@anvia/studio 0.7.3 → 0.7.5

package/dist/index.d.ts

@@ -158,9 +158,25 @@ type StudioEvalSuiteConfig = {
     description?: string;
     caseCount: number;
     metricNames: string[];
+    casePreviewCount?: number;
+    casePreviews?: StudioEvalCasePreview[];
+    metricSummaries?: StudioEvalMetricSummary[];
     concurrency?: number;
     metadata?: JsonObject;
 };
+type StudioEvalCasePreview = {
+    id: string;
+    input?: JsonValue;
+    expected?: JsonValue;
+    metadataKeys?: string[];
+};
+type StudioEvalMetricSummary = {
+    name: string;
+    dataType?: "NUMERIC" | "CATEGORICAL" | "BOOLEAN";
+    configId?: string;
+    scoreConfigId?: string;
+    metadataKeys?: string[];
+};
 type StudioEvalRunRequest = {
     concurrency?: number;
 };
@@ -840,4 +856,4 @@ declare function createSqliteSessionStore(options?: SqliteSessionStoreOptions):
 
 declare function createInMemoryStudioStore(): StudioSessionStore & StudioTraceStore & StudioPipelineLogStore & StudioPipelineRunStore;
 
-export { type AgentRunRequest, type AgentRunResponse, type AgentRunStreamEvent, type AgentTraceInfo, type AgentTraceOptions, type AnviaStudio, type SqliteSessionStoreOptions, Studio, type StudioAgent, type StudioAgentConfig, type StudioAgentKnowledgeConfig, type StudioAgentMcpServerMetadata, type StudioAgentMcpToolMetadata, type StudioAgentMcpsSummary, type StudioAgentModelPolicy, type StudioAgentModelPolicyConfig, type StudioAgentModelsSummary, type StudioAgentRuntimeSummary, type StudioAgentToolApprovalMetadata, type StudioAgentToolMetadata, type StudioAgentToolSource, type StudioAgentToolsSummary, type StudioCapability, type StudioCapabilityConfig, type StudioConfig, type StudioErrorCode, type StudioErrorResponse, type StudioEvalRunRequest, type StudioEvalRunResponse, type StudioEvalSuite, type StudioEvalSuiteConfig, type StudioKnowledgeEvidence, type StudioKnowledgeEvidenceDocument, type StudioKnowledgeItem, type StudioKnowledgeItemKind, type StudioKnowledgeItemsPage, type StudioKnowledgeSourceKind, type StudioKnowledgeSourceSummary, type StudioKnowledgeSummary, type StudioMemoryAppendInput, type StudioMemoryContext, type StudioMemoryConversationMessages, type StudioMemoryConversationSteps, type StudioMemoryConversationSummary, type StudioMemoryConversationsPage, type StudioMemoryErrorInput, type StudioMemoryStore, type StudioMemoryUserSummary, type StudioMemoryUsersPage, type StudioModelConfig, type StudioModelDefinition, type StudioModelModalities, type StudioModelModality, type StudioModelProvider, type StudioModelProviderConfig, type StudioModelRef, type StudioModelSummary, type StudioModelsConfig, type StudioObservabilityEvent, type StudioObservabilityEventType, type StudioOptions, type StudioPipeline, type StudioPipelineConfig, type StudioPipelineDetail, type StudioPipelineFinalEvent, type StudioPipelineLogAppendInput, type StudioPipelineLogCategory, type StudioPipelineLogEntry, type StudioPipelineLogEvent, type StudioPipelineLogLevel, type StudioPipelineLogListOptions, type StudioPipelineLogStore, type StudioPipelineReplayRequest, type StudioPipelineRunGetOptions, type StudioPipelineRunListOptions, type StudioPipelineRunRecord, type StudioPipelineRunRequest, type StudioPipelineRunResponse, type StudioPipelineRunSaveInput, type StudioPipelineRunStatus, type StudioPipelineRunStore, type StudioServeOptions, type StudioSession, type StudioSessionCreateInput, type StudioSessionListOptions, type StudioSessionLogAppendInput, type StudioSessionLogCategory, type StudioSessionLogEntry, type StudioSessionLogEvent, type StudioSessionLogLevel, type StudioSessionLogListOptions, type StudioSessionRunStatus, type StudioSessionRunTranscriptInput, type StudioSessionStore, type StudioSessionSummary, type StudioSessionTraceListOptions, type StudioStaticKnowledgeDocument, type StudioStatusSummary, type StudioStores, type StudioTarget, type StudioToolApproval, type StudioToolApprovalDecision, type StudioToolApprovalRequestEvent, type StudioToolApprovalResultEvent, type StudioToolApprovalStatus, type StudioToolApprovalTranscript, type StudioToolQuestion, type StudioToolQuestionAnswer, type StudioToolQuestionChoice, type StudioToolQuestionPrompt, type StudioToolQuestionRequestEvent, type StudioToolQuestionResultEvent, type StudioToolQuestionStatus, type StudioToolQuestionTranscript, type StudioToolRunRequest, type StudioToolRunResponse, type StudioTrace, type StudioTraceListOptions, type StudioTraceObservation, type StudioTraceObservationKind, StudioTraceObserver, type StudioTraceObserverOptions, type StudioTraceStatus, type StudioTraceStore, type StudioTraceSummary, type StudioTranscriptAttachment, type StudioTranscriptChatEntry, type StudioTranscriptChildAgentEvent, type StudioTranscriptEntry, type StudioTranscriptReasoningEntry, type StudioTranscriptToolEntry, type StudioUiOptions, createInMemoryStudioStore, createSqliteSessionStore, traceSummary };
+export { type AgentRunRequest, type AgentRunResponse, type AgentRunStreamEvent, type AgentTraceInfo, type AgentTraceOptions, type AnviaStudio, type SqliteSessionStoreOptions, Studio, type StudioAgent, type StudioAgentConfig, type StudioAgentKnowledgeConfig, type StudioAgentMcpServerMetadata, type StudioAgentMcpToolMetadata, type StudioAgentMcpsSummary, type StudioAgentModelPolicy, type StudioAgentModelPolicyConfig, type StudioAgentModelsSummary, type StudioAgentRuntimeSummary, type StudioAgentToolApprovalMetadata, type StudioAgentToolMetadata, type StudioAgentToolSource, type StudioAgentToolsSummary, type StudioCapability, type StudioCapabilityConfig, type StudioConfig, type StudioErrorCode, type StudioErrorResponse, type StudioEvalCasePreview, type StudioEvalMetricSummary, type StudioEvalRunRequest, type StudioEvalRunResponse, type StudioEvalSuite, type StudioEvalSuiteConfig, type StudioKnowledgeEvidence, type StudioKnowledgeEvidenceDocument, type StudioKnowledgeItem, type StudioKnowledgeItemKind, type StudioKnowledgeItemsPage, type StudioKnowledgeSourceKind, type StudioKnowledgeSourceSummary, type StudioKnowledgeSummary, type StudioMemoryAppendInput, type StudioMemoryContext, type StudioMemoryConversationMessages, type StudioMemoryConversationSteps, type StudioMemoryConversationSummary, type StudioMemoryConversationsPage, type StudioMemoryErrorInput, type StudioMemoryStore, type StudioMemoryUserSummary, type StudioMemoryUsersPage, type StudioModelConfig, type StudioModelDefinition, type StudioModelModalities, type StudioModelModality, type StudioModelProvider, type StudioModelProviderConfig, type StudioModelRef, type StudioModelSummary, type StudioModelsConfig, type StudioObservabilityEvent, type StudioObservabilityEventType, type StudioOptions, type StudioPipeline, type StudioPipelineConfig, type StudioPipelineDetail, type StudioPipelineFinalEvent, type StudioPipelineLogAppendInput, type StudioPipelineLogCategory, type StudioPipelineLogEntry, type StudioPipelineLogEvent, type StudioPipelineLogLevel, type StudioPipelineLogListOptions, type StudioPipelineLogStore, type StudioPipelineReplayRequest, type StudioPipelineRunGetOptions, type StudioPipelineRunListOptions, type StudioPipelineRunRecord, type StudioPipelineRunRequest, type StudioPipelineRunResponse, type StudioPipelineRunSaveInput, type StudioPipelineRunStatus, type StudioPipelineRunStore, type StudioServeOptions, type StudioSession, type StudioSessionCreateInput, type StudioSessionListOptions, type StudioSessionLogAppendInput, type StudioSessionLogCategory, type StudioSessionLogEntry, type StudioSessionLogEvent, type StudioSessionLogLevel, type StudioSessionLogListOptions, type StudioSessionRunStatus, type StudioSessionRunTranscriptInput, type StudioSessionStore, type StudioSessionSummary, type StudioSessionTraceListOptions, type StudioStaticKnowledgeDocument, type StudioStatusSummary, type StudioStores, type StudioTarget, type StudioToolApproval, type StudioToolApprovalDecision, type StudioToolApprovalRequestEvent, type StudioToolApprovalResultEvent, type StudioToolApprovalStatus, type StudioToolApprovalTranscript, type StudioToolQuestion, type StudioToolQuestionAnswer, type StudioToolQuestionChoice, type StudioToolQuestionPrompt, type StudioToolQuestionRequestEvent, type StudioToolQuestionResultEvent, type StudioToolQuestionStatus, type StudioToolQuestionTranscript, type StudioToolRunRequest, type StudioToolRunResponse, type StudioTrace, type StudioTraceListOptions, type StudioTraceObservation, type StudioTraceObservationKind, StudioTraceObserver, type StudioTraceObserverOptions, type StudioTraceStatus, type StudioTraceStore, type StudioTraceSummary, type StudioTranscriptAttachment, type StudioTranscriptChatEntry, type StudioTranscriptChildAgentEvent, type StudioTranscriptEntry, type StudioTranscriptReasoningEntry, type StudioTranscriptToolEntry, type StudioUiOptions, createInMemoryStudioStore, createSqliteSessionStore, traceSummary };

package/dist/index.js

@@ -3368,17 +3368,71 @@ function publicApproval(approval) {
 }
 
 // src/runtime/eval-config.ts
+var CASE_PREVIEW_LIMIT = 5;
+var PREVIEW_VALUE_MAX_LENGTH = 500;
 function evalConfig(suite) {
+  const casePreviews = suite.cases.slice(0, CASE_PREVIEW_LIMIT).map(casePreview);
+  const metricSummaries = suite.metrics.map(metricSummary);
   return compact({
     id: suite.id ?? suite.name,
     name: suite.name,
     description: suite.description,
     caseCount: suite.cases.length,
     metricNames: suite.metrics.map((metric) => metric.name),
+    casePreviewCount: casePreviews.length,
+    casePreviews,
+    metricSummaries,
     concurrency: suite.concurrency,
     metadata: suite.metadata
   });
 }
+function casePreview(testCase, index) {
+  if (!isRecord2(testCase)) {
+    return {
+      id: `case-${index + 1}`,
+      input: previewValue(testCase)
+    };
+  }
+  return compact({
+    id: typeof testCase.id === "string" && testCase.id.length > 0 ? testCase.id : `case-${index + 1}`,
+    input: "input" in testCase ? previewValue(testCase.input) : void 0,
+    expected: "expected" in testCase ? previewValue(testCase.expected) : void 0,
+    metadataKeys: metadataKeys(testCase.metadata)
+  });
+}
+function metricSummary(metric, index) {
+  if (!isRecord2(metric)) {
+    return { name: `metric-${index + 1}` };
+  }
+  return compact({
+    name: typeof metric.name === "string" && metric.name.length > 0 ? metric.name : `metric-${index + 1}`,
+    dataType: metricDataType(metric.dataType),
+    configId: typeof metric.configId === "string" ? metric.configId : void 0,
+    scoreConfigId: typeof metric.scoreConfigId === "string" ? metric.scoreConfigId : void 0,
+    metadataKeys: metadataKeys(metric.metadata)
+  });
+}
+function previewValue(value) {
+  const jsonValue = toJsonValue(value);
+  const serialized = JSON.stringify(jsonValue);
+  if (serialized.length <= PREVIEW_VALUE_MAX_LENGTH) {
+    return jsonValue;
+  }
+  return `${serialized.slice(0, PREVIEW_VALUE_MAX_LENGTH)}...`;
+}
+function metadataKeys(value) {
+  if (!isRecord2(value)) {
+    return void 0;
+  }
+  const keys = Object.keys(value).filter((key) => value[key] !== void 0);
+  return keys.length > 0 ? keys : void 0;
+}
+function metricDataType(value) {
+  return value === "NUMERIC" || value === "CATEGORICAL" || value === "BOOLEAN" ? value : void 0;
+}
+function isRecord2(value) {
+  return value !== null && typeof value === "object" && !Array.isArray(value);
+}
 
 // src/runtime/tool-metadata.ts
 import { ToolSet } from "@anvia/core/tool";
@@ -3765,7 +3819,7 @@ async function knowledgeItemsPage(agent, sourceId, request) {
   return void 0;
 }
 function inspectFn(index) {
-  if (!isRecord2(index) || typeof index.inspect !== "function") {
+  if (!isRecord3(index) || typeof index.inspect !== "function") {
     return void 0;
   }
   const inspect = index.inspect;
@@ -3801,7 +3855,7 @@ function nonInspectablePage(agentId, sourceId, kind) {
   };
 }
 function dynamicContextItem(item) {
-  const text = isRecord2(item.document) && typeof item.document.text === "string" ? item.document.text : typeof item.document === "string" ? item.document : void 0;
+  const text = isRecord3(item.document) && typeof item.document.text === "string" ? item.document.text : typeof item.document === "string" ? item.document : void 0;
   return {
     id: item.id,
     kind: "dynamic_context",
@@ -3810,8 +3864,8 @@ function dynamicContextItem(item) {
   };
 }
 function dynamicToolItem(item) {
-  const document = isRecord2(item.document) ? item.document : {};
-  const definition = isRecord2(document.definition) ? document.definition : {};
+  const document = isRecord3(item.document) ? item.document : {};
+  const definition = isRecord3(document.definition) ? document.definition : {};
   const toolName = typeof document.toolName === "string" ? document.toolName : typeof definition.name === "string" ? definition.name : item.id;
   const description = typeof definition.description === "string" ? definition.description : "";
   return {
@@ -3825,7 +3879,7 @@ function dynamicToolItem(item) {
   };
 }
 function parameterKeys(parameters) {
-  if (!isRecord2(parameters) || !isRecord2(parameters.properties)) {
+  if (!isRecord3(parameters) || !isRecord3(parameters.properties)) {
     return [];
   }
   return Object.keys(parameters.properties);
@@ -3866,7 +3920,7 @@ async function recentKnowledgeEvidence(traceStore, limit) {
 }
 function evidenceFromTrace(trace) {
   return trace.observations.flatMap((observation) => {
-    if (observation.kind !== "generation" || !isRecord2(observation.input)) {
+    if (observation.kind !== "generation" || !isRecord3(observation.input)) {
       return [];
     }
     const documents = Array.isArray(observation.input.documents) ? observation.input.documents.flatMap((document) => evidenceDocument(document)) : [];
@@ -3919,7 +3973,7 @@ function messageText(value) {
   if (typeof value === "string") {
     return value.trim();
   }
-  if (!isRecord2(value)) {
+  if (!isRecord3(value)) {
     return "";
   }
   if (typeof value.text === "string") {
@@ -3937,7 +3991,7 @@ function contentText(value) {
   if (typeof value === "string") {
     return value.trim();
   }
-  if (!isRecord2(value)) {
+  if (!isRecord3(value)) {
     return "";
   }
   if (typeof value.text === "string") {
@@ -3946,12 +4000,12 @@ function contentText(value) {
   return "";
 }
 function evidenceDocument(value) {
-  if (!isRecord2(value)) {
+  if (!isRecord3(value)) {
     return [];
   }
   const id = typeof value.id === "string" ? value.id : void 0;
   const text = typeof value.text === "string" ? value.text : void 0;
-  const additionalProps = isRecord2(value.additionalProps) ? jsonObjectFromRecord(value.additionalProps) : void 0;
+  const additionalProps = isRecord3(value.additionalProps) ? jsonObjectFromRecord(value.additionalProps) : void 0;
   if (id === void 0 && text === void 0 && additionalProps === void 0) {
     return [];
   }
@@ -3960,7 +4014,7 @@ function evidenceDocument(value) {
   ];
 }
 function evidenceToolName(value) {
-  if (!isRecord2(value) || typeof value.name !== "string") {
+  if (!isRecord3(value) || typeof value.name !== "string") {
     return [];
   }
   return [value.name];
@@ -3968,7 +4022,7 @@ function evidenceToolName(value) {
 function jsonObjectFromRecord(value) {
   return compactJsonObject(value);
 }
-function isRecord2(value) {
+function isRecord3(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }