Chat

The Chat section is used for conversations with AI agents that are created and configured in the Agents section of the platform. Users can start new conversations, browse the history of previous chats, use prompt suggestions, upload supporting files, and provide feedback on responses.

Chat Input and File Uploads

The chat input supports both written instructions and attached context. Users can type a message, upload files from the attachment button, or drag and drop files directly into the chat area. This is useful when the agent should work with a document, spreadsheet, image, exported report, screenshot, or another file that is not already available through Data, Memory, or a connected system.

When attaching files, describe what the agent should do with them. For example, ask the agent to summarize a PDF, compare two documents, extract action items from meeting notes, read a screenshot, or analyze a spreadsheet. Clear instructions help the agent focus on the right parts of the uploaded content.

If the selected agent has Sandbox enabled, it can stage an attached file in an isolated workspace, transform it, and explicitly publish a finished file back to the conversation. Availability depends on the deployed environment and the agent configuration. See Sandbox file workflow and verify sensitive or business-critical output before using it externally.

Planning and To-dos

Agents with Agent harness settings can expose a reviewable workflow for complex requests:

1. The agent proposes a plan before starting the planned work.
2. Review the scope and approve it, or ask the agent to revise it.
3. After approval, the agent can create conversation-scoped To-do items and update them as work progresses.
4. Open the progress chip or To-dos panel to see completed items, the current item, and remaining work.
5. The final response should summarize the result and identify anything that remains incomplete.

These concepts have different lifecycles:

Item	Purpose	Lifecycle
Plan	Proposes how the agent will approach the current request.	Conversation content reviewed before the planned work begins.
Harness To-do	Tracks an internal step of the agent's work.	Progress state associated with the conversation, not an organizational work record.
Base task	Records durable work for a team or project.	A separate Siesta Base object created only when the user explicitly requests task creation.

Background agents can appear as additional work in the harness panel when the agent configuration and runtime support delegation. Do not assume that every request will be delegated or run in parallel.

Sandbox file workflow

When the selected agent has Sandbox enabled, it can stage a named chat attachment into its isolated workspace, transform it, and explicitly publish a finished file back to the conversation. For example, attach sales.csv and ask the agent to clean it and publish cleaned-sales.csv. A workspace file is not a downloadable artifact until the agent publishes it.

For a report, presentation, or chart, ask the agent to render and visually check the result before publishing. Its viewImage check is separate from the preview you open on the published artifact.

Sandbox status and tool detail show whether a command completed, failed, timed out, or reported that the workspace was reset. An approval control is shown only when the deployed tool configuration requires that call to be confirmed. See the Sandbox tool reference for the workspace and file lifecycle.

Generate Images in Chat

Chat can show a generated image directly in the agent response when the selected agent has Image generation enabled.

1. Start a New Chat and choose an image-enabled agent.
2. Describe the subject and add any important style, layout, color, aspect-ratio, or text requirements.
3. Send the request and wait for the image to appear in the response.
4. Select the image to preview it, or use Download to save the generated PNG.
5. If the result needs work, reply with one clear change instead of repeating the whole request.

Example request:

Generate an image of a friendly orange cat sitting beside a window in warm morning light. Use a square composition and do not add text or logos.

The screenshot below uses xCode and the shorter request Generate an image of a cat.

If the agent says that it cannot create an image, confirm that you selected the intended agent and start a new chat after its configuration was saved. An agent editor can follow Enable Image Generation. Organization safety rules can still block a request even when the tool is enabled, so simplify or revise a blocked prompt instead of trying to bypass the restriction.

Conversation files

Use Conversation files in the chat header to find files without scrolling back through messages:

1. Open a saved conversation that contains an attachment or published output.
2. Select Conversation files beside the conversation title.
3. Review Outputs for generated/published artifacts and Sources for user-uploaded attachments. Repeated references to the same file are grouped; separately published outputs remain separate files.
4. Select a file to open its available preview, or use its download action. An unsupported preview can fall back to downloading the file.

Sources appears only when input attachments are present. Outputs can show No created files yet. If the conversation has no files, the header control is hidden. A newly uploaded item can remain unavailable while its message is being saved; wait for saving to finish and reopen the saved conversation if necessary. For an unavailable file, check current conversation access and whether the source still exists.

This list does not turn a file into an artifact or attach it to a workspace. A file still only on the sandbox disk is not listed as a published output. Documents retrieved from Data are separate document reference badges, not user attachment Sources. Opening a listed file still requires the applicable access.

Documents Used in an Answer

When an agent searches an assigned Data collection, the internal chat can show the retrieved source documents below the assistant answer. These document references help you verify where retrieved context came from; they are not a claim that every sentence in the answer is supported by every displayed document.

Each reference badge shows a file-type icon when the extension is recognized, otherwise the data-source image when one is available, and finally a generic file icon as a fallback. The visible label is the file name; hover over a truncated label to see the stored document path.

Up to five references appear directly. If an answer has more than five, the first four remain visible and the fifth position becomes More documents +N. Open that badge to page through the remaining references, see the current position, and choose Open document.

Opening a reference opens a panel to preview or download the stored original. Use Open document details to open its full Data detail in a new tab. A reference contains the document ID, document name, data-source ID, collection ID, and optional source image; it does not grant new access. Current Data permissions still apply. A removed or inaccessible document can show an access or not-found state. If the original was not stored, the panel explains that the original file is unavailable; indexed text in Data may still be available.

The search action detail also lists every referenced document under the tool result with its icon, file name, full-path tooltip, and actions to inspect it. Use this list when the answer badges are collapsed or when reviewing exactly what one search action returned.

References are carried through all supported conversation paths:

• streaming chat accepts document-reference metadata while the search action completes;
• realtime tool events carry the same metadata into the conversation;
• saved messages persist the references so they can be reconstructed when the conversation is reopened.

Document links are enabled only in authenticated internal chat and internal conversation review. Shared or public conversation views hide these internal Data links; sharing a conversation must not expose a collection document to a viewer who lacks Data access.

Do not confuse the four file surfaces:

Surface	Meaning
Document reference badge	A structured pointer to a Data collection document used by a search action.
Markdown link	Ordinary answer text controlled by the rendered content; it has no Data-reference metadata by itself.
User attachment	Input uploaded with a user message for that conversation.
Artifact	A generated or published output attached to an assistant/tool message for opening or download.

For retrieval and processing diagnostics, continue with Processing, Sync, and Troubleshooting. For administrative review, see Conversations.

Starting a New Conversation

Open Work → Chat, then select the New Chat icon beside Conversations in the sidebar. An opened conversation also offers New Chat in its header.

At the start of a conversation, the chat shows the selected agent, the message composer, available input actions, and any prompt suggestions prepared for the user. Prompt suggestions work as reusable starting points: the user can click one instead of writing the first instruction from scratch, then adjust the text before sending it.

Choose the agent in the composer before attaching files and sending the first message. Use + to attach files, or drag them into the chat. Microphone dictation is available when the browser supports speech recognition.

When realtime voice is available for the selected agent, users can start from an already open conversation or directly from the empty-state composer before the first text message is sent.

After the first message is sent, the conversation appears in the list on the left. Open its … menu to rename it.

Chat Interface

The main screen is divided into two parts:

• On the left, the sidebar shows navigation and the accessible conversation list, with titles, status indicators, an agent filter, and New Chat.
• On the right, the selected conversation shows the messages and the composer used to continue the chat.

Type a request in the composer and select the send arrow or press Enter. Use Shift+Enter for a new line. You can also attach files with +, drag files into the conversation, or use microphone dictation when supported.

The agent responds in real-time, with each message being stored within the given conversation.

Public chat, widget embedding, authenticated widget mode, privacy links, and public realtime access are configured on the agent itself. See Agents > Interfaces for those settings.

Navigate a Long Conversation

Completing a streamed response preserves the paragraph you are reading. If you are following the bottom of the conversation, it stays at the end as content grows. No new setting is required.

When a conversation contains at least five user turns, internal chat shows a compact transcript outline beside the message stream. The trigger contains up to six markers sampled across the conversation, so a long thread stays compact instead of adding one permanent marker for every turn. The active marker follows the user turn nearest the current scroll position.

Open the outline to see every user turn. Each item uses the trimmed message text as its label. If a turn has no text, the outline uses its attachment file names; if neither is available, it uses an untitled-turn fallback. Assistant responses are not separate outline items.

Select an item to scroll its user message smoothly to the start of the conversation viewport. The selected location is highlighted with aria-current. The trigger and list items are buttons, so keyboard users can reach them with Tab and activate them with Enter or Space. Navigation scrolls and highlights the destination; it does not move keyboard focus into the message content.

The outline is derived from the current message list. It appears when a streamed conversation reaches its fifth user turn and updates as later turns arrive.

Conversation surface	Transcript outline
Authenticated internal chat	Available from five user turns
Full shared-conversation page	Available from five user turns
Share-dialog preview	Hidden
Public chat widget	Hidden
Authenticated embedded widget	Hidden

Filtering and Managing Conversations

• Select a conversation title in the sidebar to reopen it. The selected conversation is highlighted; scroll down to load more conversations automatically. Nested workspace lists use Load more when more items are available.
• Open … beside a conversation to Rename it or, when permitted, Delete it with confirmation.
• Open the agent filter beside Conversations to search for and select an agent. To show conversations for all agents again, open the filter and choose Reset filter. The selection applies immediately.

Submitting a Query

• Type your request and add any files with + or drag and drop. Wait for attachments to finish uploading before sending.
• Send with the arrow or Enter; Shift+Enter inserts a new line.
• The microphone dictates text into the draft when supported by the browser. Realtime voice is a separate control available for supported agents.
• During a response, the send control becomes Stop. Sending can also be unavailable while a tool action awaits approval or a realtime session is active.

Realtime and Voice Conversations

Siesta AI supports realtime conversation flows for interactive voice-style experiences. A realtime session is always connected to a specific conversation, so messages, transcripts, tool activity, feedback, and audit context remain tied to the same conversation history as standard chat.

Realtime can be used in two contexts:

• Internal chat: authenticated users start a realtime session from an existing or newly created conversation.
• Public chat: public or embedded chat can use realtime only when public chat and realtime public access are enabled for the organization and agent.

The realtime flow has two steps:

1. The client creates a realtime-session for the conversation.
2. The client connects to the returned WebSocket path and streams audio/events for that session.

In the UI, realtime controls can expose:

• a voice start action directly in the composer,
• a stop action that lets the user end the active realtime turn without leaving the conversation,
• persona or voice selection controls,
• audio activity indicators,
• approval waiting state when a tool call must be confirmed before the session can continue.

Session behavior is controlled by platform limits. The current backend defaults include a short session token lifetime, an idle timeout of 60 seconds, a maximum session duration of 1,800 seconds, up to 2 concurrent sessions per user, and up to 100 concurrent sessions per organization. The supported input and output audio format is pcm16, and the default supported voice is alloy.

If the session expires, becomes idle, exceeds the maximum duration, or violates origin/security checks, the client should create a new session before reconnecting. Realtime sessions remain tied to the same conversation, so transcripts, tool events, and persisted messages stay in the conversation history instead of being stored separately.

For public chat and widgets, administrators should review public chat settings, allowed origins, content safety, prompt shield settings, and privacy disclosures before enabling realtime for customer-facing use.

Feedback on Responses

Under each agent response, it is possible to copy the response or click on the thumbs up or down icon, allowing the user to provide quick feedback on the agent's response.

After submitting a rating, a Send feedback window will also appear, where a specific comment on the agent's response can be added. This feedback is automatically sent to the admin interface upon submission.

Review submitted ratings and comments in Admin → Conversations → Feedbacks, or in the selected agent's Feedbacks. These views follow their own permissions and agent write access. Feedback does not automatically update the agent's instructions.

Sharing a Conversation

To share a saved conversation, select Share in its header. Review the conversation preview, choose an available Access option, such as Public or Organization only, and select Create share link when ready. The available options depend on your organization and permissions.

If the shared conversation contains generated artifacts or realtime messages, verify the preview before sending the link so the recipient sees the intended final state.

Preview and Download Generated Outputs

An agent can return several kinds of output in a chat. They are not interchangeable:

Output	What you can do
Generated image	Open the image in a larger preview or download it as a PNG. See Generate Images in Chat.
Artifact	Open or download a file that an agent or tool explicitly published to the conversation.
Generated code	Download the generated file when the response provides one, then review it before running, sharing, or adding it to a project.
HTML preview	View supported generated HTML in an isolated preview directly in chat.

The HTML preview is isolated from the Siesta AI application. Output that is unsupported or should not be rendered directly stays in a code-style fallback instead. A preview helps you inspect a result; it does not make the content trusted or safe for external use.

Before sharing a conversation, publishing an artifact, or using generated code in another system, inspect the final output and confirm that it contains no unintended data, instructions, or actions. Downloaded code remains your responsibility to review and test in an appropriate environment.

To create images, the selected agent must have Image generation enabled. See Generate Images in Chat for the user workflow and Enable Image Generation for agent configuration.

Find published files from accessible conversations and workspaces in the global Artifacts catalog. Open a file's source to return to its context. Adding a chat attachment does not automatically publish it to a workspace; see workspace files.

Start from Work → Workspaces to create a conversation with an explicit workspace scope. Its published output artifacts can be attached to workspace Artifacts automatically, subject to edit permission and write access. An ordinary agent chat does not acquire workspace scope merely because the agent is assigned there. See Workspaces.

File preview formats

Select Open on a supported published file to inspect it in Chat. Use Download original to save the actual file. Preview is read-only and does not change the original.

File	Preview behavior
Raster images	Preview PNG, JPEG, WebP, GIF, and AVIF within the source-size limit. Download the original if decoding fails or the file exceeds the preview limit.
PDF	Pages with navigation, zoom, and fit controls. Password-protected PDFs require downloading and opening in a PDF application.
Word .docx	Paginated document preview for a broad range of compatible Office files. Complex layout or formatting can be simplified; check the original before sharing.
PowerPoint .pptx	Slide navigation and zoom. Some content, including charts, may be incomplete.
Excel .xlsx	Worksheet data and supported charts across a broad range of compatible workbooks. Complex charts, formatting, or very wide sheets can be simplified or partially previewed. Formulas are not recalculated; a missing saved result is shown as unavailable.
CSV	Read-only table with delimiter and header controls. Values are treated as text.
Text and code	Readable source with text wrapping; supported languages can have syntax highlighting.
HTML and Markdown files	Switch between the rendered Preview and Source. Unsupported rendered content falls back to source.
Other formats	Download the original when no preview is available. Legacy .doc, .xls, .ppt and macro-enabled Office files do not use the Office previews above.

SVG and MDX files are shown as source rather than rendered documents. Generated images can also be opened directly from their message. The shared file viewer supports the raster formats listed above.

Preview has file-size, page/slide-count, and content-complexity limits. A file can be valid for download while exceeding its preview limit. Preview simplification or an error does not change the original file: use Download original to inspect the authoritative document in its native application. If loading fails, use Retry or Download original; if access has changed, sign in again or ask the source owner to check permissions.

Raster and source preview limits

The shared file viewer validates the filename, reported content type, and raster file signature before decoding an image. Renaming an unsupported file to a supported extension does not make it previewable.

The current shared-viewer image budget is 20 MiB of source bytes, a maximum of 8192 pixels for either dimension, and 40 million pixels in total. An image can fit the byte budget but exceed the dimension/pixel budget. Invalid, inconsistent, unsupported, or oversized content uses the unavailable/error state and the original-download fallback. Preview limits are separate from Data storage quotas, chat upload limits, and model context capacity.

SVG is displayed as source text in this shared viewer. Source previews can highlight supported code languages and show line numbers; large or complex sources stay as plain text when highlighting limits are exceeded. If highlighting cannot load, the source remains readable. Highlighting does not execute code or modify the file.

Chat images can also open in a separate image preview directly from a message or the conversation file list. Do not assume that this separate image experience uses every rule or control of the shared file viewer.

Context Usage and Capacity

Chat can show how much of the current conversation context has been used and the capacity limit that applies to it. This helps you recognize when a long conversation, large attachments, or broad instructions leave less room for the next request.

The agent can summarize older context to continue a long conversation and retry a request after context compaction when the model rejects an oversized input. A summary may omit details, so retain important source material in its original file, Data, or Memory.

If the conversation approaches its capacity, start a new chat with a focused summary of the needed context, narrow the request, or attach only the files relevant to the next task. Context capacity is separate from organization and connection token limits; administrators edit token policies through Admin → Limits → Model limits and inspect utilization in Analytics → Limits.