Knowledge imports
Upload an export of your ChatGPT or Claude history and Sparky distils what it says about your brands. Your raw conversations stay private to you; only the findings are shared.
/importsAny organization member, when the feature is switched on. Applying a finding needs Editor, Manager or Owner.
Availability
This feature is off unless a platform administrator has enabled it for your installation. When it is off, the sidebar entry is hidden and the address redirects to your brands. If you cannot find it, that is why.
With it on, a Manager or Owner sees every import in the organization. Everyone else sees only their own.
What is shared and what is not
This is the part worth understanding before uploading anything. The conversations you upload are readable only by you. What reaches your colleagues is the findings: short distilled statements about a brand.
A finding can never quote your conversation. Verbatim text is rejected rather than stored, which is what makes a finding safe to show a manager while the conversation behind it stays yours.
The point is that a manager can approve a statement about the brand's voice without reading the chat where you worked it out.
Running an import
Four steps in one dialog.
| Step | What it asks |
|---|---|
| Source | ChatGPT or Claude. Gemini and Grok are listed but not supported yet. |
| Upload | The .zip export from that service, up to 150 MB. The dialog gives the exact steps for getting one out of your account. |
| Pick conversations | A searchable, filterable table of what the export contains, with the credit estimate updating as you select. This is where the cost is decided. |
| Target | Where the knowledge belongs: let Sparky route each finding, restrict it to brands you name, or both. |
Credits are charged on the size of what you select, not per conversation or per import, and the running total is shown before you commit. A member with a personal allowance sees how much of it is left.
After it runs
A large selection is split into several sessions and worked through in order. The card shows progress live, and a crash partway keeps everything already found rather than starting over.
Sparky may ask a clarifying question mid-import. It does not block: the question arrives as a notification, and answering it feeds into the next batch or resumes a finished import.
When it finishes, the import needs review. Each finding is applied or dismissed, and applying writes it into the brand's knowledge.
A finding never overwrites something you wrote yourself. An agent write onto a user-authored document is kept as a revision and the live text is left alone.
A finding scoped to the organization is appended to the organization's instructions, never replacing them.
Worth knowing
Contributing is not approving. Uploading and selecting need only the ability to run automations; applying a finding to a brand needs Editor and access to that brand.
Every imported conversation also becomes a dated entry in the brand's memory, using the conversation's own date rather than the date you imported it, so a nine-month history stays in order.
Deleting a brand does not destroy an import whose findings reached other brands.