Theme mode
06 / audience acl

Audience ACL Security

Cross-tenant injection defense is enforced in SQL, Qdrant filters, and hydrate re-checks.

Phase 1 Beta
Step 1
source chunk
Step 2
audience tag
Step 3
scoped key
Step 4
Qdrant retrieval filter
Step 5
blocked result

Security posture widget

The cross-tenant moat uses Postgres RLS before retrieval, Qdrant payload filters during vector search, Qdrant collection payloads for audience tags, Qdrant result hydration checks, Qdrant cache keys with scope hashes, and Qdrant false-positive evals. Qdrant is never trusted alone: SQL scope, Qdrant scope, and response assembly must all agree before a chunk reaches the model. This cross-tenant path is red-team tested against malicious chunks and malicious prompts.

Cross-tenant denial probe

Switch workspace, run the query, and the MSW recall endpoint returns only that tenant's scoped result or a cross-tenant denial.