Quberi

Quberi is a knowledge base that AI agents can read, with permissions enforced at retrieval time, down to a single document.

Most knowledge bases stop at workspace-level access: if you can open the space, you can read everything in it. That breaks as soon as one space holds documents with different audiences — a company handbook everyone reads, and a customer drawing with a part number and a customer name on it. Same space, same folder, different audiences.

What's different:

• Per-document ACL with deny-wins. A grant cannot override an admin's decision to hide one document.
• Filtering happens inside the vector index, not after retrieval. You get a full result set instead of a shorter one with holes, and nothing leaks the existence of a document you are not allowed to see.
• Revoked access kills already-issued links: a preview URL returns 404 on the next request.
• Uploads go through review. Unapproved content is never vectorized, so it cannot be searched, cited, or counted.

Writes are scoped: an agent only saves into spaces you can edit. Read, write and upload are separate OAuth scopes, so a read-only workflow never asks for write permission.

Bring your own model — OpenAI, DeepSeek, Moonshot/Kimi and others. Your documents are not used to train anything.

Core is open source (AGPL); the DeepSeek Harness plugin is MIT. Try it as a guest with no signup.

Not there yet: no SSO beyond Feishu and no SAML/SCIM; image search is weak on technical line drawings; a few retrieval paths have no relevance floor.