BlogPII Redaction With Concentrate: What to Filter Before Model Inference
Concentrate enables teams to filter sensitive data before and after model inference requests.
BlogConcentrate enables teams to filter sensitive data before and after model inference requests.
Concentrate.ai
Concentrate can apply redaction to selected sensitive entities in prompts, responses or both. That gives teams a shared place to configure filtering for requests passing through the gateway. Application controls still matter: a gateway cannot inspect a request that bypasses it or a tool call sent directly to another service.
Prompt redaction runs before the configured request reaches the model provider. Response redaction filters selected fields before the result returns to the application. These protect different paths and should be tested separately.
For example, a customer-support assistant may need an email address removed from the prompt but still need the product name and error message. A rule that removes too much can make the answer less useful. Test both sensitive-data detection and the task’s output quality.
Concentrate’s redaction controls can apply configured rules to gateway traffic without each application maintaining its own detector. This is useful when several services call different models but share the same filtering requirements.
Maintain application-side data minimization as well. Avoid collecting or sending fields the model does not need. Route intended traffic through the gateway and separately review retrieval systems, tools, error reports and application logs that may also receive sensitive content.
Use synthetic examples of names, identifiers and free-form notes. Include punctuation changes, unusual formats, partial identifiers and text the detector should preserve. Record false positives and missed entities instead of assuming that one successful example proves complete coverage.
Automated redaction is not a guarantee of de-identification. Replacing a name may still leave information that identifies a person through context. Do not treat a redaction toggle as permission to send PHI to any model or as proof that privacy obligations no longer apply.
For healthcare use, establish the BAA-covered service and approved model routes before PHI is sent. Review prompt and response retention, fallback destinations and logging independently. HHS cloud guidance explains why the service arrangement and safeguards matter alongside technical features.
Concentrate supports BAA-covered inference for approved workloads. Its managed gateway brings model access and the configured data controls together; that does not make every available model eligible for the same healthcare workload.
Review which redaction rule ran and whether prompt, response or both were covered. Limit access to operational records and check whether raw content appears anywhere else in your application’s debugging path.
Review your intended data flow with Concentrate, then test it with synthetic data. The goal is a clear, repeatable request path with documented limits, rather than an unsupported promise that a filter catches everything.
LLM Gateway
Change models, providers, add fallbacks, and more by running all AI usage through one platform.


