Retrieval Architectures for Governed Knowledge
- Publication date
- Apr 9, 2026
- Author
- SkyNet AI Research
- Status
- Validated
Abstract
Retrieval over organizational knowledge is only useful if access control survives the retrieval step. This paper compares architectures for permission-aware retrieval and reports on the failure modes that appear when access control is applied after ranking rather than before.
Technical documentation
The ordering problem
If retrieval ranks all documents and then filters by permission, the ranking leaks information: result counts and similarity scores reveal the existence of restricted material. Filtering before ranking avoids this but complicates index design.
Architectures compared
We compare partition-level indexes, per-document permission filters applied pre-ranking, and query-time policy rewriting. Partition-level indexing performed best on latency while preserving isolation.
Provenance
Every retrieved passage carries its source, version, and retrieval timestamp. Answers without provenance were not accepted in the evaluated deployments.
Conclusion
Permission-aware retrieval should be treated as a data architecture problem, not a prompting problem.
References
- [1]Lewis et al., Retrieval-Augmented Generation, NeurIPS 2020
- [2]SkyNet AI Core knowledge services specification (2026)