heimdall systems.

Experiment

The tenant filter you can't forget

In a shared Qdrant collection, isolation can depend on one filter in every query. We ran the same search with and without Ten-Grd's restriction.

If you run multi-tenant RAG on one Qdrant collection, with a tenant field on each point, isolation comes down to adding that tenant filter to every query, in every tool, every time. Forget it once and the query still works. It just returns other customers' documents.

The experiment

We seeded a shared in-memory Qdrant collection with two tenants. Both hold a document with the same ID, and the other tenant's vector is the closer match to the query. Then we sent the same search two ways: straight to Qdrant filtered only by topic, and through Ten-Grd. This is the output from running the script on 2 November 2026 with ten-grd 0.3.2:

Query: topic == 'refunds', nearest to the refunds vector

Without Ten-Grd:
  globex  globex refunds within 90 days.
  acme    acme refunds within 30 days.

With Ten-Grd as acme:
  [ok] acme    acme refunds within 30 days.

With Ten-Grd as globex:
  [ok] globex  globex refunds within 90 days.

The data and the query are the same. The only difference is the tenant restriction. The vectors here are fixed toy values, not real embeddings, so this shows the mechanism and says nothing about retrieval quality or performance.

What the integration does

Your code supplies only its own filter:

with tenant_scope(context):
    hits = docs.search(query_vector, predicate=Eq("topic", "refunds"), limit=3)

The boundary

Ten-Grd guards the calls that go through it. Code that queries Qdrant directly with raw credentials bypasses it, and your application still owns authentication and authorization. A supported adapter is not a certification of the backend or of your deployment. To check a path you didn't route through the SDK, use the leak probe.

← All research