Who sees your prompts
For independent-seller endpoints: Seller can see prompts. The independent seller receives your request content to run inference; this is not guaranteed seller-blind inference. Listed providers, including providers used with your own API key (BYOK), have their own API/inference data policies. Review the dated provider summaries and primary policy links before sending content: Provider policies.
Each endpoint in /v1/models/{author}/{slug}/endpoints shows zalmir.prompt_visibility. Responses carry X-Zalmir-Prompt-Privacy: not-confidential.
Limit where requests go
| Setting | Effect |
|---|---|
provider.data_collection: "deny" | Only endpoints recorded as neither training on nor retaining your data. Unknown counts as no. |
provider.zdr: true (or top-level zdr) | Only endpoints recorded as zero data retention. |
curl "$BASE_URL/chat/completions" \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "<model id>",
"messages": [{ "role": "user", "content": "Hello" }],
"provider": { "data_collection": "deny", "zdr": true }
}'Set either one as an account default on your Account page and it applies to every request; a request cannot turn it off.
Confidential inference
Not available yet. A request with X-Zalmir-Require-Privacy: confidential or X-Zalmir-Require-Model-Proof: attested is refused (503 confidential_inference_unavailable, 503 model_attestation_unavailable) rather than served without it.
Saved answers
After a charged answer, Zalmir keeps an encrypted copy so you can fetch it if your connection dropped. Delete it with DELETE /v1/requests/{request_id}/response. See Errors and headers.