Asking questions well
Chat answers come from your organization's sources, with citations and a confidence score. A little care in how you ask goes a long way.
Phrasing that works
Use exact identifiers. Part numbers, SKUs, account names, and ticket IDs are looked up exactly, and they carry into follow-ups. "What is the lead time on the seal kit?" works after you have named the kit once.
Name the source when you have one in mind. There is no special syntax. Say it in plain words: "in our ERP", "on the pricing page", "the spreadsheet John added last month", "from our Jira desk". If no source matches, the answer says so, names the closest sources it found, and asks which you meant.
Ask for prices and stock as such. Prices and availability come only from records, never from document text. If the record does not have the value, the answer says it needs to look it up rather than guessing.
Ask for counts directly. "How many open orders over $10k?" is computed exactly from records when the data allows it. If it cannot be computed, the answer says it is answering from documents and figures are approximate.
Multi-part questions are fine. On Thorough and Research effort they can be split into sub-questions; on Fast and Standard they are answered together.
Choosing an agent
The agent button at the bottom left of the composer, next to +, shows the name of the agent for this chat. Click it to pick another. An agent defines which knowledge can be read, so pick the one built for the job. Choosing an agent also switches to its default model and answer mode, when it has them set. Set your usual agent under Account settings → Preferences → Default agent.
Mode, effort, and model
The button at the bottom right of the composer, next to send, opens Model, thinking effort & answer mode.
Mode sets how strictly the answer sticks to your sources:
Strict — answers only what it can support from organization knowledge.
Balanced — prefers cited evidence; clearly marks any general-knowledge additions.
Flexible — uses organization knowledge as context without treating it as a hard boundary.
Effort is a slider from Faster to Smarter, with Off at the far left:
Auto — adjusts search and thinking per question. A safe default.
Fast — little searching, little thinking. Fastest, but can miss things.
Standard — a moderate amount of both. Reliable for most questions.
Thorough — searches wider and thinks longer. Slower, more complete.
Research — the most searching and thinking. Slowest, for the toughest questions.
Off — does not search your knowledge; replies from this chat and attachments only.
Auto picks Fast for small talk and direct record lookups, Thorough for comparisons, troubleshooting, reports, and multi-part work, and Standard otherwise. Fine-tune splits effort into Depth (sources) and Thinking (Auto, Low, Medium, Max).
Model lists the models your organization has enabled. Council lets several models or agents answer and cross-check for decisions that matter; it costs more and takes longer.
Attachments
Click the + button (Add files & capabilities) and choose Add files or photos, or drop files on the composer. Images, PDF, DOCX, XLSX, CSV, HTML, JSON, Markdown, and plain text are read; other types are attached but not read. Files can be up to 50 MB. The same menu lists the skills the agent can use. If the composer is empty, clicking one starts your message with "Use {skill} to".
Follow-ups and context
Each chat keeps its own thread: recent messages are kept word for word and older ones are summarized. Phrases like "last time" or "you said" pull in up to three of your earlier chats, and "where did that come from?" brings back how the previous answer was produced. If you named a source earlier, follow-up questions usually stay inside it until you name a different one or change topic.
Press Enter to send and Shift+Enter for a new line. While an answer is streaming, sending again queues your message by default; you can switch it to interrupt and steer under Account settings → Preferences → Sending while an answer streams. ⌘/Ctrl+Enter always does the other one. Press ↑ in an empty composer to edit your last message. Stop generating keeps the partial answer.
When the answer says it does not know
In Strict mode, an answer without supporting sources is declined rather than guessed. Balanced does the same unless the answer clearly marks the uncited part as general knowledge. The usual wording is "I don't have enough information in the organization's sources to answer this confidently, and I'd rather not guess." Declines usually show the No sources badge.
What to do:
Rephrase with a more specific identifier or source.
Raise effort to Thorough.
If the topic should be covered, add the source material. Questions that score low and are asked repeatedly show up under Knowledge gaps in Analytics, which is how Editors spot what to write next.
Feedback
Under each answer: Good answer or Wrong or unhelpful — flags a knowledge gap. A thumbs down asks What was wrong? with Incorrect, Outdated, or Missing source, an optional note, and Flag it. Admins review flagged answers under Settings → Quality, and the helpful percentage shows in Analytics.
Other actions: Copy (with or without citations), Retry this answer, and Save to knowledge, which files the answer as a draft note.
Private chat
In a new chat, turn on Private chat (the button in the chat header) to keep the conversation out of history. Attachments are disabled in Private chat, and no Creations are made.