The best support prompts are intentionally redundant. They include context the model may need for edge cases, even when only a few sections matter for the current reply. That is good for quality, but it makes every escalation, refund request, and troubleshooting answer more expensive than it needs to be.
TicketPoliciesAccount notesCompressAnswer
What to send through Rose 1
Ticket
Customer message, attachments, product logs, and prior replies.
Policy
Refund, security, escalation, and account rules.
History
Recent orders, plan details, support notes, and open issues.
Answer
The final model call that writes the support response.
Why it works for support
- Keep generous retrieval while reducing the final model input.
- Drop repeated boilerplate from policies and past replies.
- Preserve the user ticket and other protected text that must remain intact.
- Log a receipt with saved tokens, compression ratio, latency, and risk flags.
Try one ticket without an account
Paste a real support-style context block into the capped demo, set the query to the reply you want, and inspect the receipt before creating a workspace.
curl -s https://api.adola.app/v1/demo/compress \
-H 'content-type: application/json' \
--data '{
"model": "rose-1",
"query": "Draft the next support reply and explain the escalation risk.",
"input": "Ticket, policy text, account notes, and prior replies...",
"compression": { "target_ratio": 0.35 }
}'