Guardrail fixes enhancements

src/prerna/engine/impl/guardrail/AggressiveSelfHarmGuardrailEngine.java

@@ -46,6 +46,7 @@
 import prerna.engine.impl.model.message.ResponseMessage;
 import prerna.om.Insight;
 import prerna.om.InsightStore;
+import prerna.om.ThreadStore;
 import prerna.sablecc2.om.GenRowStruct;
 import prerna.sablecc2.om.NounStore;
 import prerna.sablecc2.om.nounmeta.GuardrailNounMetadata;
@@ -63,15 +64,18 @@
  * Optional SMSS keys:
  * <ul>
  *   <li>{@code SYSTEM_PROMPT} — override the default classification instruction</li>
+ *   <li>{@code BLOCKED_MESSAGE} — override the default message returned when UNSAFE</li>
  * </ul>
  */
 public class AggressiveSelfHarmGuardrailEngine extends AbstractGuardrailReactorFunctionEngine {
 
 	private static final Logger classLogger = LogManager.getLogger(AggressiveSelfHarmGuardrailEngine.class);
 
 	public static final String MODEL_ENGINE_ID_KEY = "MODEL_ENGINE_ID";
+	public static final String SYSTEM_PROMPT_KEY = "SYSTEM_PROMPT";
+	public static final String BLOCKED_MESSAGE_KEY = "BLOCKED_MESSAGE";
 
-	private static final String SYSTEM_PROMPT =
+	private static final String DEFAULT_SYSTEM_PROMPT =
 			"You are a content safety classifier. Your sole task is to detect whether a user message "
 			+ "expresses first-person intent to harm themselves or others, suicidal ideation, or aggressive threats.\n\n"
 			+ "Respond with ONLY one word:\n"
@@ -82,7 +86,13 @@ public class AggressiveSelfHarmGuardrailEngine extends AbstractGuardrailReactorF
 			+ "third-party reports, or general medical questions\n\n"
 			+ "Do not explain your reasoning. Output only SAFE or UNSAFE.";
 
+	private static final String DEFAULT_BLOCKED_MESSAGE =
+			"I'm sorry, I'm not able to help with that. If you or someone you know is in crisis, "
+			+ "please call or text 988 to reach the Suicide & Crisis Lifeline, or call 911 for immediate emergency assistance.";
+
 	private String modelEngineId;
+	private String systemPrompt;
+	private String blockedMessage;
 
 	public AggressiveSelfHarmGuardrailEngine() {
 		this.keysToGet = new String[] { "prompt" };
@@ -98,6 +108,14 @@ public void open(Properties smssProp) throws Exception {
 		}
 		this.modelEngineId = this.modelEngineId.trim();
 
+		String systemPromptProp = this.smssProp.getProperty(SYSTEM_PROMPT_KEY);
+		this.systemPrompt = (systemPromptProp != null && !systemPromptProp.trim().isEmpty())
+				? systemPromptProp.trim() : DEFAULT_SYSTEM_PROMPT;
+
+		String blockedMessageProp = this.smssProp.getProperty(BLOCKED_MESSAGE_KEY);
+		this.blockedMessage = (blockedMessageProp != null && !blockedMessageProp.trim().isEmpty())
+				? blockedMessageProp.trim() : DEFAULT_BLOCKED_MESSAGE;
+
 		this.functionDescription = "Detects aggressive or self-harm content by asking a configured LLM to classify "
 				+ "the prompt as SAFE or UNSAFE.";
 		this.parameters = new ArrayList<>();
@@ -109,6 +127,9 @@ public void open(Properties smssProp) throws Exception {
 	public GuardrailNounMetadata execute(NounStore ns, GenRowStruct curRow) {
 		Map<String, String> keyValue = organizeKeys(ns, curRow);
 		String prompt = keyValue.get("prompt");
+		if (prompt == null) {
+			throw new IllegalArgumentException("No prompt has been defined");
+		}
 
 		classLogger.info("AggressiveSelfHarmGuardrail: classifying prompt (length={}) via model={}",
 				prompt.length(), this.modelEngineId);
@@ -121,13 +142,16 @@ public GuardrailNounMetadata execute(NounStore ns, GenRowStruct curRow) {
 		Insight classificationInsight = new Insight();
 		InsightStore.getInstance().put(classificationInsight);
 
+		String savedJobId = ThreadStore.getJobId();
+		ThreadStore.setJobId(null);
+
 		String classification;
 		try {
 			Room room = RoomUtils.createRoomIfNotExists(UUID.randomUUID().toString(), classificationInsight, modelEngine, prompt);
 			Map<String, Object> params = new HashMap<>();
 			params.put("use_history", false);
 			InputMessage msg = InputMessage.builder(room)
-					.withSystemPrompt(SYSTEM_PROMPT)
+					.withSystemPrompt(this.systemPrompt)
 					.withText(prompt)
 					.withModelType(modelEngine.getModelType())
 					.withParamMap(params)
@@ -136,6 +160,7 @@ public GuardrailNounMetadata execute(NounStore ns, GenRowStruct curRow) {
 			Object responseObj = response.getModelEngineResponse().toMap().get("response");
 			classification = responseObj != null ? responseObj.toString().trim() : "";
 		} finally {
+			ThreadStore.setJobId(savedJobId);
 			InsightStore.getInstance().remove(classificationInsight.getInsightId());
 		}
 
@@ -147,11 +172,7 @@ public GuardrailNounMetadata execute(NounStore ns, GenRowStruct curRow) {
 		details.put("classification", classification);
 		details.put("modelEngineId", this.modelEngineId);
 
-		String returnPrompt = pass ? prompt
-				: "Respond only with the following message, word for word: "
-				+ "'I'm sorry, I'm not able to help with that. "
-				+ "If you or someone you know is in crisis, please call or text 988 to reach the Veterans Crisis Line, "
-				+ "or call 911 for immediate emergency assistance.'";
+		String returnPrompt = pass ? prompt : this.blockedMessage;
 
 		return new GuardrailNounMetadata(pass, returnPrompt, details);
 	}

src/prerna/engine/impl/guardrail/PolicyComplianceGuardrailEngine.java

@@ -47,6 +47,7 @@
 import prerna.engine.impl.model.responses.AbstractModelEngineResponse;
 import prerna.om.Insight;
 import prerna.om.InsightStore;
+import prerna.om.ThreadStore;
 import prerna.sablecc2.om.GenRowStruct;
 import prerna.sablecc2.om.NounStore;
 import prerna.sablecc2.om.nounmeta.GuardrailNounMetadata;
@@ -181,6 +182,8 @@ private String classify(String textToJudge, String policyDescription) throws Exc
 		// Room/model calls require an Insight to be registered, this one is never tied to a real user session.
 		Insight classificationInsight = new Insight();
 		InsightStore.getInstance().put(classificationInsight);
+		String savedJobId = ThreadStore.getJobId();
+		ThreadStore.setJobId(null);
 		try {
 			Room room = RoomUtils.createRoomIfNotExists(UUID.randomUUID().toString(), classificationInsight,
 					judgeEngine, textToJudge);
@@ -196,6 +199,7 @@ private String classify(String textToJudge, String policyDescription) throws Exc
 			Object responseObj = response.getModelEngineResponse().toMap().get("response");
 			return responseObj != null ? responseObj.toString().trim() : "";
 		} finally {
+			ThreadStore.setJobId(savedJobId);
 			InsightStore.getInstance().remove(classificationInsight.getInsightId());
 		}
 	}

src/prerna/engine/impl/pipeline/PipelineInvocationHandler.java

@@ -68,6 +68,8 @@
 import prerna.engine.api.IEngine;
 import prerna.engine.impl.model.Room;
 import prerna.engine.impl.model.responses.AbstractModelEngineResponse;
+import prerna.engine.impl.model.responses.AskModelEngineResponse;
+import prerna.engine.impl.model.responses.AskStringModelEngineResponse;
 import prerna.logging.IgnoreEngineLogging;
 import prerna.logging.LoggingEngineSerializer;
 import prerna.logging.LoggingIReactorSerializer;
@@ -129,6 +131,7 @@ public class PipelineInvocationHandler implements InvocationHandler {
 	// took
 	private static final String GUARDRAIL_ACTION_MASK = "MASK";
 	private static final String GUARDRAIL_ACTION_BLOCK = "BLOCK";
+	private static final String GUARDRAIL_ACTION_RESPOND = "RESPOND";
 
 	private final ZoneId UTC_ZONE_ID = ZoneId.of("UTC");
 	private final Map<String, Pipeline> pipelinesMap = new HashMap<>();
@@ -315,9 +318,12 @@ public Object invoke(Object proxy, Method method, Object[] args) throws Throwabl
 							.get(PipelineReactorUtils.INTERIM_RESULT);
 					boolean pass = (boolean) resultMap.get(PipelineReactorUtils.PASS);
 					boolean masked = Boolean.TRUE.equals(resultMap.get(PipelineReactorUtils.MASKED));
+					String cannedResponse = (String) resultMap.get(PipelineReactorUtils.SHORT_CIRCUIT_RESPONSE);
 					// MASK when the guardrail neutralized content, BLOCK when it stopped the
-					// request, null when it ran clean - queryable via the GUARDRAIL_ACTION column
-					String guardrailAction = masked ? GUARDRAIL_ACTION_MASK : (!pass ? GUARDRAIL_ACTION_BLOCK : null);
+					// request, RESPOND when it supplied the answer itself, null when it ran clean
+					// - queryable via the GUARDRAIL_ACTION column
+					String guardrailAction = cannedResponse != null ? GUARDRAIL_ACTION_RESPOND
+							: masked ? GUARDRAIL_ACTION_MASK : (!pass ? GUARDRAIL_ACTION_BLOCK : null);
 
 					String request = null;
 					String response = null;
@@ -334,6 +340,16 @@ public Object invoke(Object proxy, Method method, Object[] args) throws Throwabl
 					logEngineCall(engineSpecificLogger, start, end, pass, request, response,
 							reactor.getClass().getSimpleName(), null, null, null, guardrailAction);
 
+					if (cannedResponse != null) {
+						if (!AskModelEngineResponse.class.isAssignableFrom(method.getReturnType())) {
+							throw new SemossPixelException(
+									"Unable to process this request due to content policy (guardrail input exception)");
+						}
+						classLogger.warn("Guardrail {} short-circuited the model call with a canned response",
+								reactor.getClass().getSimpleName());
+						return new AskStringModelEngineResponse(cannedResponse, 0, 0);
+					}
+
 					if (!pass) {
 						throw new SemossPixelException(
 								"Unable to process this request due to content policy (guardrail input exception)");

src/prerna/reactor/interceptor/GenericGuardrailInputReactor.java

@@ -191,7 +191,17 @@ public NounMetadata execute() {
 			}
 		}
 
-		Map<String, Object> resultMap = createInterimResult(output, this.getClass().getName(), masked);
+		// When configured to respond (rather than mask or block), hand the guardrail's
+		// message back as the model's answer. The real model call is skipped entirely,
+		// so no version of the prompt reaches the provider.
+		String cannedResponse = null;
+		Boolean respondWithGuardrailMessage = helper.getConfigParameter("respondWithGuardrailMessage", Boolean.class);
+		if (Boolean.TRUE.equals(respondWithGuardrailMessage) && !output.isPass()) {
+			cannedResponse = output.getReturnPrompt();
+		}
+
+		Map<String, Object> resultMap = createInterimResult(output, this.getClass().getName(), masked,
+				cannedResponse);
 
 		// Update the processedArguments with the interim result
 		processedArguments.put(PipelineReactorUtils.INTERIM_RESULT, resultMap);
@@ -206,13 +216,16 @@ public NounMetadata execute() {
 	 * @return
 	 */
 	private Map<String, Object> createInterimResult(GuardrailNounMetadata results, String interceptorName,
-			boolean masked) {
+			boolean masked, String cannedResponse) {
 		Map<String, Object> resultMap = new HashMap<>();
 		resultMap.put(PipelineReactorUtils.INTERCEPTOR, interceptorName);
 		// when we masked the input we neutralized the failure, so let it pass downstream
 		resultMap.put(PipelineReactorUtils.PASS, masked || results.isPass());
 		resultMap.put(PipelineReactorUtils.PASS_DETAILS, results.getValue());
 		resultMap.put(PipelineReactorUtils.MASKED, masked);
+		if (cannedResponse != null) {
+			resultMap.put(PipelineReactorUtils.SHORT_CIRCUIT_RESPONSE, cannedResponse);
+		}
 
 		return resultMap;
 	}

src/prerna/reactor/interceptor/PipelineReactorUtils.java

@@ -48,6 +48,7 @@ public final class PipelineReactorUtils {
 	public static final String PASS = "pass";
 	public static final String PASS_DETAILS = "passDetails";
 	public static final String MASKED = "masked";
+	public static final String SHORT_CIRCUIT_RESPONSE = "shortCircuitResponse";
 
 	private PipelineReactorUtils() {
 		// private constructor to prevent instantiation