Azure AI Language PII alternative: keep the client, change the endpoint
ShinrAI implements the supported Azure text PII request and response contracts. Point your existing client at the ShinrAI base URL, keep your data in Germany, and add restorable replacements when you are ready. Azure-class quality, with published evidence.
Why teams look for an Azure PII alternative
- Data location. Contracts, works councils and public-sector rules often require processing in Germany or in the EU, by a European provider.
- Restoration. Azure redacts text. A reversible, customer-held mapping is not documented in the cited PII policy. AI workflows need the original values back in the response.
- Useful text. Masks and labels remove context. Realistic replacements keep names, places and dates readable for the model.
- Budget shape. Azure bills per text record, with support purchased separately. ShinrAI publishes monthly plans, non-expiring packs and optional onboarding.
- Own hardware. Azure offers PII containers with container billing conditions. ShinrAI offers offline packs with signed activation for disconnected deployments.
The compatibility adapter
Your Azure client keeps working. Set the endpoint to https://azure.api.shinrai.innovius.io and put your ShinrAI API key in the Ocp-Apim-Subscription-Key header. Only the endpoint changes; the paths stay Azure's paths. The same adapter also runs on the API host under https://api.shinrai.innovius.io/v1/azure. The adapter implements these text PII contracts:
The Azure contract limits what ShinrAI can return: types, confidence, media, sessions and jobs. Results can therefore be weaker than ShinrAI can deliver. For full quality, use the native PII API v2.
| Contract | Path |
|---|---|
| Modern synchronous | POST /language/:analyze-text?api-version=2023-04-01 or 2026-05-01 |
| Modern asynchronous | POST /language/analyze-text/jobs?api-version=2023-04-01 or 2026-05-01 |
| Legacy synchronous | POST /text/analytics/v3.1/entities/recognition/pii |
| Legacy asynchronous | POST /text/analytics/v3.1/analyze |
All supported versions map Person, Organization, Email, URL, PhoneNumber, IPAddress, Address, CreditCardNumber and InternationalBankingAccountNumber. The 2026-05-01 contract adds City, DateOfBirth, NationalId, ZipCode and LicensePlate, with category inclusion and exclusion, confidence thresholds, value exclusions and character, entity or no-mask redaction. Synchronous requests take up to 5 documents of 5,120 grapheme clusters; asynchronous jobs take up to 25 documents and 125,000 text elements within 1 MiB.
Compatibility concerns protocol behaviour, not identical predictions. The default model is ShinrAI 1.4; dated Azure model identifiers, the phi domain, context synonyms and Azure synthetic replacement are rejected explicitly. The migration guide lists every supported option, limit and warning.
Restoration mapping: the step beyond redaction
The Azure adapter returns redacted text in the Azure format. When your workflow needs the original values back, switch that call to the native ShinrAI API. Request mode: replace with include_mapping: true. The response contains protected text, a mapping and the detected entities. Your application stores the mapping, sends the protected text to the AI model, and restores the original values in the reply. The mapping never leaves your application, and restoration uses no API records.
Many teams migrate in two phases: first the endpoint swap through the adapter, then the native call for the workflows that need restoration.
EU hosting: Germany, or your own infrastructure
The managed API runs on STACKIT in Germany, operated by Innovius UG in Teltow. Synchronous content is processed in memory, without a request-content archive or model training. Asynchronous job content is encrypted and expires 24 hours after submission. The data processing agreement covers Article 28 GDPR, security measures and subprocessors.
For disconnected environments, an offline pack with a signed activation file binds usage to your installation; processing stays local. See the on-premises section.
Quality, with the context attached
ShinrAI 1.4 reaches a median strict-span F1 of 94.8 across 15 locales on 3,000 business, medical and administrative texts. On the 27-message customer-chat set with overlap matching, ShinrAI 1.4 scores 84.5 and Azure GA 81.1; the Azure preview model leads that set with 88.8. The sample is small, and the two evaluations are not directly comparable. Run a pilot on representative texts. Methodology and versions are on the benchmark board.
Migration steps
- Create a key. Choose a plan on the plans page, log in with your work email and create an API key in the dashboard.
- Swap endpoint and header. Set the endpoint to
https://azure.api.shinrai.innovius.io(sandbox:https://azure.api-sbx.shinrai.innovius.io) and pass the key inOcp-Apim-Subscription-Key. - Check categories and language codes. Confirm that your categories are in the supported list and that your language codes are accepted: de, en, fr, es, it, pl, pt, ru, uk, tr, ar, he, ja, ko.
- Handle warnings. Native findings without a verified Azure mapping produce
ShinrAIUnmappedEntityTypeswarnings. Use the native API when those findings must be kept. - Evaluate on your documents. Compare predictions on a representative sample before you switch production traffic.
- Add restoration. Move workflows that need original values to the native API with
include_mapping.
The Azure compatibility page in the developer portal has the complete request and response examples. Google Cloud DLP and AWS Comprehend adapters are documented next to it.
Plan a measured pilot
Start on a monthly plan from €39, or ask a certified partner to review your integration during optional onboarding.
Frequently asked questions
Do I have to rewrite my Azure PII client?
No. The adapter implements the supported synchronous and asynchronous text PII contracts for api-version 2023-04-01, 2026-05-01 and the legacy v3.1 endpoints. Change the endpoint to https://azure.api.shinrai.innovius.io and put your ShinrAI key in the Ocp-Apim-Subscription-Key header.
Does ShinrAI return the same entities as Azure?
No. Compatibility concerns the protocol, not identical predictions. Supported categories are mapped; unsupported categories and options fail explicitly. Native findings without an Azure mapping produce a warning. Evaluate on your own documents before switching.
Where is my data processed?
On STACKIT in Germany. Synchronous content is processed in memory, without a request-content archive or model training. Asynchronous job content is encrypted and expires after 24 hours. On-premises deployment with signed offline activation is available.
How do I get the original values back after the AI model replies?
Use the native API with mode replace and include_mapping. Your application stores the mapping and restores the values locally in the response. The Azure adapter itself returns redacted text in the Azure format.
Which Azure PII features are not supported?
Dated Azure model identifiers, the phi domain, context synonyms, disabling detector validation and Azure synthetic replacement are rejected. Conversation and document PII have their own supported scope; see the developer portal for the release matrix.