New Chat is the main DEEIX Chat composer. Users choose a model, type a message, upload files, and enable model options, MCP tools, HTML visual layout, or voice input when the task needs them.
Where It Is#
Click “New Chat” in the sidebar to open a clean conversation. The composer at the bottom is more than a text box: it submits the current message together with the selected model, attachments, model options, tools, and visual layout setting.
When text is present, the lower-right button sends the message. During generation, it becomes a pause button so users can stop the response. While files or requests are processing, the composer is temporarily disabled to avoid duplicate submission.
The send shortcut follows the user conversation setting. IME composition does not trigger accidental sending.
Choose a Model#
The model picker sits on the lower right of the composer. It shows the selected platform model name and icon. When opened, available models are grouped by vendor. Names, icons, pricing hints, and capabilities come from admin configuration.
If a model is missing, it is usually not enabled, not allowed for the account, limited by balance or policy, or not exposed by an administrator.
| Task | Choose |
|---|---|
| General Q&A, writing, summarization | A chat-capable text model. |
| Screenshot or image analysis | A vision-capable chat model, then upload the image. |
| Image generation | An image generation model; with no attachments, the turn enters image generation mode. |
| Image editing | An image-editing model, with the image attached. |
| MCP tool use | A normal chat model; MCP tools are hidden in image generation or image editing mode. |
Some models show a pricing icon. Hover it to inspect token, per-call, or duration pricing. Long context, images, tool calls, and retries can increase usage.
Upload Files#
The plus button on the lower left opens the tools menu, currently including file upload and screenshot capture. Users can also drag files into the composer or paste images and files from the clipboard.
Uploaded files appear as attachment cards with file name, size, processing status, and a remove button. Click a card to preview the file. Wait for upload or processing to finish before sending when the file matters.
| File state | What happens |
|---|---|
| Image + image-editing model | The turn enters image edit mode. |
| Image + chat model | The image becomes conversation context for recognition, analysis, or extraction. |
| Documents or other non-image files | They are used as source context according to platform file policy. |
| Image generation model + attachments | Image generation rejects attachments; remove them first. |
| Non-image files mixed into an image task | The composer shows an invalid-file warning; keep only images. |
If RAG file mode is enabled and all document attachments have RAG disabled, the composer shows a warning. Confirm whether the files should participate in retrieval before sending.
Configure Parameters#
Model Options on the left side of the composer adjust parameters for the current model request. The button appears only when admin policy and the selected model allow user-controlled options.
Options are not fixed. They can vary by model and protocol. Users may adjust output randomness, context-related controls, provider-native tool switches, or restore defaults. Most chats should use defaults unless output style, stability, length, or model behavior needs control.
| Scenario | Recommendation |
|---|---|
| Need stable, repeatable answers | Lower randomness-related controls or restore defaults and retry. |
| Need more varied expression | Increase creative controls carefully and expect more variation. |
| Unsure what a setting means | Keep defaults and avoid changing multiple settings at once. |
| Output behaves oddly | Restore defaults first, then check the prompt or model choice. |
Model options affect the current request. They are not admin-side model, routing, or pricing configuration.
MCP Tools#
The MCP Tools entry appears on the left side of the composer when tools are available for the account and the current task is not image generation or image editing.
Tools are grouped by MCP Server. Users can search tools, expand a server, select individual tools, or select a group. The button badge shows the number of selected tools.
The maximum selected tool count is controlled by platform limits. If the limit is exceeded, the UI asks users to reduce selection. Select only the tools needed for the current turn.
| Scenario | Recommendation |
|---|---|
| Query internal systems, knowledge bases, or business data | Select only the required tools under the relevant MCP Server. |
| Unsure what a tool does | Check its name, description, and server before enabling it. |
| Sensitive data or write-capable tools | State the boundary clearly, for example “query only, do not modify.” |
| Tool results are critical | Ask the model to distinguish tool-returned facts from inference. |
MCP tool selection applies to the current message. Recheck tools when switching conversations or tasks.
HTML Visual Layout#
The blocks icon on the left is the visual layout prompt toggle. It appears only in normal chat mode and is hidden for image generation or image editing.
When enabled, the current turn can inject safe inline HTML layout guidance where appropriate, helping the model render complex structures more clearly. The setting is stored locally in the browser and reused later.
Use it for page analysis, product plans, operations reviews, requirement breakdowns, cards, grouped sections, comparisons, flows, and dashboard-like answers.
Leave it off for simple Q&A, short rewrites, code snippets, plain tables, or tasks that only need a direct answer.
Voice Input#
When the text box is empty, the lower-right button shows the voice input icon. Clicking it asks the browser for microphone permission and continuously writes recognized speech into the composer. Click again to stop.
Voice input depends on browser SpeechRecognition or webkitSpeechRecognition. It follows the browser language setting. Unsupported browsers disable the button and show an unsupported hint.
Voice input works well for quick thoughts, meeting questions, or longer natural-language descriptions. Review the recognized text before sending, especially model names, file names, numbers, proper nouns, and mixed Chinese/English content.
Before Sending#
Before sending, check that the model matches the task, files have finished processing, parameters are expected, only necessary MCP tools are selected, HTML visual layout is useful, and voice-recognized text is accurate.
For customer data, account details, secrets, contracts, payments, or personal information, confirm the material is allowed on the platform before typing or attaching it.