Strata PlatformDocs
↑↓ navigate ↵ open esc close

The chat interface

The chat page is where you do most of your work in Strata — asking questions, generating reports, and reviewing what the AI did to answer you.

Layout at a glance

Open Chat from the left sidebar. The page has three main areas:

  • Conversation area — the running transcript of your messages and Strata's replies.
  • Composer — the input box at the bottom. A plus button sits on the left; the model picker, a dictate (microphone) button, a voice-mode button, and the send button sit on the right.
  • App sidebar — the collapsible left rail shared across Strata. It lists your previous conversations (with a Show all chats link) under Search, New chat, Inbox, Agents and Customize.

If a SQL Server extension is connected, you may also see the artifact panel slide in on the right side when the AI opens a SQL workspace or builds a dashboard. See /docs/chat/artifacts.

Every chat also has a workspace panel, opened from the Workspace toggle in the top-right corner. It holds the chat's sub-agents, its checklist, the files the chat produced, and the places Strata looked for information. See /docs/chat/workspace-panel.

Sending a message

  1. Type your question in the input box at the bottom.
  2. Press Enter to send, or Shift+Enter for a new line.
  3. The send button (upward arrow) on the right of the composer activates once you have text. Click it to submit if you prefer the mouse.

While Strata is generating, the send button becomes a Stop button (a filled square). Click it to interrupt the current response.

Unsent messages

What you have typed but not sent is kept on this device, for each chat: open the chat again in a new tab, after a reload, or after the browser closed unexpectedly, and the text is back in the message box. Only the text is kept, not attached files, for 14 days after you last changed it and for up to 30 chats. It never leaves your browser, is not kept while you are in incognito, and is deleted when you sign out.

Mentioning with @

Type @ at the start of the message or after a space to point that one message at something. The menu opens just above the line you are typing, its edge at the @, over the top of the message box if it has to. It lists, in groups:

  • Agents — your agents and the ones your organization published. Picking one writes @AgentName, and that agent answers this one message.
  • Extensions — the extensions you have connected and your organization allows you: Outlook, OneDrive, SQL Server, Jira and Confluence. Picking one writes @Outlook (for example) with Outlook's logo just before it: in the message box the mention is blue, in the message once sent it is bold, on no background either way, and the AI looks there first for this message. SQL Server is always your active connection; a mention never switches it.
  • Connectors — the connectors you have turned on (and signed in to, when a connector uses your own account). A connector's mention carries its logo too, or its initial when Strata has no logo for it.
  • Files — your OneDrive files, once OneDrive is connected: your recent files for a bare @, and a search once you have typed two letters. Picking one adds it to the message as a chip, and the AI reads it from OneDrive before answering. It counts toward the 20-file limit.

Keep typing to narrow the menu (names with accents or several words match too), use the arrow keys to move, and press Enter or Tab to choose. A mention deletes as one piece: Backspace right after it removes the whole name.

A mention only points the AI at something this chat can already use. It never turns on an extension or connector your organization has switched off for you; if one is no longer available when you send, the AI is simply not pointed at it.

Reading a streamed response

Strata streams text into the conversation as the model produces it. A few things may appear above or inside the response:

  • Reasoning trail. When the model takes a real beat to reason, a Thinking line appears with a running timer and the newest one-line summary of what it is working through. When it has finished, it collapses to its last summary line; click it to open a card titled Thought for N s (N is the time spent thinking) that lists one line per step of its reasoning. Short, simple prompts don't show one. See /docs/chat/models.
  • Action steps. A collapsible trail listing what Strata is doing — running a SQL query, loading a skill, searching the web, running an analysis, running sub-agents in parallel, etc. While a step runs, one line shows the working indicator, then the icon of the tool in use (a globe for a web search, a database for a query, a connector's own logo), then what it is doing and the seconds it has taken; between steps the line shows no icon. Once a step has finished, that line opens with a click to show the steps so far. While Strata writes a document, the step says how far it has got — Writing the report with Section 3 · Revenue by region · about 1,240 words on a quiet line under it (Planning the deck with the slide it is on, Preparing the workbook with the rows so far) — and keeps a receipt such as Wrote 6 sections · 2,310 words when it is done. Collapsed, the trail sums up what happened in one sentence: the first two things it did, then how many steps are left, for example Queried 1 source, handled 2 emails, and 3 more steps. It counts each file once however many review rounds it went through: Created 1 file and reviewed the file. Opened, each step is one line with no icon or timing: what it did in a muted tone, then what it did it to (Searched the web and the query, Queried and the table). A connector's step keeps the connector's logo. Queries or analyses that ran back to back fold into one line, such as Ran 3 queries, that opens to them. Click a step to see the SQL, code, stdout, web results, or sub-agent detail behind it; a connector's step opens to its Request and Response, and a loaded skill does not open. See /docs/chat/subagents for how Strata splits a big request across several parallel workers. The Outlook add-in draws the same trail, with each step's name but not its detail; see Following what Strata does.
  • Query results. A SQL step's detail shows the rows the database returned in a table you can sort by any column, with a search field once there are 10 or more rows, Export CSV, and an Expand button that opens the table full screen.
  • Rich blocks. Charts, comparison cards, progress bars, timelines, correlation matrices, email drafts, Mermaid diagrams, and SQL results render inline as interactive components. See /docs/reference/blocks.
  • Citations. Where the answer draws on retrieved data — query results, files, emails, or the web — small numbered markers appear inline. Hover one to see the source; web answers also carry a Sources button in the action row. See /docs/chat/web-search.
  • Callouts. Highlighted boxes prefixed with a label such as Info, Warning, Success, Note, Tip, Key, Data, or Alert call out important context. See /docs/reference/callouts.
  • Agent draft cards. Ask Strata to set up an agent ("every Monday at 9am, summarize my open Jira issues") and it answers with a draft card showing the agent's name, schedule, tools, where results go and model, and the action trail shows Drafted an agent. Nothing is created until you choose Create agent on the card. See Create an agent from a chat.

Strata checks its work. Before it answers a question it used tools for, Strata re-reads your request and checks the answer against what its tools returned: every part of the request done, or named as not done; every figure exactly as a tool gave it; and every file, chart, dashboard or change it says it made actually made. A mistake it can fix, it fixes before answering, and anything still missing appears in a Warning callout. It does this quietly: you see the corrected answer, not a checklist.

When the response is finished, suggested follow-up questions may be attached to it. They do not render under the answer — they surface in a panel that expands above the prompt input when you focus it while it is empty. Click any suggestion to send it as your next message. Suggestions are optional, so a short factual answer often has none.

The plus menu

The plus button (+, named Add files and more) on the left of the composer opens a menu with:

  • Add file or photo — attach a file or image to your next message. See /docs/chat/files.
  • Add from OneDrive — only shown when OneDrive is connected and allowed for you. Lists your recent OneDrive files; pick one to add it to your next message as a chip. Search OneDrive… puts an @ in the message box so you can find a file by name (see Mentioning with @ above).
  • Mode — switch between Auto, Technical, and Simple. See /docs/chat/modes.
  • Permission — only shown when a SQL Server extension is connected. Sets whether the AI may run reads only (Viewer), reads and writes with approval (Editor), or DDL (Admin).
  • Agents — your agents; pick one to write @AgentName into the message (that agent answers that one message), or Manage agents to open the Agents page.
  • Skills — list of your pinned skills plus a Manage in Directory link. See /docs/chat/skills.
  • Extensions — opens the extensions modal where you connect or disconnect OneDrive, Outlook, SQL Server, and Atlassian (one connection covering Jira and Confluence). (If your organization runs Atlassian through one organization account, the connection belongs to the organization — there is nothing for you to connect or disconnect.) See /docs/integrations/atlassian.
  • Plan mode — a toggle that lets the AI propose a plan before executing complex tasks.
  • Web search — a toggle that lets the AI query the public web. See /docs/chat/web-search.

If your organization or your administrator locks the mode or the SQL permission, that row appears read-only with a padlock and shows "Set by your organization" instead of a picker. When your organization pins web search with its padlock, the Web search row is locked the same way: it shows a padlock and "Set by your organization", its tick shows whether your organization has web search on, and it can't be changed.

Other composer controls

  • Model picker — the model name with a chevron, to the right of the input box. Click it to switch the AI model for this conversation. The list shows the models you may use: those your organization has enabled in its Model Catalog, narrowed by your role and by any limit an administrator has set for you, with older models grouped under Legacy models. A legacy model that is due to retire shows its date there, for example Retires Oct 19. If your access changes while the page is open and a message runs on another model, the action trail says so, for example GPT-5.6 Sol isn’t available to you — answered with Claude Opus 4.8. See /docs/chat/models.
  • Effort and thinking — pinned to the bottom of the model picker, the Effort row sets how hard the model reasons (Low, Medium, High, Extra, Max) and includes a Thinking toggle for visible extended reasoning. The current effort shows on the model pill. Both are saved to your account and apply to every new message. See /docs/chat/models.
  • Dictate (microphone icon) — push-to-talk dictation that converts your speech into text in the input box. Supported in Chrome, Edge, and Safari.
  • Voice mode (waveform icon) — opens an overlay for a hands-free spoken conversation. See /docs/chat/voice.
  • Incognito — a separate button in the empty-state greeting area that starts a one-off conversation that is not saved to your history. See /docs/chat/incognito.
  • Slash menu — type / at the start of the input to open a quick menu of pinned skills and "Attach file or photo".

Attaching files

Four ways to attach:

  1. Click the plus button and choose Add file or photo, or
  2. Drag-and-drop a file directly onto the chat, or
  3. Paste an image from your clipboard, or
  4. With OneDrive connected, pick a OneDrive file with @ or with Add from OneDrive in the plus menu. It is not uploaded: the AI reads it from OneDrive.

Strata accepts .pdf, .csv, .xlsx, .xls, .json, .xml, .txt, .md, .sql, .log, .png, .jpg, .jpeg, .gif, .webp, .docx, .pptx, .doc, and .ppt. See /docs/chat/files.

Each file shows as a chip with its type and size (for example PDF · 240 KB), with a spinner while it uploads. In a saved chat, files you attach and files Strata makes stay usable later in the same conversation, even in a later session: Strata can save them to OneDrive, attach them to an email or edit them without making them again. See Files from earlier in the conversation.

Past chats and pinning

The app sidebar lists your conversations: Pinned first, then your most recent chats grouped by day (Today, Yesterday, Previous 7 days, Previous 30 days, Older). A Show all chats link at the bottom opens the dedicated All Chats page with every conversation.

  • Pin a chat — hover a conversation row, open its three-dot menu, and choose Pin chat. Pinned chats stay at the top under the Pinned group until you unpin them.
  • Rename — open the three-dot menu and choose Rename.
  • Export — open the three-dot menu and choose Export to download the conversation as a PDF.
  • Delete — open the three-dot menu and choose Delete, then confirm. The conversation is removed from your history immediately, and if it is the chat you are in, a new chat takes its place. How long the underlying record is retained before it is permanently purged is governed by your organization's chat-retention policy.
  • Search — press Cmd+K (or click Search at the top of the sidebar) to open the search modal, or use the search bar on the All Chats page to filter by title.

If you'd rather a conversation never be saved to your history in the first place, use Incognito chat.

Starting a new conversation

Click New chat in the sidebar. A fresh conversation begins with a greeting and a grid of topic starters (for example, "Executive briefing" or "Anomaly investigation") that fill the input with a ready-made prompt when clicked. Your current chat stays in the sidebar.

Stopping a generation

Click the Stop button (the filled square on the right of the composer that replaces the send arrow while the AI is producing output) to halt the current response immediately. Strata saves whatever was already produced and marks the response as stopped in the action-step trail.

A response you stopped ends with the line Strata's response was interrupted. and two buttons: Edit prompt opens the message that asked for it so you can change it and send it again, and Try again runs the same message again.

In the web app, Strata knows an answer did not finish. On your next message it sees where the answer you stopped was cut off (or where an error cut one off), so asking it to continue picks up from there, and it never treats half an answer as a finished one.

Leaving while Strata works

You don't need to keep the tab open while a response runs. In a saved chat, Strata keeps working on the server if you close the tab, reload, or go to another page, and saves the full answer when it finishes. Come back to the chat while it is still running and your question and the response pick up live where they are, even in a chat you started moments before leaving; come back later and the finished answer is there.

Switch to another tab while an answer is being written and come back, and Strata glides the chat down to where the answer is now, if you were following it at the bottom. If you had scrolled up to read something, it stays where you left it. Reloading the page, including with Reload on the new-version notice, opens the chat where you were reading, or at its end if that is where you were.

Two exceptions:

  • Incognito and other chats that are not saved have nothing on the server to continue, so the response ends when you leave. Your browser asks you to confirm before leaving in that case.
  • An action that needs your approval (sending an email, a write query, a sub-agent's proposal) waits up to ten minutes for an answer. If nobody answers, it is not done, and the response says so. With browser notifications allowed, Strata notifies you about a waiting approval while its tab is in the background, including approvals waiting in another chat. Allow them from Notify me when a decision is waiting in the Approvals menu at the top of the Sub-agents card in the workspace panel.
  • A question Strata asks in the middle of an answer waits 90 seconds. This is a different thing from an approval, and it behaves the opposite way — see below.

When Strata asks you something mid-answer

Occasionally Strata gets partway through your request and hits a fork it cannot pick for you — two equally reasonable readings of what you asked, or a required input with no sensible default, where guessing wrong would waste the work already done. When that happens a short card appears in the response as it is being written, with the question and two to four one-click choices. Pick one, or type your answer in the message box and press Enter (while the card waits the box reads Answer the question above, or pick an option): the same response carries on from exactly where it was, keeping everything it had already worked out. A message too long to be an answer, or one with files attached, is not taken as the answer: it waits in the queue and sends as your next message, and a line under the card says so.

The last row on the card is always Something else — I'll explain. Pick it and Strata stops, asks in one plain sentence, and waits for your next message.

You are never blocked by one of these:

  • The card says which choice it will continue with, and after 90 seconds with no answer it continues with that one and tells you in a line which default it used, so you can redirect it.
  • Pressing Stop ends the response, as it always does. It never falls back to the default.
  • Only one such question is asked per response.

Questions are rare by design, and only appear in the Strata web app — in Teams and the Outlook add-in Strata picks its own default and says so instead.

A response ended by something other than your Stop (a newer message sent to the same chat from another tab, or a response that stopped making progress) is marked as interrupted, not as stopped by you.

Queueing your next message

You don't have to wait for a long answer to finish before typing the next thing. Send a second message while Strata is still working and it is queued rather than dropped: it appears as a greyed bubble at the bottom of the thread and is sent automatically the moment the current response completes.

Click the × on a queued bubble to pull it back — the text returns to the composer so you can edit or discard it.

Queued messages run in order, and each one sees the completed answer to the message before it. When a response ends with a question for you, queued messages wait until you answer or skip it; a message you send in the meantime waits behind them, so they still go in the order you wrote them.

Editing a message you sent

Hover a message you sent and click the pencil (Edit). The message becomes a box with its words in it. Change them and press Enter, or click Send. Shift+Enter adds a line; Escape or Cancel leaves the message as it was.

The edit is sent as a new version of the message, with the same attachments, and Strata answers it from there: it reads the conversation above the message, not what came after it. Nothing is deleted. The earlier version, its answer and everything after it stay in the chat, and arrows under the message (‹ 2/2 ›) switch between the versions. A regenerated answer works the same way, with its own arrows under the answer.

When you open the chat again, it shows the version you last sent a message in. Anything that reads the chat back for a model (the API's conversation_id, voice and Teams) reads only the version on screen.

You can edit once the current answer has finished or you have stopped it. A message whose words were not kept (because of your retention settings) and a message that answered Strata's questions cannot be edited.

Reading an answer aloud

Click the speaker under an answer (Read aloud) to hear it in your organization's voice, the one voice mode uses. Strata reads the words and skips code, charts and tables. While it reads, the button is Stop reading; it also stops by itself at the end, and when you open another chat. Read aloud uses no microphone, and it is metered like voice mode. It is offered wherever voice mode is: if voice is not set up for your organization, the button is not there.

Replying to part of an answer

Select words in an answer and a Reply button appears over them. It puts the passage in the prompt box as a quote (Quote from the answer), so Strata reads exactly that passage along with your next message. The quote is sent like a pasted text: it shows above your message, and you can remove it before sending.

Switching models mid-conversation

Changing the model from the composer's model picker does not start a new conversation. The new model reads the entire thread above, which is usually what you want — you can escalate a hard question to a stronger model without repeating yourself.

Because that surprises people, Strata shows a short notice when you pick a model that hasn't been used in this conversation yet: Continuing with <model> — your conversation carries over. It stays until you dismiss it.

Two things worth knowing:

  • Because the new model can read what came before, it may summarize an earlier result rather than re-running the work. Ask it to redo a step if you want genuinely fresh output.
  • Switching back to a model that has already answered in this conversation shows no notice — nothing about the context has changed for it.

If you want the model to work without any of the earlier context, start a new chat instead.

What can go wrong

  • "Send" stays disabled. The button only activates once the input has text and no file uploads are still in progress. Wait for any in-flight upload chip to finish.
  • Response stalls without text appearing. Action steps may still be running; expand the action-step trail to confirm the AI is in a tool call (for example, a "Querying the database" step). If nothing moves for a long time, click Stop and re-send.
  • "I couldn't process that request. Please try rephrasing your question." The AI provider's content filter blocked the request. Rephrase it; asking the same thing again is blocked again. If something had already happened before the block (an email sent, a file saved), the message says part of the request went through instead: check the result before asking again, so nothing is done twice. When your organization records blocked requests, this message (and the one shown when one of your organization's own data rules blocks a request) adds Blocked requests are recorded for your organization's administrators., and your privacy settings say so too. The record names you, where you asked and why it was blocked, never what you wrote.
  • An attached file's chip reads Failed. The upload failed (network or unsupported format). Remove the chip with its x button and re-attach. For unsupported formats, convert to one of the accepted types.
  • "You've reached your usage limit". A limit stopped the message. The card names it and when it resets, and says Your org admin has been notified only when a notice really went to your admins (your organization's monthly AI spend ceiling, or its credit allowance outside a trial). When every shared budget you draw on has run out, it says Your team's shared budget has run out, and so does the banner over the message box. At your own monthly spend limit, or your team's shared budget, it offers Request more usage, which asks your admins for more: the request waits for them on the Usage page, and you are emailed the answer, which also appears in your inbox. View your usage opens Settings → Usage.
  • An orange notice appears above the prompt box. Strata has lost contact with its servers. You're offline means your own network dropped; Can't reach Strata. Trying again. means your network is fine but Strata isn't answering — click Retry now to check again immediately. You can keep reading and typing, and the notice clears by itself as soon as the connection is back. See The connection notice. If your session has actually expired or been revoked, Strata signs you out and prompts you to sign in again. See /docs/sign-in.

Related