Qdrant

Qdrant holds only vectors and the minimum metadata retrieval needs. Users, companies, document records, and authorisation live on the relational side, in PostgreSQL. This hybrid model is a deliberate choice.

When a query arrives, the question is converted into a vector and the nearest passages are searched within that company’s collection. The returned passages are handed to the language model as context, and the answer is produced from them.

Qdrant is not an access control system. Who can see what is decided in the relational layer; the vector layer only searches the collection it is asked about.

A separate collection is created for each company. Isolation is therefore enforced by a collection boundary rather than a query filter: one company’s vectors are never present in another company’s search space.

That is a stronger guarantee than filter-based isolation. A filter can be forgotten or written incorrectly; a collection boundary cannot. In multi-tenant systems this choice structurally lowers the risk of data leakage.

It also helps operationally: removing a departing customer’s data means dropping one collection, re-indexing does not affect other customers, and audit questions get a concrete answer.

Vector counts reach multiples of document counts because each document is split into several chunks. The document limits in plan tiers are therefore not vector counts.

Memory use and query latency scale with that number. When problems appear, the first place to look is assistant scope rather than index parameters; an over-broad scope inflates both latency and imprecision at once.

In a self-hosted deployment Qdrant runs in your own infrastructure, and backup, version upgrades, and capacity planning become your responsibility.

A useful way to evaluate qdrant is to follow one real question from the source document to the final answer. Record which file was selected, what context reached the model, and what a reviewer would need to verify. This turns a definition into an operational check and makes the result comparable across teams.

The same check should include ownership and change management. Decide who updates the relevant documents, how an outdated result is reported, and which access boundary applies. KobiGPT can provide the assistant and the cited document context, but the organisation still owns the source material, permissions, and the decision made from the answer.