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

# The chat surface

> The composer, answer modes, voice input, answer actions, and how citations are presented

Everything on this page is the Libra chat UI as it renders **inside the embedded widget**. Where the widget deliberately differs from the Libra Workspace, it says so.

## The composer

<Frame caption="The composer: Tools, Research, the model-tier control, voice, and send">
  <img src="https://mintcdn.com/libra-sdk/urMcTJk7LWyFarvE/assets/images/sdk/suggestion-chip.png?fit=max&auto=format&n=urMcTJk7LWyFarvE&q=85&s=0f96d4b176a1d0b183fcc5a76e56c27e" alt="Libra widget composer showing a document suggestion chip, the Tools and Research buttons, the Claude Opus 5 tier control, microphone and send" width="360" data-path="assets/images/sdk/suggestion-chip.png" />
</Frame>

| Control         | What it does                                                                                                                                                                                           |
| --------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **Tools** (`+`) | Attach, Assistants, Workflows, Current Project, Add context. See [Context and sources](/capabilities/context-and-sources#the-tools-menu).                                                              |
| **Research**    | Turns research on and picks sources; shows the logos of the active ones.                                                                                                                               |
| **Model tier**  | Standard, Deep Thinking or Fast — see below.                                                                                                                                                           |
| **Improve**     | Rewrites the user's draft prompt into a better one **without sending it**, so they can review it first. An **Undo** button replaces it afterwards, and disappears as soon as the text is edited again. |
| **Microphone**  | Voice input — see below.                                                                                                                                                                               |
| **Send**        | Becomes a **Stop** button while an answer is generating.                                                                                                                                               |

Inside the widget the composer is pinned to its compact layout, so Tools and Research are icon-only at every panel width. See [Panel and placement](/guides/panel-and-placement#width).

## Answer modes

One control in the composer selects how Libra answers. Its accessible name is *Select model tier*, and it always offers three Libra AI tiers:

<CardGroup cols={3}>
  <Card title="Standard" icon="circle">
    Best for most tasks. The default.
  </Card>

  <Card title="Deep Thinking" icon="brain">
    Complex reasoning and analysis.
  </Card>

  <Card title="Fast" icon="zap">
    Quick, everyday questions.
  </Card>
</CardGroup>

**Deep Thinking** routes the turn to a reasoning model and skips the planning step, so answers take longer and go deeper. The first time a user picks it, a confirmation dialog warns that complex queries take 2–3 minutes and deep research 5 minutes or more.

Two caveats worth knowing:

* The control's **visible label is not always a tier name.** If the user has enabled *Show Model Selector* in their Libra settings (off by default), the same dropdown also lists individual models, and picking one makes the button read that model's name — for example `Claude Opus 5`.
* The control is **hidden** while an assistant or a workflow is active, and on model-locked deployments.

<Note>
  Deep Thinking is **not** restricted by subscription tier. Older material describing it as a paid-tier feature is out of date — see [Plans and limits](/capabilities/plans-and-limits).
</Note>

## While an answer is generating

The microphone and send buttons collapse into a single **Stop** button that ends the response where it is.

Progress appears in a card above the answer. Collapsed and running, it reads *Analyzing…* then *Composing answer…*; when finished it reads *Thought for \<time>*. Expanding it reveals the step list under the heading **Actions**, with steps named after what actually ran — *Analyzed query*, *Searching Wolters Kluwer Online*, *Searched a database*, *Searching the web*, *Documents read*.

## Answer actions

Under every finished answer:

| Action           | Behaviour                                                                                                                                                                                                                                               |
| ---------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Copy**         | Copies immediately when the answer cites nothing. When it does cite sources, opens a menu: *Copy with footnotes* or *Copy plain text*. Either way the clipboard carries both rich and plain text, so pasting into Word or Outlook keeps the formatting. |
| **Regenerate**   | Produces a new version. Available on the **latest answer only**. Once an answer has versions, a counter (`2/3`) with arrows appears and stays visible while the new version streams.                                                                    |
| **Export**       | Opens a menu: Word or PDF (the choice is remembered), an optional *With footnotes* toggle when the answer cites something, then *Export entire chat* or *Export this message*. Downloads as `Libra export <date>.docx` or `.pdf`.                       |
| **Create email** | Opens the user's default mail client with the answer as plain text. It does not attach a file.                                                                                                                                                          |

## Citations

Citations render as **inline chips** — source name plus source logo — at the exact point in the text where the answer cites them, not gathered at the end of the paragraph. Numbered footnotes with a bibliography appear only when the answer is copied or exported.

When two or more chips for the same source type sit next to each other, they collapse into one chip with a **`+N`** suffix, where N is how many *additional* citations are behind it.

<Steps>
  <Step title="Hover a chip — the citation card">
    Shows the source logo and name, the cited content's title, optional metadata such as page or date, and an excerpt of the cited passage. A merged chip adds prev/next arrows and an *N of M citations* counter. The footer carries **Citation list** and **To source**.
  </Step>

  <Step title="Open the cited-sources row — the citation list">
    Every answer that used sources ends with an **`N Cited Sources:`** row. In the widget this opens a modal titled *Citation list* (the Workspace shows it as a side-panel tab). One card per source, one expanded at a time, revealing the cited snippets plus **View source** and **Find in chat**.
  </Step>
</Steps>

<Warning>
  **The widget has no document viewer.** Citations into *user-uploaded* documents therefore show no **To source** or **View source** button, and generated Word documents offer download only, with no in-widget preview. Citations into research sources and host-supplied external documents are unaffected — those route through `onCitationClick`. This is a deliberate difference from the Libra web app.
</Warning>

Where those citation clicks go is the integrating product's decision: see [Citation clicks](/guides/citations).

## Voice input

The microphone button starts dictation. The composer is replaced by a recording bar with a live waveform, an elapsed timer, **Cancel recording** and **Confirm recording**. Confirming shows a *Processing* state, then **appends** the transcript to whatever is already in the composer, so the user can edit before sending.

Sending stays manual unless the user turns on *Settings → Audio → Instant Auto-Send After Voice Transcription*, which is off by default. No language is chosen in the UI or sent with the request; transcription relies on the speech model's own language detection.

<Warning>
  If your page sends a `Permissions-Policy` header it must include `microphone=(self)`, or the microphone is blocked inside the iframe. See [Security headers](/guides/security-headers).
</Warning>