fix: bedrock guardrail handling update

plugins/bedrock/index.ts

@@ -149,34 +149,41 @@ export const pluginHandler: PluginHandler<
       )
     );
 
-    const interventionData =
-      results.find(
-        (result) => result && result.action === 'GUARDRAIL_INTERVENED'
-      ) ?? results[0];
+    const interventionData = results.find(
+      (result) => result && result.action === 'GUARDRAIL_INTERVENED'
+    );
+    if (interventionData) {
+      verdict = false;
+    }
 
     const flaggedCategories = new Set();
 
-    results.forEach((result) => {
-      if (!result) return;
-      if (result.assessments[0].contentPolicy?.filters?.length > 0) {
+    let hasTriggeredPII = false;
+    for (const result of results) {
+      if (!result) continue;
+      // adding other guardrail categories to the set, required for PII redaction check.
+      if (result.assessments[0]?.contentPolicy) {
         flaggedCategories.add('contentFilter');
       }
-      if (result.assessments[0].wordPolicy?.customWords?.length > 0) {
-        flaggedCategories.add('wordFilter');
-      }
-      if (result.assessments[0].wordPolicy?.managedWordLists?.length > 0) {
+      if (result.assessments[0]?.wordPolicy) {
         flaggedCategories.add('wordFilter');
       }
-      if (
-        result.assessments[0].sensitiveInformationPolicy?.piiEntities?.length >
-        0
-      ) {
-        flaggedCategories.add('piiFilter');
+
+      if (hasTriggeredPII) {
+        continue;
       }
-    });
 
-    let hasPii = flaggedCategories.has('piiFilter');
-    if (hasPii && redact) {
+      const sensitiveInfo = result.assessments[0]?.sensitiveInformationPolicy;
+      const sensitiveInfoKeys = Object.keys(sensitiveInfo ?? {});
+      sensitiveInfoKeys.forEach((key: string) => {
+        if ((sensitiveInfo as any)[key].length > 0) {
+          flaggedCategories.add('piiFilter');
+          hasTriggeredPII = true;
+        }
+      });
+    }
+
+    if (hasTriggeredPII && redact) {
       const maskedTexts = textArray.map((text, index) =>
         redactPii(text, results[index])
       );
@@ -185,10 +192,8 @@ export const pluginHandler: PluginHandler<
       transformed = true;
     }
 
-    if (hasPii && flaggedCategories.size === 1 && redact) {
+    if (hasTriggeredPII && flaggedCategories.size === 1 && redact) {
       verdict = true;
-    } else if (flaggedCategories.size > 0) {
-      verdict = false;
     }
     data = interventionData;
   } catch (e) {

plugins/bedrock/type.ts

@@ -79,12 +79,12 @@ export interface PIIFilter<T = any> extends BedrockAction<T> {
 export interface BedrockResponse {
   action: 'NONE' | 'GUARDRAIL_INTERVENED';
   assessments: {
-    wordPolicy: {
+    wordPolicy?: {
       customWords: WordPolicy[];
       managedWordLists: (WordPolicy & { type: 'PROFANITY' })[];
     };
-    contentPolicy: { filters: ContentPolicy[] };
-    sensitiveInformationPolicy: {
+    contentPolicy?: { filters: ContentPolicy[] };
+    sensitiveInformationPolicy?: {
       piiEntities: PIIFilter<'ANONYMIZED' | 'BLOCKED'>[];
       regexes: (Omit<PIIFilter, 'type'> & {
         name: string;

plugins/bedrock/util.ts

@@ -118,7 +118,7 @@ export const redactPii = (text: string, result: BedrockResponse | null) => {
       maskedText = data?.text;
     } else {
       // Replace the all entires of each filter sent from api.
-      result.assessments[0].sensitiveInformationPolicy.piiEntities.forEach(
+      result.assessments[0].sensitiveInformationPolicy.piiEntities?.forEach(
         (filter) => {
           maskedText = replaceMatches(filter, maskedText, false);
         }