Chat

The platform includes a built-in chat interface, available to everyone in your organization. It connects to the Model Gateway and uses the platform's own user and tenant management, so people can chat with the models they're allowed to use — no setup, no separate login.

Using chat

• Pick a model or a preset agent — the picker lists every chat-capable model you have access to, and every preset agent you may use. Agents come after the models, shown as Preset agent: <name>. When you pick a preset agent, the Tools and Knowledge bases buttons are disabled because the agent has its own tools and knowledge bases — you cannot change them in chat.
• Upload files, images, and audio. Ask questions about a document, summarize it, translate text, or transcribe a recording before sending it to the selected model or preset agent. See Sending audio for the transcription flow and supported formats. You can also paste an image straight into the message box. Copy a screenshot or a picture, click the message box, and press Ctrl+V (⌘+V on a Mac). The image is added to your message like any other attachment, and you can remove it before you send. Only images can be pasted this way. Use the upload button for other files. Images work only with models that accept them. See Sending images.
• Manage conversations — full history, with forking, message editing, and retry / regenerate.

Assistant instructions

Chat adds fixed formatting guidance and the current UTC date to each request, including follow-ups and retries. For direct models, it also uses the tenant's assistant instructions. The default is a built-in helpful-assistant prompt. Preset agents use their configured instructions together with Chat's formatting guidance and the date; they do not inherit the tenant assistant prompt.

Open Chat Settings in the Portal sidebar, then select Assistant Instructions. The sidebar entry and page are available only to administrators of the selected tenant.

Above the editable text box, a status reads Using the platform default instructions. or Custom instructions. A saved empty value counts as custom instructions.

Save an empty value to omit the tenant prompt, or choose Reset to Default to restore the built-in prompt. Changes apply to subsequent requests, including follow-ups in existing conversations. They do not change preset-agent instructions. Switching tabs with unsaved changes prompts you to save, discard, or keep editing.

If a read fails, the page shows Assistant instructions could not be loaded. and a Try again button. Select Try again to reload the instructions. An initial read failure leaves the editor unavailable; a later failed reload keeps the previously loaded text or current draft visible. A failed read does not change the saved instructions. If the gateway also cannot load the tenant instructions, direct-model Chat requests fail before model execution.

The Platform Additions tab shows the fixed formatting guidance in a read-only text box. You can select and copy the text, but cannot edit it. This guidance is maintained by the platform and applies to both direct models and preset agents in Chat. Chat renders Markdown, supports \( ... \) for inline math and \[ ... \] for display math, and uses language tags on fenced code blocks for syntax highlighting. Display math delimiters should be on their own lines. Single-dollar math is disabled so currency stays readable.

The gateway also supplies execution context, which is not displayed in Chat Settings:

• The current UTC date is captured once per Chat request and stays the same across its tool turns, even if they cross UTC midnight. Later requests receive the current date.
• The currently callable tool names are supplied on each model dispatch, for both Chat and API requests. They reflect the request, agent configuration, and existing permissions; they do not grant access to tools.

The date and tool context cannot be edited or disabled in Chat Settings.

The former deployment-level CHAT_SYSTEM_PROMPT setting is no longer used. Ordinary API requests receive these Chat-specific additions only if the caller opts into Chat mode; callable-tool context remains mandatory.

Sending images

You can attach an image only when the selected model accepts images. Not every model does.

The same rule applies to a preset agent. It accepts images when the model it runs on accepts them.

• Add an image with the upload button, by dragging it onto the message box, or by pasting it. All three work the same way.
• When the model does not accept images, Chat does not add the image. It shows The selected model does not accept images. Documents and audio files still work. above the message box. When Chat has no transcription model it can use for you, the message reads The selected model does not accept images. Documents still work. See Transcription model access.
• Point at the upload button, or move to it with the keyboard, to read the same message before you try.
• Documents and audio work with every model. Chat turns them into text before it sends your message, so only images depend on the model you selected.
• If you select a different model while images are still attached, Chat does not send the message. It shows The selected model does not accept images. Remove the attached images or select a model that accepts them. above the message box.
• The same happens when you edit a message that has images, and when you retry an answer in a conversation you moved to a model that does not accept images.
• Nothing is lost. Remove the images, or select a model that accepts images, and send again.

An administrator decides which models accept images. See How a model's capabilities are set.

Sending audio

Attach a recording with the upload button, as you would attach a document. Chat sends the recording to the transcription model and waits for the text before it can send your message.

• Attach one audio file per message. If you attach more, Chat shows Attach only one audio file per message. and does not add the new audio files.
• The message box stays locked while Chat transcribes the recording. It unlocks when the transcript is ready.
• The message sent to the selected model or preset agent contains text only. If you also typed text, Chat sends it first, followed by an Audio transcript: section and the transcript. Without typed text, Chat sends only the transcript.
• The selected model or preset agent never receives the audio file. Chat does not store the recording in the conversation.

Supported formats

Chat supports WAV, MP3, FLAC, and Ogg Vorbis recordings up to 127 MiB. A larger file produces Audio files must be 127 MiB or smaller. Save a shorter recording, or split it into files within the limit and attach one file per message.

For M4A, AAC, WMA, and Opus recordings, behavior depends on the file metadata. If the browser identifies the file as audio, Chat sends it to the transcription model, which decides whether it can decode the recording. If the browser does not identify the file as audio, Chat rejects these filename extensions before sending the file to the model. In either case, an unsupported recording produces This audio format is not supported. Use WAV, MP3, FLAC, or Ogg Vorbis. Save the recording in one of the supported formats and attach it again.

Transcription model access

Chat selects the transcription model automatically. When an administrator has set the platform default transcription model, Chat uses it; otherwise exactly one model must be available to you. Administrators set the default on the platform Models page with Configure defaults, under GPU & model management.

• No transcription model is available for this workspace. You cannot use audio yet. Ask an administrator to give you access to a transcription model.
• Chat cannot choose between transcription models. Ask an administrator to set the default transcription model, or to grant you access to it. You have access to more than one model and no configured default applies to you: either no default is set, or the default is no longer available to you.
• Chat could not load transcription models. Remove the audio file and try again. Chat could not read the model list. Remove the failed attachment and attach the file again.

When a recording fails

Chat shows the reason above the message box and marks the audio file with a warning. The message box stays locked, so you cannot send the message until you address the failed attachment.

• Select the remove button next to the file. The message box unlocks.
• Attach an audio file again to retry. The new attempt replaces the failed attachment.

Removing a recording while Chat is still processing it stops the transcription request.

Running several chats at the same time

You do not have to wait for one answer to finish. Send your message, open another conversation from the list, and send a message there too. Both answers keep arriving.

• The message box is locked only in the conversation that is still writing an answer. In every other conversation you can type and send as usual.
• The conversation list shows a small dot next to every conversation that is still writing an answer. The dot disappears when the answer is finished. A screen reader reads the conversation name followed by Streaming….
• Stop ends only the answer in the conversation you are looking at. Answers in other conversations keep arriving.
• Open a conversation again at any time. It appears right away, and an answer that is still arriving carries on where it left off.
• Deleting a conversation stops the answer it is still writing.
• Closing the tab or losing your connection stops the stream. When you open the conversation again, you see the answer up to the last part the server saved. The answer does not continue from there. For agents with output guardrails, an interrupted request leaves no partial answer if the checks have not finished.

Chatting with a preset agent

A preset agent is a saved setup with its own model, its own instructions, and its own tools. Pick one in the picker and write your message as usual.

• The person who created the agent chooses its model, its instructions, and its tools. You cannot change them in chat.
• You do not pick tools for an agent. It always uses the tools it was set up with. The Tools and Knowledge bases buttons stay on screen while an agent is selected, but you cannot use them. Point at a button, or move to it with the keyboard, to see why.
• You can send an image to a preset agent when the agent's model accepts images. The person who created the agent chose that model. When the model does not accept images, Chat refuses the image and shows the reason above the message box, just as it does for a model you select yourself. Point at the upload button, or move to it with the keyboard, to read the reason first. Documents and audio always work, because Chat turns them into text before it sends your message.
• You do not need access to the agent's model or to its tools. Access to the agent is enough.
• An agent can check its final answer before showing it. With output guardrails on, you wait for the agent to finish generating its answer and for the checks to finish. During that wait, no answer text appears gradually, and you receive no reasoning or tool progress updates. The final answer or, if a check blocks it, a short block message then appears all at once. This can make the wait before seeing an answer longer. See Output guardrails.
• The conversation stays with the agent you started it on. Its entry in the conversation list reads <title> · Preset agent: <agent id>, and the picker returns to that agent when you open the conversation again — also after a page reload.
• You can still move the conversation to a model or to another agent. When the conversation already has messages, chat asks you to confirm first, because the change can affect the quality of the answers.

If the agent is deleted or renamed, opening the conversation shows Preset agent no longer available. Pick a model or another agent in the dialog and select Continue. The conversation carries on with your choice, and chat does not ask again.

Some of an agent's tools belong to an outside service that you sign in to yourself. When the agent needs one of them, Chat shows a dialog titled Connect service to continue. It names the agent and service, and shows the address where sign-in will open. Select Connect service. A new browser tab opens on that service's sign-in page. Sign in there. The tab closes and the agent finishes answering the message you already sent. You do not have to send it again. You sign in through the agent, so you do not need your own permission on the tool server. If the agent needs more than one service, Chat asks for each one in turn.

Select Cancel or press Esc to stop the connection attempt. The answer shows Connection cancelled. Connect the required service to continue. Send the message again when you are ready to connect. Stop also closes the sign-in tab and ends the request.

Chat runs entirely within your environment and respects the same tenant boundaries and access rules as the rest of the platform.