Private RAG boundary and retrieval setup note
An architecture exploration for document search that keeps the processing boundary explicit.
- Owner
- content-team
- Reviewed
- 2026-09-12 · Next review 2027-03-12
- Evidence level
- Architecture
- Citation scope
- Cite as an in-development architecture note only; do not present it as a benchmark or security certification.
Hypothesis
A private retrieval path can keep document payloads within the agreed boundary while leaving updates, telemetry and support paths explicit.
Setup and hardware
Synthetic document corpus and a proposed on-premise processing boundary; no customer documents or credentials.
Edge appliance selected after site assessment · Runtime and model selected per workload
Method
Map payload, retrieval, logs, updates and support separately, then define the human review and citation checks needed for a pilot.
Result
The boundary can be described and reviewed, but no retrieval-recall or grounded-answer performance result is published.
Failed attempts
- Treating all network traffic as document payload obscured the separate telemetry and update decisions.
Limitations
- Architecture describes a deployment boundary, not a production performance guarantee.
- Remote support and update paths remain subject to the agreed operating scope.
Next steps
- Version a representative corpus and technician-authored question set.
- Run a scoped Solution Architecture workshop before any pilot claim.
Citation scope
Cite as an in-development architecture note only; do not present it as a benchmark or security certification.