Direct answer. Do not approve a Mistral-powered knowledge assistant on the strength of "EU endpoint" or "zero data retention" alone. Each control has a narrow, documented scope. Approve only when you have demonstrated, in your own environment, the data flow, retrieval permissions, log scope, model status and licence, and operating responsibility.
Method and boundary. This is public-source analysis of three vendor documentation pages. Their statements are provider assertions, not independent validation. The extracts carry no capture dates, so treat pricing, endpoints and feature limits as time-sensitive and re-check them before any decision. The matrix below is my proposed method. It describes no deployment I performed and implies no Mistral partnership or certification.
What zero data retention covers, and what it excludes
When ZDR is enabled, Mistral does not store or log inputs and outputs for supported API requests longer than required to generate the output Zero data retention. It applies to stateless API calls, and the listed endpoints include chat completions, embeddings and OCR Zero data retention.
The exclusions matter for a knowledge assistant. ZDR does not apply to Agents, Batch processing files, Conversations, Libraries or /v1/files Zero data retention. It also excludes Labs and Public Preview models Zero data retention. Traces sent to Studio Observability are stored when it is enabled Zero data retention. ZDR is not automatic: Mistral reviews each request and may approve or deny it Zero data retention.
ZDR and training opt-out are separate controls, and you do not need ZDR to opt out of training Zero data retention. The page says ZDR is available on pay-as-you-go for supported stateless API calls Zero data retention. Confirm that your contract tier matches.
My analysis: a design that keeps documents in your own store and sends only retrieved passages to a stateless endpoint is the design most likely to fall within ZDR scope. That holds only if the model is neither Labs nor Public Preview and ZDR is approved and visible in your admin settings. A design that uploads manuals to a vendor file or library service is outside the documented scope.
What regional inference does and does not fix
Mistral offers regional inference through dedicated endpoints that process requests on systems in a chosen geography, currently the European Union or the United States Regional inference. If you do not specify a region, the API uses the global endpoint Regional inference. Regional inference is billed at 1.1× standard list pricing Regional inference.
The limits are explicit. Account configuration, API keys, billing, access management, usage analytics and other operational metadata may still be handled by Mistral systems outside the selected geography Regional inference. Function calling is the only supported regional tool, and stateful features including Agents, Batch and the Files API are not available on regional endpoints Regional inference. Available models vary by region Regional inference. If a workload depends on an unsupported feature, the regional processing guarantees do not apply to it Regional inference.
My analysis: inference location is one question. Retention, control-plane handling, model licence and who operates the service are separate questions. Open weights alone would not settle them either.
Access controls on a third-party hosted endpoint
As a neutral comparison point, Hugging Face Inference Endpoints offers three access levels: Private (default), Public and Authenticated Configuration. Endpoints deployed on an AWS instance can use AWS PrivateLink to restrict access to a specific VPC Configuration. By default an endpoint is accessible from the Internet and secured with TLS/SSL Configuration.
The extract does not say which Mistral models are offered there, their licence terms, who operates the underlying infrastructure, or what retention applies. Obtain that evidence before treating this route as an option.
Deployment decision matrix (proposed method)
| Dimension | Question to answer | Acceptance evidence | Stop criterion |
|---|---|---|---|
| Data flows | Which endpoints does each assistant function call? | Network capture and endpoint inventory; regional base URL pinned; only endpoints you have confirmed in scope | Any call reaches the global endpoint, or a Files, Libraries or Agents call carries plant content |
| Model status | Is the chosen model outside Labs and Public Preview, and available in the target region? | Model list queried against the regional base URL; written status for the exact model | Model is Labs or Public Preview, or not served in region |
| Retrieval permissions | Are user entitlements enforced before text enters a prompt? | Tests with restricted documents and unprivileged users | A restricted passage appears in any response |
| Logs | Where do prompts, outputs and traces persist? | Written inventory of your own logs; Observability disabled or scoped | Traces with document text stored outside agreed retention |
| Model licence | Do the licence terms of the exact model version permit your use and deployment location? | Licence text for that version, reviewed by counsel | Licence not obtained or use not permitted |
| Operational responsibility | Who owns patching, access reviews, incident response and ZDR status checks? | Named owner per control; ZDR visible in admin privacy settings | No named owner, or ZDR request not approved |
| Metadata | Is control-plane handling outside the region acceptable? | Documented risk acceptance | Policy requires all data, including metadata, to stay in region |
Owners and tests are my recommendations. The three sources do not establish licence terms, certifications or performance, so each row needs buyer-obtained evidence.
Trade-offs and when this does not apply
- Vendor API with ZDR and EU endpoint: fastest to test, but you inherit the documented feature gaps and rely on provider statements.
- Retrieval in your own store, stateless calls only: keeps documents under your control, at the cost of building and securing the retrieval layer yourself.
- Third-party hosted endpoint with network restriction: access can be narrowed, but operating responsibility, model availability, licence and retention must each be established from evidence the extracts do not supply.
This does not apply if the assistant handles only public documentation, where a lighter review may do. Evidence that would change my view: a signed data-processing agreement covering control-plane metadata, or documented regional support for the file and retrieval features you need.
Next step
If your team cannot yet fill the matrix, that is the consequential decision. Bring your document classes, intended endpoints, entitlement model and licence questions to a Physical AI Product Decision Review at Hyperion.
