@vellumai/assistant 0.10.4-dev.202607010541.092b857 → 0.10.4-dev.202607010743.70cd9f0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.10.4-dev.202607010541.092b857",
+  "version": "0.10.4-dev.202607010743.70cd9f0",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/plugin-import-boundary-guard.test.ts

@@ -187,6 +187,7 @@ const BASELINE: Record<string, readonly string[]> = {
     "../../../persistence/embeddings/messages-lexical-index.js",
     "../../../persistence/embeddings/qdrant-client.js",
     "../../../persistence/embeddings/qdrant-manager.js",
+    "../../../persistence/job-utils.js",
     "../../../persistence/jobs-store.js",
     "../../../persistence/jobs-worker.js",
     "../../../persistence/memory-lifecycle-hooks.js",

package/src/plugins/defaults/memory/embeddings.ts

@@ -0,0 +1,77 @@
+import { getConfig } from "../../../config/loader.js";
+import type { AssistantConfig } from "../../../config/types.js";
+import type {
+  EmbeddingInput,
+  EmbeddingRequestOptions,
+} from "../../../persistence/embeddings/embedding-types.js";
+import { resolveQdrantUrl as resolveQdrantUrlWithConfig } from "../../../persistence/embeddings/qdrant-client.js";
+
+/**
+ * Memory-internal accessor for the shared embeddings / vector subsystem.
+ *
+ * The embeddings operations live in `persistence/embeddings` (and
+ * `persistence/job-utils`) and take the full `AssistantConfig` as their first
+ * argument — they read both the `memory` slice (Qdrant collection and vector
+ * settings) and the `llm` slice (embedding-backend selection). Memory code
+ * reaches those operations through this accessor: the self-contained ones
+ * (`embedAndUpsert`, `selectedBackendSupportsMultimodal`, `resolveQdrantUrl`)
+ * resolve the live config internally, so their call sites need not hold it.
+ * `embedWithBackend` keeps its `config` parameter — it is a primitive whose
+ * vectors the caller stores alongside its own config-derived metadata (cache
+ * keys, vector-size checks, Qdrant collection), so it must embed with the
+ * caller's exact config snapshot rather than a re-read that could diverge from
+ * that metadata mid-operation.
+ *
+ * The embed operations are loaded via dynamic `import()` inside each wrapper so
+ * that importing this module for a single operation does not eagerly pull the
+ * whole embed/vector import graph (`job-utils`, `embedding-backend`) into the
+ * consumer. An eager pull would force every one of those modules' named exports
+ * to resolve at instantiation, which breaks the intentional partial module
+ * mocks in memory tests (a store that mocks only the exports it uses). Only the
+ * synchronous `resolveQdrantUrl` is imported statically.
+ */
+
+type EmbeddingTargetType = Parameters<
+  typeof import("../../../persistence/job-utils.js").embedAndUpsert
+>[1];
+
+/** Embed a target and upsert its vector to Qdrant. */
+export async function embedAndUpsert(
+  targetType: EmbeddingTargetType,
+  targetId: string,
+  input: EmbeddingInput,
+  extraPayload?: Record<string, unknown>,
+): Promise<void> {
+  const { embedAndUpsert: withConfig } =
+    await import("../../../persistence/job-utils.js");
+  return withConfig(getConfig(), targetType, targetId, input, extraPayload);
+}
+
+/** Embed one or more inputs via the selected embedding backend. */
+export async function embedWithBackend(
+  config: AssistantConfig,
+  inputs: EmbeddingInput[],
+  options?: EmbeddingRequestOptions,
+): Promise<
+  Awaited<
+    ReturnType<
+      typeof import("../../../persistence/embeddings/embedding-backend.js").embedWithBackend
+    >
+  >
+> {
+  const { embedWithBackend: withConfig } =
+    await import("../../../persistence/embeddings/embedding-backend.js");
+  return withConfig(config, inputs, options);
+}
+
+/** Whether the active embedding backend handles multimodal inputs. */
+export async function selectedBackendSupportsMultimodal(): Promise<boolean> {
+  const { selectedBackendSupportsMultimodal: withConfig } =
+    await import("../../../persistence/embeddings/embedding-backend.js");
+  return withConfig(getConfig());
+}
+
+/** Resolve the Qdrant base URL for this process. */
+export function resolveQdrantUrl(): string {
+  return resolveQdrantUrlWithConfig(getConfig());
+}

package/src/plugins/defaults/memory/graph/bootstrap.ts

@@ -21,10 +21,7 @@ import {
   setMemoryCheckpoint,
 } from "../../../../persistence/checkpoints.js";
 import { getDb } from "../../../../persistence/db-connection.js";
-import {
-  initQdrantClient,
-  resolveQdrantUrl,
-} from "../../../../persistence/embeddings/qdrant-client.js";
+import { initQdrantClient } from "../../../../persistence/embeddings/qdrant-client.js";
 import {
   enqueueMemoryJob,
   hasActiveJobOfType,
@@ -38,6 +35,7 @@ import {
 } from "../../../../persistence/schema/index.js";
 import { getLogger } from "../../../../util/logger.js";
 import { getWorkspaceDir } from "../../../../util/platform.js";
+import { resolveQdrantUrl } from "../embeddings.js";
 import { runGraphExtraction } from "./extraction.js";
 import { countNodes } from "./store.js";
 
@@ -86,7 +84,7 @@ export async function bootstrapFromHistory(
   // Initialize Qdrant client for inline embedding
   try {
     initQdrantClient({
-      url: resolveQdrantUrl(config),
+      url: resolveQdrantUrl(),
       collection: config.memory.qdrant.collection,
       vectorSize: config.memory.qdrant.vectorSize,
       onDisk: config.memory.qdrant.onDisk ?? true,

package/src/plugins/defaults/memory/graph/extraction.ts

@@ -1157,7 +1157,7 @@ export async function runGraphExtraction(
     const { embedGraphNodeDirect } = await import("./graph-search.js");
     for (const node of createdNodes) {
       try {
-        await embedGraphNodeDirect(node, config);
+        await embedGraphNodeDirect(node);
       } catch (err) {
         const msg = err instanceof Error ? err.message : String(err);
         log.warn(

package/src/plugins/defaults/memory/graph/graph-search.ts

@@ -4,7 +4,6 @@
 
 import { getConfig } from "../../../../config/loader.js";
 import type { AssistantConfig } from "../../../../config/types.js";
-import { selectedBackendSupportsMultimodal } from "../../../../persistence/embeddings/embedding-backend.js";
 import type { EmbeddingInput } from "../../../../persistence/embeddings/embedding-types.js";
 import { isQdrantBreakerOpen } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
@@ -13,7 +12,6 @@ import {
   type QdrantSearchResult,
   type QdrantSparseVector,
 } from "../../../../persistence/embeddings/qdrant-client.js";
-import { embedAndUpsert } from "../../../../persistence/job-utils.js";
 import { asString } from "../../../../persistence/job-utils.js";
 import {
   enqueueMemoryJob,
@@ -21,6 +19,10 @@ import {
   type MemoryJob,
 } from "../../../../persistence/jobs-store.js";
 import { getLogger } from "../../../../util/logger.js";
+import {
+  embedAndUpsert,
+  selectedBackendSupportsMultimodal,
+} from "../embeddings.js";
 import { loadImageRefData } from "./image-ref-utils.js";
 import { getNode } from "./store.js";
 import type { MemoryNode } from "./types.js";
@@ -165,10 +167,7 @@ function formatNodeForEmbedding(node: MemoryNode): string {
  * (text queries match image memories in the same vector space). Falls back
  * to text embedding with image description suffixes otherwise.
  */
-export async function embedGraphNodeDirect(
-  node: MemoryNode,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedGraphNodeDirect(node: MemoryNode): Promise<void> {
   if (node.fidelity === "gone") return;
 
   const text = formatNodeForEmbedding(node);
@@ -181,7 +180,7 @@ export async function embedGraphNodeDirect(
   };
 
   if (node.imageRefs && node.imageRefs.length > 0) {
-    const multimodalAvailable = await selectedBackendSupportsMultimodal(config);
+    const multimodalAvailable = await selectedBackendSupportsMultimodal();
     if (multimodalAvailable) {
       const imageData = await loadImageRefData(node.imageRefs[0]);
       if (imageData) {
@@ -191,7 +190,7 @@ export async function embedGraphNodeDirect(
             data: imageData.data,
             mimeType: imageData.mimeType,
           };
-          await embedAndUpsert(config, "graph_node", node.id, input, {
+          await embedAndUpsert("graph_node", node.id, input, {
             ...extraPayload,
             has_image: true,
           });
@@ -209,33 +208,24 @@ export async function embedGraphNodeDirect(
     // Fallback: text embedding with image description suffix
     const descSuffix = node.imageRefs.map((r) => r.description).join("; ");
     const textWithImages = `${text}\n[images: ${descSuffix}]`;
-    await embedAndUpsert(
-      config,
-      "graph_node",
-      node.id,
-      textWithImages,
-      extraPayload,
-    );
+    await embedAndUpsert("graph_node", node.id, textWithImages, extraPayload);
     return;
   }
 
-  await embedAndUpsert(config, "graph_node", node.id, text, extraPayload);
+  await embedAndUpsert("graph_node", node.id, text, extraPayload);
 }
 
 /**
  * Job handler: embed a graph node and upsert to Qdrant.
  */
-export async function embedGraphNodeJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedGraphNodeJob(job: MemoryJob): Promise<void> {
   const nodeId = asString(job.payload.nodeId);
   if (!nodeId) return;
 
   const node = getNode(nodeId);
   if (!node) return;
 
-  await embedGraphNodeDirect(node, config);
+  await embedGraphNodeDirect(node);
 }
 
 /**
@@ -262,8 +252,7 @@ export async function embedGraphTriggerJob(
   const { eq } = await import("drizzle-orm");
   const { memoryGraphTriggers } =
     await import("../../../../persistence/schema/index.js");
-  const { embedWithBackend } =
-    await import("../../../../persistence/embeddings/embedding-backend.js");
+  const { embedWithBackend } = await import("../embeddings.js");
 
   const db = getDb();
   const row = db

package/src/plugins/defaults/memory/graph/retriever.ts

@@ -11,12 +11,10 @@ import { getConfiguredProvider } from "@vellumai/plugin-api";
 
 import type { AssistantConfig } from "../../../../config/types.js";
 import { embedWithRetry } from "../../../../persistence/embeddings/embed.js";
-import {
-  generateSparseEmbedding,
-  selectedBackendSupportsMultimodal,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { generateSparseEmbedding } from "../../../../persistence/embeddings/embedding-backend.js";
 import type { QdrantSparseVector } from "../../../../persistence/embeddings/qdrant-client.js";
 import { getLogger } from "../../../../util/logger.js";
+import { selectedBackendSupportsMultimodal } from "../embeddings.js";
 import { extractToolUse, userMessage } from "../llm-helpers.js";
 import { searchGraphNodes } from "./graph-search.js";
 import type { InContextTracker } from "./injection.js";
@@ -973,7 +971,7 @@ export async function retrieveForTurn(
 
   if (imageBlocks.length > 0) {
     try {
-      const isMultimodal = await selectedBackendSupportsMultimodal(opts.config);
+      const isMultimodal = await selectedBackendSupportsMultimodal();
       if (isMultimodal) {
         const maxImageQueries = 2;
         for (

package/src/plugins/defaults/memory/indexer.ts

@@ -9,7 +9,6 @@ import {
   setMemoryCheckpoint,
 } from "../../../persistence/checkpoints.js";
 import { getDb } from "../../../persistence/db-connection.js";
-import { selectedBackendSupportsMultimodal } from "../../../persistence/embeddings/embedding-backend.js";
 import {
   enqueueMemoryJob,
   isMemoryEnabled,
@@ -24,6 +23,7 @@ import type { TrustClass } from "../../../runtime/actor-trust-resolver.js";
 import { enqueueAutoAnalysisIfEnabled } from "../../../runtime/services/auto-analysis-enqueue.js";
 import { isAutoAnalysisConversation } from "../../../runtime/services/auto-analysis-guard.js";
 import { getLogger } from "../../../util/logger.js";
+import { selectedBackendSupportsMultimodal } from "./embeddings.js";
 import { isMemoryRetrospectiveConversation } from "./memory-retrospective-enqueue.js";
 import { maybeEnqueueRetrospective } from "./memory-retrospective-trigger-check.js";
 import { segmentText } from "./segmenter.js";
@@ -93,8 +93,7 @@ export async function indexMessageNow(
     (b) => b.type === "image",
   );
   const mediaBlocks =
-    candidateMediaMeta.length > 0 &&
-    (await selectedBackendSupportsMultimodal(getConfig()))
+    candidateMediaMeta.length > 0 && (await selectedBackendSupportsMultimodal())
       ? candidateMediaMeta
       : [];
 

package/src/plugins/defaults/memory/job-handlers/embedding.test.ts

@@ -41,25 +41,12 @@ mock.module("../../../../persistence/job-utils.js", () => ({
 }));
 
 import { resetDbForTesting } from "../../../../__tests__/db-test-helpers.js";
-import { DEFAULT_CONFIG } from "../../../../config/defaults.js";
-import type { AssistantConfig } from "../../../../config/types.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import { initializeDb } from "../../../../persistence/db-init.js";
 import type { MemoryJob } from "../../../../persistence/jobs-store.js";
 import { mediaAssets } from "../../../../persistence/schema/index.js";
 import { embedMediaJob } from "./embedding.js";
 
-const TEST_CONFIG: AssistantConfig = {
-  ...DEFAULT_CONFIG,
-  memory: {
-    ...DEFAULT_CONFIG.memory,
-    extraction: {
-      ...DEFAULT_CONFIG.memory.extraction,
-      useLLM: false,
-    },
-  },
-};
-
 describe("embedMediaJob", () => {
   // initializeDb runs the full migration chain (hundreds of steps); under
   // parallel CI load it can exceed bun's default 5s hook timeout, so allow more.
@@ -90,12 +77,12 @@ describe("embedMediaJob", () => {
   }
 
   test("skips when assetId is missing", async () => {
-    await embedMediaJob(makeJob({}), TEST_CONFIG);
+    await embedMediaJob(makeJob({}));
     expect(embedAndUpsertCalls).toHaveLength(0);
   });
 
   test("skips when asset is not found", async () => {
-    await embedMediaJob(makeJob({ assetId: "nonexistent" }), TEST_CONFIG);
+    await embedMediaJob(makeJob({ assetId: "nonexistent" }));
     expect(embedAndUpsertCalls).toHaveLength(0);
   });
 
@@ -122,7 +109,7 @@ describe("embedMediaJob", () => {
       })
       .run();
 
-    await embedMediaJob(makeJob({ assetId: "asset-registered" }), TEST_CONFIG);
+    await embedMediaJob(makeJob({ assetId: "asset-registered" }));
     expect(embedAndUpsertCalls).toHaveLength(0);
   });
 
@@ -149,7 +136,7 @@ describe("embedMediaJob", () => {
       })
       .run();
 
-    await embedMediaJob(makeJob({ assetId: "asset-image" }), TEST_CONFIG);
+    await embedMediaJob(makeJob({ assetId: "asset-image" }));
 
     expect(embedAndUpsertCalls).toHaveLength(1);
     const call = embedAndUpsertCalls[0];
@@ -191,7 +178,7 @@ describe("embedMediaJob", () => {
       })
       .run();
 
-    await embedMediaJob(makeJob({ assetId: "asset-audio" }), TEST_CONFIG);
+    await embedMediaJob(makeJob({ assetId: "asset-audio" }));
 
     expect(embedAndUpsertCalls).toHaveLength(1);
     const call = embedAndUpsertCalls[0];
@@ -224,7 +211,7 @@ describe("embedMediaJob", () => {
       })
       .run();
 
-    await embedMediaJob(makeJob({ assetId: "asset-video" }), TEST_CONFIG);
+    await embedMediaJob(makeJob({ assetId: "asset-video" }));
 
     expect(embedAndUpsertCalls).toHaveLength(1);
     const call = embedAndUpsertCalls[0];

package/src/plugins/defaults/memory/job-handlers/embedding.ts

@@ -2,10 +2,9 @@ import { readFile } from "node:fs/promises";
 
 import { eq } from "drizzle-orm";
 
-import type { AssistantConfig } from "../../../../config/types.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import type { EmbeddingInput } from "../../../../persistence/embeddings/embedding-types.js";
-import { asString, embedAndUpsert } from "../../../../persistence/job-utils.js";
+import { asString } from "../../../../persistence/job-utils.js";
 import type { MemoryJob } from "../../../../persistence/jobs-store.js";
 import { extractMediaBlocks } from "../../../../persistence/message-content.js";
 import {
@@ -14,11 +13,9 @@ import {
   memorySummaries,
   messages,
 } from "../../../../persistence/schema/index.js";
+import { embedAndUpsert } from "../embeddings.js";
 
-export async function embedSegmentJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedSegmentJob(job: MemoryJob): Promise<void> {
   const segmentId = asString(job.payload.segmentId);
   if (!segmentId) return;
   const db = getDb();
@@ -28,7 +25,7 @@ export async function embedSegmentJob(
     .where(eq(memorySegments.id, segmentId))
     .get();
   if (!segment) return;
-  await embedAndUpsert(config, "segment", segment.id, segment.text, {
+  await embedAndUpsert("segment", segment.id, segment.text, {
     conversation_id: segment.conversationId,
     message_id: segment.messageId,
     created_at: segment.createdAt,
@@ -36,10 +33,7 @@ export async function embedSegmentJob(
   });
 }
 
-export async function embedSummaryJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedSummaryJob(job: MemoryJob): Promise<void> {
   const summaryId = asString(job.payload.summaryId);
   if (!summaryId) return;
   const db = getDb();
@@ -50,7 +44,6 @@ export async function embedSummaryJob(
     .get();
   if (!summary) return;
   await embedAndUpsert(
-    config,
     "summary",
     summary.id,
     `[${summary.scope}] ${summary.summary}`,
@@ -63,10 +56,7 @@ export async function embedSummaryJob(
   );
 }
 
-export async function embedMediaJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedMediaJob(job: MemoryJob): Promise<void> {
   const assetId = asString(job.payload.assetId);
   if (!assetId) return;
 
@@ -88,7 +78,7 @@ export async function embedMediaJob(
     mimeType: asset.mimeType,
   };
 
-  await embedAndUpsert(config, "media", asset.id, input, {
+  await embedAndUpsert("media", asset.id, input, {
     created_at: asset.createdAt,
     kind: asset.mediaType,
     subject: asset.title,
@@ -96,10 +86,7 @@ export async function embedMediaJob(
   });
 }
 
-export async function embedAttachmentJob(
-  job: MemoryJob,
-  config: AssistantConfig,
-): Promise<void> {
+export async function embedAttachmentJob(job: MemoryJob): Promise<void> {
   const messageId = asString(job.payload.messageId);
   const blockIndex = job.payload.blockIndex as number;
   if (!messageId || typeof blockIndex !== "number") return;
@@ -124,7 +111,7 @@ export async function embedAttachmentJob(
 
   // Use messageId + blockIndex as targetId for uniqueness
   const targetId = `${messageId}:${blockIndex}`;
-  await embedAndUpsert(config, "media", targetId, input, {
+  await embedAndUpsert("media", targetId, input, {
     created_at: message.createdAt,
     message_id: messageId,
     conversation_id: message.conversationId,

package/src/plugins/defaults/memory/job-handlers/index-maintenance.ts

@@ -1,8 +1,6 @@
 import { eq, isNotNull, like, ne } from "drizzle-orm";
 
-import { getConfig } from "../../../../config/loader.js";
 import { getDb } from "../../../../persistence/db-connection.js";
-import { selectedBackendSupportsMultimodal } from "../../../../persistence/embeddings/embedding-backend.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
 import { getQdrantClient } from "../../../../persistence/embeddings/qdrant-client.js";
 import {
@@ -24,6 +22,7 @@ import {
   messages,
 } from "../../../../persistence/schema/index.js";
 import { getLogger } from "../../../../util/logger.js";
+import { selectedBackendSupportsMultimodal } from "../embeddings.js";
 
 const log = getLogger("memory-jobs-worker");
 
@@ -50,7 +49,7 @@ export async function rebuildIndexJob(): Promise<void> {
   // Re-enqueue multimodal embedding jobs only when the resolved embedding
   // backend supports multimodal inputs. Without this gate, embed_media and
   // embed_attachment jobs would all fail for text-only backends.
-  if (await selectedBackendSupportsMultimodal(getConfig())) {
+  if (await selectedBackendSupportsMultimodal()) {
     const assets = db
       .select({ id: mediaAssets.id })
       .from(mediaAssets)

package/src/plugins/defaults/memory/job-handlers/index-message-lexical.ts

@@ -11,7 +11,6 @@ import {
   type MessagesLexicalIndex,
 } from "../../../../persistence/embeddings/messages-lexical-index.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
-import { resolveQdrantUrl } from "../../../../persistence/embeddings/qdrant-client.js";
 import { asString } from "../../../../persistence/job-utils.js";
 import {
   enqueueMemoryJob,
@@ -21,6 +20,7 @@ import {
 import { messages } from "../../../../persistence/schema/index.js";
 import { getLogger } from "../../../../util/logger.js";
 import { isPluginDisabled } from "../../../disabled-state.js";
+import { resolveQdrantUrl } from "../embeddings.js";
 import memoryPkg from "../package.json" with { type: "json" };
 
 const log = getLogger("messages-lexical-enqueue");
@@ -62,7 +62,7 @@ export function resolveLexicalIndex(
     return getMessagesLexicalIndex();
   } catch {
     return initMessagesLexicalIndex({
-      url: resolveQdrantUrl(config),
+      url: resolveQdrantUrl(),
       collection: MESSAGES_LEXICAL_COLLECTION,
       onDisk: config.memory.qdrant.onDisk,
     });

package/src/plugins/defaults/memory/job-handlers.ts

@@ -122,11 +122,11 @@ async function graphNarrativeRefineJob(
 export const memoryJobHandlers: readonly JobHandlerEntry[] = [
   {
     type: "embed_segment",
-    handler: (job, config) => embedSegmentJob(job, config),
+    handler: (job) => embedSegmentJob(job),
   },
   {
     type: "embed_summary",
-    handler: (job, config) => embedSummaryJob(job, config),
+    handler: (job) => embedSummaryJob(job),
   },
   { type: "backfill", handler: (job, config) => backfillJob(job, config) },
   { type: "rebuild_index", handler: () => rebuildIndexJob() },
@@ -134,14 +134,14 @@ export const memoryJobHandlers: readonly JobHandlerEntry[] = [
     type: "delete_qdrant_vectors",
     handler: (job) => deleteQdrantVectorsJob(job),
   },
-  { type: "embed_media", handler: (job, config) => embedMediaJob(job, config) },
+  { type: "embed_media", handler: (job) => embedMediaJob(job) },
   {
     type: "embed_attachment",
-    handler: (job, config) => embedAttachmentJob(job, config),
+    handler: (job) => embedAttachmentJob(job),
   },
   {
     type: "embed_graph_node",
-    handler: (job, config) => embedGraphNodeJob(job, config),
+    handler: (job) => embedGraphNodeJob(job),
   },
   {
     type: "embed_pkb_file",

package/src/plugins/defaults/memory/jobs/embed-concept-page.ts

@@ -23,7 +23,6 @@ import { and, eq } from "drizzle-orm";
 import type { AssistantConfig } from "../../../../config/types.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import {
-  embedWithBackend,
   generateSparseEmbedding,
   getMemoryBackendStatus,
 } from "../../../../persistence/embeddings/embedding-backend.js";
@@ -43,6 +42,7 @@ import { BackendUnavailableError } from "../../../../util/errors.js";
 import { getLogger } from "../../../../util/logger.js";
 import { getWorkspaceDir } from "../../../../util/platform.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { readPage } from "../v2/page-store.js";
 import {
   deleteConceptPageEmbedding,

package/src/plugins/defaults/memory/pkb/pkb-index.ts

@@ -15,10 +15,9 @@ import { createHash } from "node:crypto";
 import { readdir, readFile, stat } from "node:fs/promises";
 import { join, relative } from "node:path";
 
-import { getConfig } from "../../../../config/loader.js";
 import { withQdrantBreaker } from "../../../../persistence/embeddings/qdrant-circuit-breaker.js";
 import { getQdrantClient } from "../../../../persistence/embeddings/qdrant-client.js";
-import { embedAndUpsert } from "../../../../persistence/job-utils.js";
+import { embedAndUpsert } from "../embeddings.js";
 import type { PkbIndexEntry } from "./types.js";
 import { PKB_TARGET_TYPE } from "./types.js";
 
@@ -189,8 +188,6 @@ export async function indexPkbFile(
   const relPath = relative(pkbRoot, absPath);
   const chunks = chunkPkbFile(content);
 
-  const config = getConfig();
-
   const qdrant = getQdrantClient();
   const existing = await withQdrantBreaker(() =>
     qdrant.scrollByTargetType(PKB_TARGET_TYPE, {
@@ -209,7 +206,6 @@ export async function indexPkbFile(
     const targetId = `${memoryScopeId}:${relPath}#${chunkIndex}`;
     newTargetIds.add(targetId);
     await embedAndUpsert(
-      config,
       PKB_TARGET_TYPE,
       targetId,
       { type: "text", text: chunk },

package/src/plugins/defaults/memory/routes/memory-item-routes.ts

@@ -27,7 +27,6 @@ import { z } from "zod";
 import { getConfig } from "../../../../config/loader.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import {
-  embedWithBackend,
   generateSparseEmbedding,
   getMemoryBackendStatus,
 } from "../../../../persistence/embeddings/embedding-backend.js";
@@ -46,6 +45,7 @@ import {
 } from "../../../../runtime/routes/errors.js";
 import type { RouteDefinition } from "../../../../runtime/routes/types.js";
 import { getLogger } from "../../../../util/logger.js";
+import { embedWithBackend } from "../embeddings.js";
 import { createNode, deleteNode, getNode, updateNode } from "../graph/store.js";
 import type {
   Fidelity,

package/src/plugins/defaults/memory/startup.ts

@@ -20,10 +20,7 @@ import {
   initMessagesLexicalIndex,
   MESSAGES_LEXICAL_COLLECTION,
 } from "../../../persistence/embeddings/messages-lexical-index.js";
-import {
-  initQdrantClient,
-  resolveQdrantUrl,
-} from "../../../persistence/embeddings/qdrant-client.js";
+import { initQdrantClient } from "../../../persistence/embeddings/qdrant-client.js";
 import { createQdrantManager } from "../../../persistence/embeddings/qdrant-manager.js";
 import {
   enqueueMemoryJob,
@@ -32,6 +29,7 @@ import {
 import { startMemoryJobsWorker } from "../../../persistence/jobs-worker.js";
 import { getLogger } from "../../../util/logger.js";
 import { getWorkspaceDir } from "../../../util/platform.js";
+import { resolveQdrantUrl } from "./embeddings.js";
 import { sweepConceptPageFrontmatter } from "./v2/frontmatter-sweep.js";
 import {
   maybeRebuildMemoryV2Concepts,
@@ -41,7 +39,7 @@ import {
 const log = getLogger("memory-startup");
 
 export async function runMemoryStartup(config: AssistantConfig): Promise<void> {
-  const qdrantUrl = resolveQdrantUrl(config);
+  const qdrantUrl = resolveQdrantUrl();
   log.info({ qdrantUrl }, "Daemon startup: initializing Qdrant");
   const manager = createQdrantManager({ url: qdrantUrl });
   const QDRANT_START_MAX_ATTEMPTS = 3;

package/src/plugins/defaults/memory/v2/activation.ts

@@ -35,11 +35,9 @@
 //     for the next turn and drop below `epsilon` if no longer relevant.
 
 import type { AssistantConfig } from "../../../../config/types.js";
-import {
-  embedWithBackend,
-  isEmbeddingDimensionAvailable,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { isEmbeddingDimensionAvailable } from "../../../../persistence/embeddings/embedding-backend.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { clampUnitInterval } from "../validation.js";
 import type { EdgeIndex } from "./edge-index.js";
 import { hybridQueryConceptPages } from "./qdrant.js";

package/src/plugins/defaults/memory/v2/cli-command-store.ts

@@ -21,12 +21,10 @@
 //     summary.
 
 import { getConfig } from "../../../../config/loader.js";
-import {
-  embedWithBackend,
-  generateSparseEmbedding,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { generateSparseEmbedding } from "../../../../persistence/embeddings/embedding-backend.js";
 import { getLogger } from "../../../../util/logger.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { buildCliCommandContent } from "./cli-command-content.js";
 import { invalidatePageIndex } from "./page-index.js";
 import {

package/src/plugins/defaults/memory/v2/qdrant.ts

@@ -27,9 +27,9 @@ import { v5 as uuidv5 } from "uuid";
 
 import { getConfig } from "../../../../config/loader.js";
 import type { SparseEmbedding } from "../../../../persistence/embeddings/embedding-types.js";
-import { resolveQdrantUrl } from "../../../../persistence/embeddings/qdrant-client.js";
 import { getLogger } from "../../../../util/logger.js";
 import { getDataDir } from "../../../../util/platform.js";
+import { resolveQdrantUrl } from "../embeddings.js";
 
 const log = getLogger("memory-v2-qdrant");
 
@@ -137,9 +137,8 @@ export async function clearReembedSentinel(): Promise<void> {
 /** Lazily create a Qdrant REST client bound to the resolved URL. */
 function getClient(): QdrantRestClient {
   if (_client) return _client;
-  const config = getConfig();
   _client = new QdrantRestClient({
-    url: resolveQdrantUrl(config),
+    url: resolveQdrantUrl(),
     checkCompatibility: false,
   });
   return _client;

package/src/plugins/defaults/memory/v2/sim.ts

@@ -33,11 +33,9 @@
 //   across turns.
 
 import type { AssistantConfig } from "../../../../config/types.js";
-import {
-  embedWithBackend,
-  isEmbeddingDimensionAvailable,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { isEmbeddingDimensionAvailable } from "../../../../persistence/embeddings/embedding-backend.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { clampUnitInterval } from "../validation.js";
 import { hybridQueryConceptPages } from "./qdrant.js";
 import { generateBm25QueryEmbedding } from "./sparse-bm25.js";

package/src/plugins/defaults/memory/v2/skill-store.ts

@@ -26,10 +26,7 @@ import { isAssistantFeatureFlagEnabled } from "../../../../config/assistant-feat
 import { getConfig } from "../../../../config/loader.js";
 import { resolveSkillStates } from "../../../../config/skill-state.js";
 import { loadSkillCatalog } from "../../../../config/skills.js";
-import {
-  embedWithBackend,
-  generateSparseEmbedding,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { generateSparseEmbedding } from "../../../../persistence/embeddings/embedding-backend.js";
 import { getCatalog } from "../../../../skills/catalog-cache.js";
 import {
   fromCatalogSkill,
@@ -37,6 +34,7 @@ import {
 } from "../../../../skills/skill-memory.js";
 import { getLogger } from "../../../../util/logger.js";
 import { applyCorrectionIfCalibrated } from "../anisotropy.js";
+import { embedWithBackend } from "../embeddings.js";
 import { invalidatePageIndex } from "./page-index.js";
 import {
   backfillKindOnPointsWithPrefix,

package/src/plugins/defaults/memory/v3/dense.ts

@@ -16,11 +16,9 @@
 // forward regardless, so a dense outage narrows recall but never breaks a turn.
 
 import type { AssistantConfig } from "../../../../config/types.js";
-import {
-  embedWithBackend,
-  isEmbeddingDimensionAvailable,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { isEmbeddingDimensionAvailable } from "../../../../persistence/embeddings/embedding-backend.js";
 import { getLogger } from "../../../../util/logger.js";
+import { embedWithBackend } from "../embeddings.js";
 import {
   getSectionDenseClient,
   SECTION_COLLECTION,
@@ -72,14 +70,11 @@ export async function denseLane(
     const vector = vectors[0];
     if (!vector || vector.length === 0) return [];
 
-    const result = await getSectionDenseClient(config).query(
-      SECTION_COLLECTION,
-      {
-        query: vector,
-        limit: k * OVERSAMPLE,
-        with_payload: true,
-      },
-    );
+    const result = await getSectionDenseClient().query(SECTION_COLLECTION, {
+      query: vector,
+      limit: k * OVERSAMPLE,
+      with_payload: true,
+    });
     points = result.points;
   } catch (err) {
     log.warn({ err }, "memory v3 dense lane failed; degrading to no hits");

package/src/plugins/defaults/memory/v3/maintain-job.ts

@@ -68,10 +68,7 @@ import {
   getMemoryCheckpoint,
   setMemoryCheckpoint,
 } from "../../../../persistence/checkpoints.js";
-import {
-  EmbeddingBackendUnavailableError,
-  embedWithBackend,
-} from "../../../../persistence/embeddings/embedding-backend.js";
+import { EmbeddingBackendUnavailableError } from "../../../../persistence/embeddings/embedding-backend.js";
 import { EmbeddingBillingBlockError } from "../../../../persistence/embeddings/embedding-billing-breaker.js";
 import type { MemoryJob } from "../../../../persistence/jobs-store.js";
 import {
@@ -81,6 +78,7 @@ import {
 import { executeDeleteManagedSkill } from "../../../../tools/skills/delete-managed.js";
 import { getLogger } from "../../../../util/logger.js";
 import { getWorkspaceDir } from "../../../../util/platform.js";
+import { embedWithBackend } from "../embeddings.js";
 import { getPageIndex } from "../v2/page-index.js";
 import { readPage } from "../v2/page-store.js";
 import { skillSlugFor } from "../v2/skill-store.js";

package/src/plugins/defaults/memory/v3/section-dense-store.ts

@@ -28,7 +28,6 @@ import type { AssistantConfig } from "../../../../config/types.js";
 import { deleteMemoryCheckpoint } from "../../../../persistence/checkpoints.js";
 import { getDb } from "../../../../persistence/db-connection.js";
 import {
-  embedWithBackend,
   geminiCacheExtras,
   getMemoryBackendStatus,
 } from "../../../../persistence/embeddings/embedding-backend.js";
@@ -37,8 +36,8 @@ import {
   writeEmbeddingCache,
 } from "../../../../persistence/embeddings/embedding-cache.js";
 import { embeddingInputContentHash } from "../../../../persistence/embeddings/embedding-types.js";
-import { resolveQdrantUrl } from "../../../../persistence/embeddings/qdrant-client.js";
 import { getLogger } from "../../../../util/logger.js";
+import { embedWithBackend, resolveQdrantUrl } from "../embeddings.js";
 import type { Section } from "./types.js";
 
 const log = getLogger("memory-v3-section-dense-store");
@@ -77,12 +76,10 @@ let _collectionReady = false;
  * Shared by both section-dense lanes (this store and the dense read lane in
  * `dense.ts`) so they reuse one client instance and one reset hook.
  */
-export function getSectionDenseClient(
-  config: AssistantConfig,
-): QdrantRestClient {
+export function getSectionDenseClient(): QdrantRestClient {
   if (_client) return _client;
   _client = new QdrantRestClient({
-    url: resolveQdrantUrl(config),
+    url: resolveQdrantUrl(),
     checkCompatibility: false,
   });
   return _client;
@@ -108,7 +105,7 @@ export async function ensureSectionCollection(
 ): Promise<void> {
   if (_collectionReady) return;
 
-  const client = getSectionDenseClient(config);
+  const client = getSectionDenseClient();
   const vectorSize = config.memory.qdrant.vectorSize;
   const onDisk = config.memory.qdrant.onDisk;
 
@@ -218,7 +215,7 @@ export async function ensureSectionCollection(
 export async function recreateSectionCollection(
   config: AssistantConfig,
 ): Promise<void> {
-  const client = getSectionDenseClient(config);
+  const client = getSectionDenseClient();
   const exists = await client.collectionExists(SECTION_COLLECTION);
   if (exists.exists) {
     await client.deleteCollection(SECTION_COLLECTION);
@@ -296,7 +293,7 @@ export async function upsertSections(
 
   if (points.length === 0) return;
 
-  await getSectionDenseClient(config).upsert(SECTION_COLLECTION, {
+  await getSectionDenseClient().upsert(SECTION_COLLECTION, {
     wait: true,
     points,
   });
@@ -418,7 +415,7 @@ export async function deleteSectionsForArticle(
 ): Promise<void> {
   await ensureSectionCollection(config);
 
-  await getSectionDenseClient(config).delete(SECTION_COLLECTION, {
+  await getSectionDenseClient().delete(SECTION_COLLECTION, {
     wait: true,
     filter: { must: [{ key: "article", match: { value: article } }] },
   });
@@ -440,7 +437,7 @@ export async function listSectionArticles(
 ): Promise<string[]> {
   await ensureSectionCollection(config);
 
-  const client = getSectionDenseClient(config);
+  const client = getSectionDenseClient();
   const articles = new Set<string>();
   let offset: string | number | undefined = undefined;
   const maxIterations = 10_000;