Tenant-isolated
It may mean your data is kept apart from other customers'. It says nothing about who can sign in through a connector, who can support it or which model sees it.
Private retrieval-augmented generation
Private RAG is an architecture you can check.
Private RAG is retrieval-augmented generation over your own approved sources. Marella grounds answers that way, and a privacy claim only means something once you can see where the documents are stored, which model reads them, who can get to them and who runs it.
How to check a private deployment
“Private” means nothing until every path across the boundary is named
Organisation boundary
- 01Approved sources
Files, pages, passages
- 02Retrieval
Pick the right passages
- 03Model path
Which model runs, and where
- 04Cited answer
Claims + source links
Inside your boundary
- Your documents
- Who can see them
- Where they are stored
- Which model provider
How RAG works
RAG finds the evidence before the model answers.
- 01
Retrieve
Search the approved documents for passages relevant to the question.
- 02
Rerank
Choose which passages should enter the answer context.
- 03
Generate
Ask the configured model path to synthesise from that context.
- 04
Cite
Resolve material claims back to the passages presented as support.
- 05
Review
Check source, version, support and remaining uncertainty.
Retrieval gives the model something to work from. It does not replace access controls.
Four words vendors use
Ask what “private” means here.
Customer cloud
It tells you who owns the servers. You still need to write down who handles the model calls, the updates, the backups and support.
Self-hosted
Names where the software package runs. It does not automatically mean on-premises, local models, air-gapped or zero-egress.
No training
This is a promise in a contract. Ask which suppliers and which models it covers.
The private RAG worksheet
Follow every copy of the data.
For each one, write down where it lives, who runs it, who can see it and how it gets deleted.
| Artefact | Questions to settle | Evidence |
|---|---|---|
| Source files | Storage, region, backup, deletion | Data-flow + retention record |
| Parsed text + embeddings | Store, isolation, rebuild, purge | Architecture + lifecycle test |
| Prompts + passages | Inference provider, transit, retention | Provider + contract coverage |
| Answers + traces | Event scope, access, export, retention | Event matrix + product inspection |
| Support artefacts | Access route, approval, expiry | Support responsibility split |
Test before you buy
Test the retrieval and the boundary.
A well-designed system can still pull the wrong passage, and a good answer can still travel a route you would not approve. Test both.
- 01
Retrieval relevance
Known answers, plausible distractors, stale and conflicting sources
- 02
Claim support
Citation correctness, completeness and no-answer behaviour
- 03
Access boundaries
Cross-scope questions, revoked access, deleted content
- 04
Data flow
Storage, inference, logs, backup, support and egress paths
Boundary conditions
The word private does not remove these limits.
- RAG does not make a language model inherently correct.
- Tenant isolation does not establish every connector permission path.
- Self-hosted does not automatically mean air-gapped or zero-egress.
- Provider and retention terms must match the selected configuration.
Private RAG FAQ
Does private RAG mean on-premises?
Not necessarily. The term may refer to managed isolation, customer cloud, self-hosting or contractual controls. Name the actual architecture.
Are documents used to train models?
Current product legal materials state a no-training position for customer data. Confirm the relevant provider and contract coverage during procurement.
Does retrieval prevent hallucinations?
No. Retrieval changes what the model has to work from without making it correct, so an answer can still overstate or distort a passage that was retrieved perfectly well. Test whether the retrieved passages are relevant, whether the claims stay supported by them, and whether the system says so when the evidence is thin.
Map first, move data second
Start with the workflow, the kinds of data involved and where it can be hosted.
Do not submit confidential documents through the public form. Agree the transfer route during evaluation.
Plan a private RAG evaluation