Asking Questions with Chat

The Chat tab lets you ask natural language questions across your entire knowledge base. Instead of browsing and filtering manually, just ask what you want to know and get an AI-powered answer grounded in your actual data.

Starting a conversation

Open the Chat tab and type your question in the message box. There is no special syntax required -- write the way you would ask a colleague. Hit send, and Upstream searches your knowledge base, finds the most relevant entities, and generates an answer.

The message box wraps and grows as you type, then scrolls internally for longer drafts. Use Expand message for more drafting space and Collapse message to return to the compact box. Toggling keeps your draft and cursor or selection. On phones, expanding the message collapses Active Memory; reopening Active Memory returns the message box to its compact, auto-growing mode. Short keyboard viewports still limit drafting space. Clearing or sending the draft returns the box to its compact size.

On phone-width screens (below 768px), Enter adds a new line; use Send message to submit. On wider screens, Enter sends and Shift+Enter adds a new line. The message box and send button are disabled while a reply is being generated.

Active Memory starts collapsed on phone-width screens. Its heading remains available while drafting: tap to review or select context, then tap again to collapse it. Switching between memory and expanded drafting keeps your draft, context selections and pins. Typing or incoming context results do not switch which panel is expanded. Wider screens start with Active Memory expanded and allow both panels to stay expanded.

On phones, one compact action row contains app navigation, Conversation history, New chat, the Conversation menu (three dots), and Settings. The separate Chat/model/status bar has been removed. Active Memory context remains available from its heading above the message box. While you have a draft, the mobile welcome illustration, introduction and suggestions are hidden; clearing the draft restores them.

Open the Conversation menu to choose the model and thinking budget for that conversation. Your selection updates immediately while it saves. Rapid changes keep your latest choice; you can close the menu and keep typing. Send waits until that choice is confirmed saved. If saving fails, your desired choice and draft stay in place: choose Retry in the menu or the chat notice, then send after saving succeeds. Confirmed choices survive switching chats or reloading and do not change other chats. Choosing settings before the first message creates one saved, empty conversation. Pending or failed choices remain in this page only; reloading restores the last saved settings.

Tap a thinking-budget preset instead of dragging: Default = 0, Low = 2,048, Medium = 4,096, and High = 8,192 tokens. These are budget shortcuts, not a guarantee of answer quality. The slider still supports 512-token steps; a custom value between presets leaves all presets unselected. Presets are also available in Settings > New chat defaults. They are disabled whenever the budget control is unavailable. Slider touch handling is local to that control; it does not disable normal navigation or guarantee suppression of Telegram's embedded-browser or OS Back gestures.

If the chat backend has not been updated, the conversation controls are disabled with an explanation. Chats, renaming and messages still work with Gemini 3.5 Flash and the default thinking budget. Your new chat defaults stay in your account but are not applied to conversations until the backend supports saving them. A save response that does not confirm your choices is reported as a persistence failure; your desired choice and draft are retained, and Send stays blocked until a retry confirms the choice.

For future conversations, use Settings > New chat defaults, then Save Settings. These preferences belong to your user account and apply to new chats only. Existing chats retain their own settings. Older conversations without saved choices use the standard chat defaults; the old browser-wide model preference is no longer used. Default thinking budget leaves the budget to the existing AI service behavior; custom budgets are available only for the models enabled in the control. See Account Settings.

On phones and smaller tablets, use Conversation history in the Chat header to search or select a saved conversation, or choose New Chat. The drawer closes when you select a conversation or start a new one. You can also dismiss it with Close, Escape, or a tap outside. On wider screens, history stays beside the conversation. Your draft and staged Active Memory context remain available when you open or close history.

This history drawer belongs to the main Chat tab. The embedded chat on the Knowledge page is a separate interface.

Each saved conversation has its own web address (/chat?session=…). Reloading the page, pressing Back after following a citation, or opening the address again returns you to that conversation rather than to an empty new chat. Switching conversations updates the address in place, so Back still leaves the Chat page as before. After leaving Chat, the application's Chat link also resumes the last saved conversation in the same browser tab, account and workspace. This keeps only a navigation pointer in tab storage, not a local transcript; when tab storage is unavailable, use the conversation address or history instead. Unsent drafts remain page-local and are not recovered after leaving the page.

You can draft while a bookmarked conversation loads, but Send and conversation settings wait until its identity and context are ready. Loading never sends your draft automatically. New Chat intentionally leaves the bookmark and starts fresh, clearing any Refine in Chat edit target from the address so reloading does not reopen it. Bookmarks also work for older conversations outside the recent history list. If loading fails, Try again retries that conversation; it does not create a replacement. Older bookmarks require the backend's single-conversation lookup; an unavailable lookup keeps sending blocked rather than dropping your history.

Saved conversations belong to the account that created them, within their organization. Another organization member cannot load, change, delete, or send messages to your saved conversation. If sending reports Not your session, select one of your own conversations or start New Chat. If it reports Session not found (for example, after deletion), start New Chat; sending does not recreate the missing conversation.

Conversation names, emoji and suggested follow-ups

Each saved conversation is named automatically after its first exchange. While that name is being decided, the bar above the messages reads Naming this conversation…. Choose Rename conversation in that bar (or the pencil in the history list) to name it yourself at any time, including while the first reply is still generating: a name you type is kept even if the automatic one finishes later. Older conversations are not renamed.

A conversation can also carry one optional emoji beside its name. After a completed exchange, Upstream may show a Suggested emoji with Accept, Change and Dismiss. Nothing is saved until you choose Accept or Save; Dismiss keeps the conversation as it is, and no emoji is a perfectly good choice. Tap the saved emoji beside the title to open its chooser, or use the smile-plus icon if none is set. Change opens the same chooser for a suggestion. Pick a displayed emoji or type one with your keyboard's emoji picker, then Save; Remove clears a saved emoji. Opening or cancelling the chooser does not save anything or request new suggestions. Saved emoji appear in the history list and survive reloads. If the chat backend has not been updated, the emoji controls do not appear.

After each completed exchange you may also see up to three Suggested follow-ups under the reply. They are written from that exchange and the knowledge it used, so they can be specific; when there is nothing specific to ask, none are shown. Choosing one fills the message box for you to review or edit. It never sends by itself and never starts a new conversation; your pinned context is kept, and the automatic Active Memory suggestions refresh for the new text just as they do when you type. If you already have a draft, the question is added below it. Suggestions are optional: if they cannot be generated (an older backend, a busy AI service, your plan's monthly limit, or more than thirty requests in an hour), the conversation works exactly as before.

Referencing artifacts

Type @ in the message box to search artifacts in the current workspace. Suggestions search the fifty most recent artifacts and show up to five matches, with previews to help distinguish similar names. Select with a tap or Arrow keys and Enter; Escape closes suggestions without sending. You can attach up to five references and remove them from their chips.

An @Artifact name · context chip supplies context, not permission to change that artifact. Before starting a conversation, choose Edit on a reference or use Refine in Chat from the artifact to make it an explicit edit target. Mentioning another artifact never replaces the selected target. The selected edit target has an Open artifact chip above the message box, outside the collapsed Artifact context details. It opens the artifact in a new tab and remains available after a completed turn and when reopening saved history. Saved turns retain reference IDs; reopening resolves the current artifact titles in the same workspace. People and canvas mentions are not supported here.

If a chat response fails, one inline notice appears above the composer, with a dismiss control. Your draft remains available. If Chat cannot produce a final answer, it reports an incomplete response rather than a successful blank reply. Tool activity does not by itself mean the answer finished. Service-unavailable errors do not mean your internet connection is broken. Check the conversation and draft before sending again; an uncertain artifact edit requires read-only reconciliation, not a blind retry.

Multi-turn conversations

Chat preserves context across messages. You can ask a question, then follow up with clarifications or dig deeper without repeating yourself. For example:

"What are the top objections we hear from enterprise prospects?"
"Which of those come up most in deals over $100K?"
"Do we have any messaging themes that address the security objection?"

Each follow-up builds on the previous exchange, so the conversation stays focused and productive.

Citations

Markdown replies can include headings, bold text, bullet lists, tables, links, and fenced code examples. If a reply contains raw HTML, it is shown as code rather than rendered, and an image in a reply appears as its description. Artifact bodies follow the same rules. On phones, assistant replies use the available message width to keep longer answers readable.

Answers can include numbered references like [1] or grouped references like [1, 2]. Each number with returned evidence is a link: Knowledge citations open the entity; retained-source citations open the document. A number without returned evidence stays plain text. A cited excerpt is not proof that the whole document was read. Citations let you check the evidence behind an answer.

Tool use during conversation

Expand Tool activity in a main Chat reply to see individual tool calls and their outcomes. Repeated calls stay separate. Returned means the tool returned a result, not that every requested change succeeded. Partial indicates reported limitations or incomplete work; Error does not rule out changes already saved. An unconfirmed outcome is not success or failure—check the affected item before sending again.

Activity uses short server-generated summaries, not raw tool arguments, source text, error details, or private model reasoning. Up to 32 calls are shown per turn; at that limit, a notice warns that later calls may be omitted. Successfully saved replies retain this activity when you reopen the conversation. Older messages have no recorded activity; older backends show only the existing stage badge. If a turn fails or disconnects, its received activity remains in the local error notice with saving explicitly unconfirmed. Reloading or dismissing that notice can lose it; failed-turn recovery and a full audit export are not provided.

The AI can do more than answer questions. During a conversation, it can:

Create artifacts -- Ask it to generate a battle card, account brief, or summary and it will build one from your knowledge base right in the conversation. Generate Artifact can also save a completed, nonempty reply. Empty responses cannot be converted; you can still intentionally start a blank draft from the Artifacts page.
Create living artifacts -- Ask for a live dashboard or a report whose numbers must stay in sync. Chat defines each figure once in a dataset and writes the views from it. Edit the dataset, then use Update linked views on it to re-render every view. If Content Spy has crawled your site, ask for a site overview; it is bound to the crawl, so Refresh from source (or asking chat to refresh it) re-imports the site, updates the dataset and the views, and says whether anything changed. For a GitHub repository whose issues are imported, ask for its repository status; it refreshes the same way, through the saved GitHub connection. Rows you set to manual in the source column are yours and are never overwritten. Any artifact can quote a dataset's numbers; ask chat what a change reaches and it lists every artifact that quotes the value, and updates them only when you say so. A bound artifact can carry a refresh cadence (daily, weekly, monthly); saving cadence proposes upkeep; an owner must approve a finite allowance before chat or an external trigger can run what is due. A custom living artifact has no bound source; its values are a snapshot until one is added, and scheduled refresh is not available yet. See Using Living Artifacts for the full guide.
Read the wiki and propose changes -- Ask about your team's wiki and the AI searches it and reads the pages it needs; under its answer you see which pages it read. Ask it to create a page or change one and it proposes the change rather than making it: the proposal waits on the Review screen, and nothing on the wiki changes until someone approves it there. To propose an edit, the AI must first read that page while answering the same message, and the proposal is tied to the version it read. Each message allows at most 3 wiki searches, 5 page reads and 3 proposals. Anyone in your workspace can write wiki text, so once the AI has read wiki pages while answering a message, it will not change an existing artifact in that answer: it describes the change and asks you to confirm it in a new message. It can still create a new artifact. See Reviewing suggested knowledge.
Propose new entities -- If the AI identifies an insight worth capturing, it can suggest adding it as a new entity to your knowledge base.
Search your knowledge -- The AI actively searches your entities, relationships, and documents to find the best answer. You will see when it pulls in relevant context.

For proposed knowledge entities, choose Approve to save the entity, Reject to dismiss it, or Revise to suggest changes. On narrow screens the card spans the full width of the conversation, and the actions wrap onto another line only when they do not fit. In the revision field, Enter submits your feedback and Escape cancels it; you can also use the submit and cancel buttons. Submitting feedback places a revision request in your message box for you to send.

Optional source screening

In an enabled workspace, ordinary Chat can discover source metadata, call screen_document with a question, then read selected evidence. Jev receives only the requested bounded retained-text page; Chat receives advisory use, skip, or inspect, coverage and read/continuation arguments—not that source body until it reads it. Screening is not citable evidence, truth certification, or automatic removal of pinned context. An incomplete capture or chunk cannot justify rejecting the whole source; unavailable screening falls back to reading, not automatic retries. This is off by default, not a per-keystroke operation, and does not expand selected-artifact edit mode, Canvas or Sprint Planner. Recorded attempts appear under Billing and Plans.

Example questions by role

Sales:

"What objections do prospects mention most about our pricing?"
"Give me a quick account brief for Acme Corp before my call tomorrow."
"What pain points do mid-market companies mention compared to enterprise?"

Marketing:

"What messaging themes resonate with technical buyers?"
"What content ideas have we captured from customer feedback this quarter?"
"How do prospects describe our product compared to Competitor X?"

Product:

"What are the most requested features from enterprise customers?"
"Are there any process gaps that multiple customers have mentioned?"
"What customer feedback have we received about the onboarding experience?"

RevOps:

"What deal context patterns appear in deals that closed this quarter?"
"Are there common objections in deals that stalled?"
"What market insights have we captured about our target segments?"

Tips for getting better answers

Be specific. "What objections do enterprise prospects raise about security?" will give you a more focused answer than "What are our objections?"
Reference timeframes. "What feature requests came in this quarter?" narrows the scope and gives you more relevant results.
Ask follow-ups. If the first answer is too broad, ask a more targeted follow-up. The AI remembers the full conversation.
Name what you need. If you want a summary, a list, or a comparison, say so. "Compare the top three competitor mentions by frequency" tells the AI exactly what format you expect.
Use entity types. Mentioning types like "pain points," "objections," or "messaging themes" helps the AI focus on the right part of your knowledge base.

Importing web context

Use Add context (+) → Search & import web to search the web yourself, then choose Import beside a result. Browse URL extracts a page whose address you already know. Both need you to be signed in, and both hand the selected content to the same import flow as an uploaded file; searching alone imports nothing.

These actions do not give the AI a web-search tool or change what it can do on its own. A search or import error is separate from any AI-service error in the conversation. While a search or URL extraction is running, pressing Enter again does not send a second request.

Using the Knowledge Browser -- Browse the entities that power chat answers
Creating Artifacts -- Generate deliverables directly from chat
Using the Canvas Workspace -- Organize ideas visually after a chat session

Still need help?

Email us at hello@getupstream.ai -- we read every message.