Features
Content screening
How every write is checked for leaked secrets, injected instructions, and personal data before it's stored.
Content you store will later be read by a model. A leaked API key or a hidden instruction such as "ignore your previous instructions" can be turned against your app long after it was written. Every write to Docklee is screened before it's stored, whether it goes to a container or a memory space.
What's checked
| Check | Looks for | Result |
|---|---|---|
| Secrets | Cloud access keys, private keys, model provider keys, bearer tokens, and lines like password = ... | Blocked, or stored and quarantined |
| Injected instructions | Text that tries to override a model, such as "ignore previous instructions", "you are now", requests to reveal a system prompt, or fake role tags | Blocked, or stored and quarantined |
| Personal data | Personal and health information | Stored and marked as unverified |
| Size | Content over the gateway's size limit | Rejected |
| Binary in memory | A PDF, image, or archive sent to a memory space | Rejected. Send files to a container instead. |
The screen never records the value of a secret it finds, only the kind of secret.
Trust labels
Every stored item gets a trust label.
| Label | Given when | Effect when it's read |
|---|---|---|
| Trusted | Nothing was found | Ranked normally |
| Unverified | Personal data was found | Ranked lower than trusted content |
| Quarantined | A secret or an injected instruction was found and the write wasn't blocked | Ranked lower than trusted content |
Low-trust content is ranked lower rather than hidden, so nothing disappears without a trace. The tags on each item record what was found.
Layers of defense
- Write. The screen labels or blocks content before it's stored.
- Read. Retrieval reads only the sources an engine is linked to, and ranks low-trust content lower.
- Answer. Each answer is checked against the chunks it came from, and invalid citations are removed. See Citations and confidence.
No single layer is relied on alone. If the screen itself fails, the write is rejected.
Screening is a safety net, not a filter for what you should store. Keep secrets out of the content you send, and use your own opaque IDs instead of personal details.
Next steps
Your content is screened on the way in. If something isn't working, start with troubleshooting.