> ## Documentation Index
> Fetch the complete documentation index at: https://docs.coreweave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Redact PII from traces

> Automatically redact Personally Identifiable Information from traces

Some organizations process Personally Identifiable Information (PII) such as names, phone numbers, and email addresses in their Large Language Model (LLM) workflows. Storing this data in Weights & Biases Weave poses compliance and security risks. Stripping this data before you log it helps keep your agent compliant with policies like [GDPR](https://gdpr.eu/) and [HIPAA](https://www.hhs.gov/hipaa/index.html).

## Enable organization-level PII redaction

Organization admins can turn on **Redact PII in Weave traces** in the organization's privacy settings. When this setting is on, the Weave server redacts the following from new Calls and Agent OTel writes before storing them:

* Email addresses
* Phone numbers
* US Social Security numbers
* Payment card numbers

Because redaction happens on the server, it applies regardless of the SDK or language you use. It uses a fixed set of entities and applies to new writes only. Turning the setting off doesn't restore values that Weave already redacted. For the steps to turn it on, see [Configure privacy settings](https://docs.coreweave.com/products/wandb/platform/hosting/privacy-settings.md#project-and-content-security).

## Redact PII with the Python SDK

> [!NOTE]
>
> Client-side PII redaction is only available in Python.

The *Sensitive Data Protection* feature lets you automatically redact Personally Identifiable Information (PII) from a [trace](https://docs.coreweave.com/products/wandb/weave/guides/tracking/tracing.md) before Weave sends it to Weave servers. This feature integrates [Microsoft Presidio](https://microsoft.github.io/presidio/) into the Weave Python SDK, so you can control redaction settings at the SDK level.

The Sensitive Data Protection feature introduces the following functionality to the Python SDK:

* A `redact_pii` setting, which you can enable or disable in the `weave.init()` call to turn on PII redaction.
* Automatic redaction of [common entities](#entities-redacted-by-default) when `redact_pii = True`.
* Customizable redaction fields with the configurable `redact_pii_fields` setting.
* Exclusion of specific entities from redaction with the `redact_pii_exclude_fields` setting.

### Enable PII redaction

This section walks you through installing the required dependencies, turning on PII redaction in `weave.init()`, and optionally tailoring which entities Weave redacts. To get started with the Sensitive Data Protection feature in Weave, complete the following steps:

1. Install the required dependencies:

   ```bash theme={"system"}
   pip install presidio-analyzer presidio-anonymizer
   ```

2. Modify your `weave.init()` call to enable redaction. When `redact_pii=True`, Weave [redacts common entities by default](#entities-redacted-by-default):

   ```python lines theme={"system"}
   import weave

   weave.init("my-project", settings={"redact_pii": True})
   ```

3. Optional: Customize redaction fields with the `redact_pii_fields` parameter:

   ```python lines theme={"system"}
   weave.init("my-project", settings={"redact_pii": True, "redact_pii_fields":["CREDIT_CARD", "US_SSN"]})
   ```

   For a full list of the entities that can be detected and redacted, see [PII entities supported by Presidio](https://microsoft.github.io/presidio/supported_entities/).

4. Optional: Exclude specific entities from redaction with the `redact_pii_exclude_fields` parameter. This is useful when you want to keep the default redaction but preserve certain entity types. The following example shows how to redact all [default entities](#entities-redacted-by-default) except `EMAIL_ADDRESS` and `PERSON`:

   ```python lines theme={"system"}
   weave.init("my-project", settings={"redact_pii": True, "redact_pii_exclude_fields":["EMAIL_ADDRESS", "PERSON"]})
   ```

After you complete these steps, Weave redacts the configured PII entities from traces before sending them to Weave servers.

### Entities redacted by default

This section lists the PII entity types that Weave redacts automatically when you enable PII redaction without specifying `redact_pii_fields`. Weave automatically redacts the following entities when you enable PII redaction:

* `CREDIT_CARD`
* `CRYPTO`
* `EMAIL_ADDRESS`
* `ES_NIF`
* `FI_PERSONAL_IDENTITY_CODE`
* `IBAN_CODE`
* `IN_AADHAAR`
* `IN_PAN`
* `IP_ADDRESS`
* `LOCATION`
* `PERSON`
* `PHONE_NUMBER`
* `UK_NHS`
* `UK_NINO`
* `US_BANK_NUMBER`
* `US_DRIVER_LICENSE`
* `US_PASSPORT`
* `US_SSN`

### Redact sensitive keys with `REDACT_KEYS`

In addition to PII redaction, the Weave SDK supports redaction of custom keys with `REDACT_KEYS`. This is useful when you want to protect additional sensitive data that might not fall under the PII category but must remain private. Examples include:

* API keys
* Authentication headers
* Tokens
* Internal IDs
* Config values

#### Pre-defined `REDACT_KEYS`

Weave automatically redacts the following sensitive keys by default:

```json theme={"system"}
[
  "api_key",
  "auth_headers",
  "authorization"
]
```

#### Add your own keys

If your application uses other field names that contain sensitive values, you can extend the default `REDACT_KEYS` list. Add your own custom keys that you want to redact from traces:

```python lines theme={"system"}
import weave
from weave.utils import sanitize

client = weave.init("my-project", settings={"redact_pii": True})

# Add custom keys to redact
sanitize.add_redact_key("client_id")
sanitize.add_redact_key("token")

client_id = "123"
token = "789"

@weave.op
def test(client_id, token):
    return client_id + token

test(client_id, token)
```

When you view them in the Weights & Biases UI, the values of `client_id` and `token` appear as `"REDACTED"`:

```python lines theme={"system"}
client_id = "REDACTED"
token = "REDACTED"
```

### Usage information

* This feature is only available in the Python SDK.
* When you enable redaction, processing time increases because of the Presidio dependency.
