Ask Growth Kit

Ask searches the records your role already lets you read in this project. It does not browse the web, reach another project or another organization, and does not answer from what a model knows about product information management in general.

Recorded text travels to the model as data, never as instructions. A transcript that says to ignore the instructions is quoted material to be described, not a command. Nothing anybody writes in a record changes what Ask does.

Under every answer is one sentence saying how many records were cited, how many matched and how many were searched. Cited two of three matched out of forty searched is a different answer to the same words than two of thirty.

Every record an answer claims to rely on is checked against the records that were actually supplied. A reference to anything else is dropped before you see it and counted, so a fabricated source cannot reach the page. When no citation survives, the answer is shown and marked unverified with the matching records beside it.

Your question is split into words and looked for in a record's title, owner, summary and the fields its kind declares. A hit in the title counts most, then the owner, then the summary, then a declared field. A question about churn will not find a record that only says cancelled, so use the vocabulary of our own transcripts.

The limits are all stated on the page: the 400 most recently updated records are searched, the first 4,000 characters of a field are scanned, at most eight records come back, and each field reaches the model cut to 1,500 characters with shortened records marked as shortened.

Ask will not use outside knowledge, invent a number, date, owner, customer or competitor, or describe a record it was not given. It writes nothing: no record is created or changed, and an answer is not stored as a finding.

When our records do not hold the answer it says so and names what is missing. That is the most useful output on the screen. Write the meeting, the evidence record or the objective it named, then ask again.

Your question is sent in the request body and never put in the address, so it stays out of browser history and request logs. The address names only the open chat, by an id that grants nothing on its own.

Every chat is kept with its questions, its answers and the count line each answer was shown with. The list beside the conversation groups them by when you last used them, pins the ones you keep returning to, and finds one by name. Moving to another chat never stops an answer that is still being written: it finishes and is there when you come back.

A chat belongs to the person who started it and nobody else can find or open it. Sharing it with the workspace lets everyone who can read this project read it under Shared; they cannot add to it, rename it or delete it, and making it private again takes it back. A copied link opens the chat for people who may read it and says it is unavailable for everybody else.

A saved answer is what was true of our records when it was written. It is not refreshed when a record changes, so ask again before acting on an old one.

An admin can switch on a nightly chat review under Settings, Usage. A model then reads the chats it is allowed to, private ones only if the admin included them, and reports what people were trying to do, which data was missing and what to change. Chats are read in memory and kept nowhere; the report names nobody, quotes nothing, and only admins can read it.