent UI server at startup with an error naming the variable, rather than quietly using the default.

**If the PC sleeps mid-reply** (S3 sleep or hibernation), that time doesn't count toward this limit, the model-stall check, a running tool's time limit or the retry budget, so the reply isn't cut off the moment the PC wakes. A local model and a running tool pick up where they left off. A hosted model's connection may have dropped during the sleep; that is handled like any other dropped connection. A model that stays silent after the PC wakes is still caught. Modern Standby, the sleep most current Windows laptops use, counts as running time.

## When GAIA closes mid-answer

A reply keeps going when you switch chats or close the window, and stops when you quit GAIA. See [Close, quit and restart](https://amd-gaia.ai/docs/guides/agent-ui/desktop.md#close-quit-and-restart).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.