Redaction

Reusable redaction for diagnostic text, log messages, and structured data.

Use redact_text before displaying untrusted text in logs or diagnostics. Use redact_sensitive_data for mappings that have an explicit safe-key list. Both functions return a redacted copy and leave their input unchanged.

`redact_sensitive_data(data, safe_keys=None)`

Return a copy of structured data with non-whitelisted fields masked.

Parameters:

Name	Type	Description	Default
`data`	`Any`	Mapping, list, tuple, or scalar to redact.	required
`safe_keys`	`Container[str] \| None`	Case-insensitive allowlist of mapping keys. Defaults to `SAFE_KEYS`.	`None`

Returns:

Type	Description
`Any`	A redacted copy of `data`; scalar values retain their original type.

Example

Keep a model identifier while masking a credential.

from gllm_multimodal.utils.redaction import redact_sensitive_data

print(redact_sensitive_data({"model_id": "vision", "api_key": "secret"}))

Output:

{'model_id': 'vision', 'api_key': 'REDACTED'}

`redact_text(value, *, limit=None)`

Redact common secrets, URLs, and local paths from a text value.

This is suitable for lower-level error messages, tracebacks, and strings written to logs. A length limit is optional so callers can set their own storage boundary without changing the redaction rules.

Parameters:

Name	Type	Description	Default
`value`	`str`	Text to redact.	required
`limit`	`int \| None`	Maximum output characters before a truncation marker. Defaults to no length limit.	`None`

Returns:

Type	Description
`str`	The redacted text, optionally bounded to `limit` characters.

Example

Redact a log message before emitting it.

from gllm_multimodal.utils.redaction import redact_text

print(redact_text("api_key=secret-value"))

Output:

api_key=[REDACTED]