Struct DlpEngine 

pub struct DlpEngine { /* private fields */ }

The compiled DLP engine. Built once per config (re)load and carried on the proxy Runtime.

Implementations

impl DlpEngine

pub fn build(cfg: &DlpCfg) -> Result<Option<DlpEngine>>

Build from [llm.dlp]. Returns Ok(None) when the mode is off (the proxy then skips DLP entirely). A bad custom regex / mode / style — or [llm.dlp.ner].enabled without the ner feature compiled in — fails here, at startup/reload.

pub fn mode(&self) -> DlpMode

pub fn scan_request(&self) -> bool

pub fn scan_response(&self) -> bool

pub fn stream_redact(&self) -> bool

Whether streamed SSE frames should be rewritten (deterministic spans only) in redact mode. Never true in reversible mode — there the response is unmasked, not redacted.

pub fn reversible(&self) -> bool

Whether reversible masking is active (redact mode + [llm.dlp].reversible). When true, an inbound finding is replaced with a placeholder recorded in a MaskMap, and the response is unmasked from that map instead of being scanned/redacted.

pub fn scan(&self, text: &str) -> Vec<Finding>

Scan text, returning every finding including the NER family when compiled/enabled. Used on the buffered request/response bodies. Findings are returned sorted by start offset, overlaps merged.

pub fn scan_stream(&self, text: &str) -> Vec<Finding>

Scan text with the deterministic families only (signature / gazetteer / entropy) — never the NER model. Used on the streaming path, where running ML over partial, mid-token frames would be both slow and unreliable. Findings are sorted by start offset, overlaps merged.

pub fn redact(&self, text: &str, findings: &[Finding]) -> String

Replace each finding’s span per the configured RedactStyle. findings must be the sorted, merged output of scan / scan_stream.

pub fn redact_reversible( &self, text: &str, findings: &[Finding], map: &mut MaskMap, ) -> String

Redact reversibly: replace each finding’s span with a placeholder minted by map (identical values reuse one placeholder), recording the placeholder→original mapping so the response can be unmasked later. findings must be sorted/merged (as scan returns). Used inbound when [llm.dlp].reversible is on; the provider sees only placeholders.