PR 2: E01 — Neural Document Intelligence workflow + canonical benchmark run

apps/backend-services/src/seed/local-dataset-sync.service.spec.ts

@@ -47,11 +47,23 @@ function makeFakeBlobStorage(prePopulated: string[] = []): FakeBlobStore {
   return { blobs, storage, writes };
 }
 
+interface DatasetVersionUpdateData {
+  storagePrefix: string;
+  manifestPath: string;
+  documentCount: number;
+}
+
 function makePrismaStub() {
-  const updates: Array<{ where: { id: string }; data: unknown }> = [];
+  const updates: Array<{
+    where: { id: string };
+    data: DatasetVersionUpdateData;
+  }> = [];
   const client = {
     datasetVersion: {
-      update: async (args: { where: { id: string }; data: unknown }) => {
+      update: async (args: {
+        where: { id: string };
+        data: DatasetVersionUpdateData;
+      }) => {
         updates.push(args);
         return null as unknown;
       },
@@ -179,11 +191,17 @@ describe("syncLocalDatasetToBlobStorage", () => {
         ),
       ).toHaveLength(1);
 
-      // DatasetVersion updated
+      // DatasetVersion updated — manifestPath stored relative to storagePrefix
+      // so the benchmark materializer's path.join(materializedPath, manifestPath)
+      // resolves to the file that materializeDataset wrote at <materializedPath>/dataset-manifest.json.
       expect(prisma.updates).toHaveLength(1);
       expect(prisma.updates[0].where.id).toBe(
         "seed-local-samples-mix-private-v1",
       );
+      expect(prisma.updates[0].data.manifestPath).toBe("dataset-manifest.json");
+      expect(prisma.updates[0].data.storagePrefix).toBe(
+        "datasets/seed-local-samples-mix-private/seed-local-samples-mix-private-v1",
+      );
     } finally {
       cleanup();
     }

apps/backend-services/src/seed/local-dataset-sync.service.ts

@@ -237,11 +237,15 @@ export async function syncLocalDatasetToBlobStorage(
 
   // Update DatasetVersion to point at the blob storage (in case the seed
   // initially wrote a local path, or to keep documentCount in sync).
+  // manifestPath is stored relative to storagePrefix — same convention as
+  // dataset.service.ts createVersion (defaults to "dataset-manifest.json").
+  // The benchmark materializer downloads files relative to the full storage
+  // prefix and joins materializedPath + manifestPath to find the manifest.
   await prisma.datasetVersion.update({
     where: { id: versionId },
     data: {
       storagePrefix: blobStoragePrefix,
-      manifestPath: manifestKey,
+      manifestPath: "dataset-manifest.json",
       documentCount: samples.length,
     },
   });

apps/shared/prisma/seed.ts

@@ -1841,9 +1841,11 @@ async function seedLocalDatasets() {
     // Blob-convention paths. The actual files are uploaded by
     // LocalDatasetSyncService on backend startup; until then, these paths
     // are placeholders that the sync service confirms / refreshes.
+    // manifestPath is stored relative to storagePrefix — same convention as
+    // dataset.service.ts createVersion (defaults to "dataset-manifest.json").
     const datasetStoragePath = `datasets/${datasetId}`;
     const versionStoragePrefix = `datasets/${datasetId}/${versionId}`;
-    const versionManifestPath = `groups/${SEED_GROUP_ID}/benchmark/${versionStoragePrefix}/dataset-manifest.json`;
+    const versionManifestPath = "dataset-manifest.json";
 
     await prisma.dataset.upsert({
       where: { id: datasetId },
@@ -2044,9 +2046,10 @@ async function seedExperimentWorkflows() {
         name: `${workflowName} Benchmark`,
         workflowVersionId: versionId,
         workflowConfigHash: `seed-${slug}`,
-        evaluatorType: "field-accuracy",
+        evaluatorType: "schema-aware",
         evaluatorConfig: {
-          metrics: ["field_accuracy", "character_accuracy", "word_accuracy"],
+          defaultRule: { rule: "fuzzy", fuzzyThreshold: 0.85 },
+          passThreshold: 0.8,
         },
         runtimeSettings: { timeout: 600, retries: 2 },
       },
@@ -2057,9 +2060,10 @@ async function seedExperimentWorkflows() {
         datasetVersionId: targetVersion.id,
         workflowVersionId: versionId,
         workflowConfigHash: `seed-${slug}`,
-        evaluatorType: "field-accuracy",
+        evaluatorType: "schema-aware",
         evaluatorConfig: {
-          metrics: ["field_accuracy", "character_accuracy", "word_accuracy"],
+          defaultRule: { rule: "fuzzy", fuzzyThreshold: 0.85 },
+          passThreshold: 0.8,
         },
         runtimeSettings: { timeout: 600, retries: 2 },
       },

apps/temporal/package.json

@@ -8,7 +8,7 @@
 		"build:logging": "cd ../../packages/logging && npm run build",
 		"start": "node dist/worker.js",
 		"start:worker": "node dist/worker.js",
-		"dev": "ts-node-dev -r tsconfig-paths/register --respawn --transpile-only --rs --watch src --ignore-watch ../../packages src/worker.ts",
+		"dev": "ts-node-dev -r tsconfig-paths/register --respawn --transpile-only --rs --watch 'src/**/*.ts' --ignore-watch ../../packages src/worker.ts",
 		"type-check": "tsc --noEmit",
 		"clean": "rm -rf dist",
 		"db:generate": "node ../shared/scripts/generate-prisma.js",