@anuma/sdk 10.5.1 → 10.5.3

package/dist/server/index.cjs

@@ -19345,6 +19345,28 @@ async function chunkRowSource(ctx, filter) {
     return fullScanChunkRowSource(ctx, filter);
   }
 }
+async function readChunkVectorRow(ctx, row, onUnreadable) {
+  const version = Number(row.updated_at) || 0;
+  const chunks = await readRawJsonField(row, "chunks", ctx.walletAddress);
+  if (chunks && chunks.length > 0) {
+    return {
+      entry: {
+        version,
+        chunks: chunks.map((chunk) => decodeChunkVector(chunk.vector, onUnreadable))
+      },
+      chunks
+    };
+  }
+  const messageVector = await readRawJsonField(row, "vector", ctx.walletAddress);
+  if (!messageVector || messageVector.length === 0) return null;
+  return { entry: { version, chunks: [], fallback: Float32Array.from(messageVector) }, chunks: [] };
+}
+function warnUnreadableChunkVectors(operation, count) {
+  if (count === 0) return;
+  getLogger().warn(
+    `${operation}: could not read ${count} chunk vectors and left them out of ranking \u2014 the stored value is corrupt, so those chunks stay unsearchable until the message is re-embedded via chunkAndEmbedAllMessages`
+  );
+}
 async function searchChunksOp(ctx, queryVector, options) {
   const {
     limit = 10,
@@ -19413,45 +19435,39 @@ async function searchChunksOp(ctx, queryVector, options) {
     for (const row of rows) {
       await yieldIfDue();
       const id = row.id;
-      const version = Number(row.updated_at) || 0;
       const order = orderById.get(id) ?? Number.MAX_SAFE_INTEGER;
-      const chunks = await readRawJsonField(row, "chunks", ctx.walletAddress);
-      if (chunks && chunks.length > 0) {
-        const vectors = chunks.map(
-          (chunk) => decodeChunkVector(chunk.vector, () => {
-            unreadableVectors++;
-          })
-        );
-        chunkCache?.set(id, { version, chunks: vectors });
-        for (let ci = 0; ci < chunks.length; ci++) {
-          const vec = vectors[ci];
-          if (vec.length === 0) continue;
-          const similarity = cosineSimilarity(queryVector, vec);
-          if (similarity >= minSimilarity) {
-            candidates.push({
-              id,
-              version,
-              order,
-              similarity,
-              chunkTextSource: { kind: "chunk", chunk: chunks[ci], siblings: chunks }
-            });
-          }
+      const read = await readChunkVectorRow(ctx, row, () => {
+        unreadableVectors++;
+      });
+      if (!read) continue;
+      const { entry, chunks } = read;
+      const { version } = entry;
+      chunkCache?.set(id, entry);
+      if (entry.fallback) {
+        const similarity = cosineSimilarity(queryVector, entry.fallback);
+        if (similarity >= minSimilarity) {
+          candidates.push({
+            id,
+            version,
+            order,
+            similarity,
+            chunkTextSource: { kind: "message" }
+          });
         }
-      } else {
-        const messageVector = await readRawJsonField(row, "vector", ctx.walletAddress);
-        if (messageVector && messageVector.length > 0) {
-          const fallback = Float32Array.from(messageVector);
-          chunkCache?.set(id, { version, chunks: [], fallback });
-          const similarity = cosineSimilarity(queryVector, fallback);
-          if (similarity >= minSimilarity) {
-            candidates.push({
-              id,
-              version,
-              order,
-              similarity,
-              chunkTextSource: { kind: "message" }
-            });
-          }
+        continue;
+      }
+      for (let ci = 0; ci < chunks.length; ci++) {
+        const vec = entry.chunks[ci];
+        if (vec.length === 0) continue;
+        const similarity = cosineSimilarity(queryVector, vec);
+        if (similarity >= minSimilarity) {
+          candidates.push({
+            id,
+            version,
+            order,
+            similarity,
+            chunkTextSource: { kind: "chunk", chunk: chunks[ci], siblings: chunks }
+          });
         }
       }
     }
@@ -19462,11 +19478,7 @@ async function searchChunksOp(ctx, queryVector, options) {
       `searchChunksOp: skipped ${staleSkipped} messages whose embedding model differs from the current model (${embeddingModel}) \u2014 re-embed via chunkAndEmbedAllMessages`
     );
   }
-  if (unreadableVectors > 0) {
-    getLogger().warn(
-      `searchChunksOp: could not read ${unreadableVectors} chunk vectors and left them out of ranking \u2014 the stored value is corrupt, so those chunks stay unsearchable until the message is re-embedded via chunkAndEmbedAllMessages`
-    );
-  }
+  warnUnreadableChunkVectors("searchChunksOp", unreadableVectors);
   const topK = candidates.sort((a, b) => b.similarity - a.similarity || a.order - b.order).slice(0, limit);
   const topIds = Array.from(new Set(topK.map((c) => c.id)));
   const fullRows = await source.loadFullRows(topIds);
@@ -33056,12 +33068,16 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const uncachedTexts = [];
     const uncachedIndices = [];
+    const resolvedVectors = /* @__PURE__ */ new Map();
     let staleReembedCount = 0;
     for (let i = 0; queryEmbedding.length > 0 && i < memories.length; i++) {
       const content = memories[i].content;
       const memoryId = memories[i].uniqueId;
       const cached = cachedRowVector(cache, memoryId, memories[i].updatedAt, content);
-      if (cached && cached.length === queryEmbedding.length) continue;
+      if (cached && cached.length === queryEmbedding.length) {
+        resolvedVectors.set(memoryId, cached);
+        continue;
+      }
       if (cached) cache.delete(memoryId);
       const storedModel = memories[i].embeddingModel;
       const modelCompatible = (storedModel ?? currentModel) === currentModel;
@@ -33069,13 +33085,9 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
         try {
           const parsed = JSON.parse(memories[i].embedding);
           if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-            cacheRowVector(
-              cache,
-              memoryId,
-              Float32Array.from(parsed),
-              memories[i].updatedAt,
-              content
-            );
+            const vector = Float32Array.from(parsed);
+            resolvedVectors.set(memoryId, vector);
+            cacheRowVector(cache, memoryId, vector, memories[i].updatedAt, content);
             continue;
           }
         } catch {
@@ -33108,20 +33120,17 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
       }
       if (newEmbeddings) {
         for (let j = 0; j < uncachedTexts.length; j++) {
-          cacheRowVector(
-            cache,
-            memories[uncachedIndices[j]].uniqueId,
-            Float32Array.from(newEmbeddings[j]),
-            memories[uncachedIndices[j]].updatedAt,
-            memories[uncachedIndices[j]].content
-          );
+          const memory = memories[uncachedIndices[j]];
+          const vector = Float32Array.from(newEmbeddings[j]);
+          resolvedVectors.set(memory.uniqueId, vector);
+          cacheRowVector(cache, memory.uniqueId, vector, memory.updatedAt, memory.content);
         }
       }
     }
     embeddedItems = memories.map((m) => ({
       id: m.uniqueId,
       content: m.content,
-      embedding: cache.get(m.uniqueId) ?? [],
+      embedding: resolvedVectors.get(m.uniqueId) ?? [],
       createdAt: m.createdAt,
       updatedAt: m.updatedAt,
       proofCount: m.proofCount,

package/dist/server/index.d.mts

@@ -4100,6 +4100,13 @@ interface ChatBackend extends ChatStorageAdapter {
         uniqueId: string;
     }): Promise<StoredMessage>;
     searchChunks(queryVector: number[], options?: ChatChunkSearchOptions): Promise<ChunkSearchResult[]>;
+    /** Fills a chunk vector cache ahead of the first search, for backends that score on the device. */
+    warmChunkCache?(options: {
+        chunkCache: ChunkVectorCache;
+        embeddingModel?: string;
+        limit?: number;
+        signal?: AbortSignal;
+    }): Promise<number>;
     getSummary(conversationId: string): Promise<StoredConversationSummary | null>;
     upsertSummary(conversationId: string, summary: string, summarizedUpTo: string, tokenCount: number): Promise<StoredConversationSummary>;
     deleteSummary(conversationId: string): Promise<void>;

package/dist/server/index.d.ts

@@ -4100,6 +4100,13 @@ interface ChatBackend extends ChatStorageAdapter {
         uniqueId: string;
     }): Promise<StoredMessage>;
     searchChunks(queryVector: number[], options?: ChatChunkSearchOptions): Promise<ChunkSearchResult[]>;
+    /** Fills a chunk vector cache ahead of the first search, for backends that score on the device. */
+    warmChunkCache?(options: {
+        chunkCache: ChunkVectorCache;
+        embeddingModel?: string;
+        limit?: number;
+        signal?: AbortSignal;
+    }): Promise<number>;
     getSummary(conversationId: string): Promise<StoredConversationSummary | null>;
     upsertSummary(conversationId: string, summary: string, summarizedUpTo: string, tokenCount: number): Promise<StoredConversationSummary>;
     deleteSummary(conversationId: string): Promise<void>;

package/dist/server/index.mjs

@@ -17637,6 +17637,28 @@ async function chunkRowSource(ctx, filter) {
     return fullScanChunkRowSource(ctx, filter);
   }
 }
+async function readChunkVectorRow(ctx, row, onUnreadable) {
+  const version = Number(row.updated_at) || 0;
+  const chunks = await readRawJsonField(row, "chunks", ctx.walletAddress);
+  if (chunks && chunks.length > 0) {
+    return {
+      entry: {
+        version,
+        chunks: chunks.map((chunk) => decodeChunkVector(chunk.vector, onUnreadable))
+      },
+      chunks
+    };
+  }
+  const messageVector = await readRawJsonField(row, "vector", ctx.walletAddress);
+  if (!messageVector || messageVector.length === 0) return null;
+  return { entry: { version, chunks: [], fallback: Float32Array.from(messageVector) }, chunks: [] };
+}
+function warnUnreadableChunkVectors(operation, count) {
+  if (count === 0) return;
+  getLogger().warn(
+    `${operation}: could not read ${count} chunk vectors and left them out of ranking \u2014 the stored value is corrupt, so those chunks stay unsearchable until the message is re-embedded via chunkAndEmbedAllMessages`
+  );
+}
 async function searchChunksOp(ctx, queryVector, options) {
   const {
     limit = 10,
@@ -17705,45 +17727,39 @@ async function searchChunksOp(ctx, queryVector, options) {
     for (const row of rows) {
       await yieldIfDue();
       const id = row.id;
-      const version = Number(row.updated_at) || 0;
       const order = orderById.get(id) ?? Number.MAX_SAFE_INTEGER;
-      const chunks = await readRawJsonField(row, "chunks", ctx.walletAddress);
-      if (chunks && chunks.length > 0) {
-        const vectors = chunks.map(
-          (chunk) => decodeChunkVector(chunk.vector, () => {
-            unreadableVectors++;
-          })
-        );
-        chunkCache?.set(id, { version, chunks: vectors });
-        for (let ci = 0; ci < chunks.length; ci++) {
-          const vec = vectors[ci];
-          if (vec.length === 0) continue;
-          const similarity = cosineSimilarity(queryVector, vec);
-          if (similarity >= minSimilarity) {
-            candidates.push({
-              id,
-              version,
-              order,
-              similarity,
-              chunkTextSource: { kind: "chunk", chunk: chunks[ci], siblings: chunks }
-            });
-          }
+      const read = await readChunkVectorRow(ctx, row, () => {
+        unreadableVectors++;
+      });
+      if (!read) continue;
+      const { entry, chunks } = read;
+      const { version } = entry;
+      chunkCache?.set(id, entry);
+      if (entry.fallback) {
+        const similarity = cosineSimilarity(queryVector, entry.fallback);
+        if (similarity >= minSimilarity) {
+          candidates.push({
+            id,
+            version,
+            order,
+            similarity,
+            chunkTextSource: { kind: "message" }
+          });
         }
-      } else {
-        const messageVector = await readRawJsonField(row, "vector", ctx.walletAddress);
-        if (messageVector && messageVector.length > 0) {
-          const fallback = Float32Array.from(messageVector);
-          chunkCache?.set(id, { version, chunks: [], fallback });
-          const similarity = cosineSimilarity(queryVector, fallback);
-          if (similarity >= minSimilarity) {
-            candidates.push({
-              id,
-              version,
-              order,
-              similarity,
-              chunkTextSource: { kind: "message" }
-            });
-          }
+        continue;
+      }
+      for (let ci = 0; ci < chunks.length; ci++) {
+        const vec = entry.chunks[ci];
+        if (vec.length === 0) continue;
+        const similarity = cosineSimilarity(queryVector, vec);
+        if (similarity >= minSimilarity) {
+          candidates.push({
+            id,
+            version,
+            order,
+            similarity,
+            chunkTextSource: { kind: "chunk", chunk: chunks[ci], siblings: chunks }
+          });
         }
       }
     }
@@ -17754,11 +17770,7 @@ async function searchChunksOp(ctx, queryVector, options) {
       `searchChunksOp: skipped ${staleSkipped} messages whose embedding model differs from the current model (${embeddingModel}) \u2014 re-embed via chunkAndEmbedAllMessages`
     );
   }
-  if (unreadableVectors > 0) {
-    getLogger().warn(
-      `searchChunksOp: could not read ${unreadableVectors} chunk vectors and left them out of ranking \u2014 the stored value is corrupt, so those chunks stay unsearchable until the message is re-embedded via chunkAndEmbedAllMessages`
-    );
-  }
+  warnUnreadableChunkVectors("searchChunksOp", unreadableVectors);
   const topK = candidates.sort((a, b) => b.similarity - a.similarity || a.order - b.order).slice(0, limit);
   const topIds = Array.from(new Set(topK.map((c) => c.id)));
   const fullRows = await source.loadFullRows(topIds);
@@ -31294,12 +31306,16 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const uncachedTexts = [];
     const uncachedIndices = [];
+    const resolvedVectors = /* @__PURE__ */ new Map();
     let staleReembedCount = 0;
     for (let i = 0; queryEmbedding.length > 0 && i < memories.length; i++) {
       const content = memories[i].content;
       const memoryId = memories[i].uniqueId;
       const cached = cachedRowVector(cache, memoryId, memories[i].updatedAt, content);
-      if (cached && cached.length === queryEmbedding.length) continue;
+      if (cached && cached.length === queryEmbedding.length) {
+        resolvedVectors.set(memoryId, cached);
+        continue;
+      }
       if (cached) cache.delete(memoryId);
       const storedModel = memories[i].embeddingModel;
       const modelCompatible = (storedModel ?? currentModel) === currentModel;
@@ -31307,13 +31323,9 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
         try {
           const parsed = JSON.parse(memories[i].embedding);
           if (Array.isArray(parsed) && parsed.length === queryEmbedding.length) {
-            cacheRowVector(
-              cache,
-              memoryId,
-              Float32Array.from(parsed),
-              memories[i].updatedAt,
-              content
-            );
+            const vector = Float32Array.from(parsed);
+            resolvedVectors.set(memoryId, vector);
+            cacheRowVector(cache, memoryId, vector, memories[i].updatedAt, content);
             continue;
           }
         } catch {
@@ -31346,20 +31358,17 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
       }
       if (newEmbeddings) {
         for (let j = 0; j < uncachedTexts.length; j++) {
-          cacheRowVector(
-            cache,
-            memories[uncachedIndices[j]].uniqueId,
-            Float32Array.from(newEmbeddings[j]),
-            memories[uncachedIndices[j]].updatedAt,
-            memories[uncachedIndices[j]].content
-          );
+          const memory = memories[uncachedIndices[j]];
+          const vector = Float32Array.from(newEmbeddings[j]);
+          resolvedVectors.set(memory.uniqueId, vector);
+          cacheRowVector(cache, memory.uniqueId, vector, memory.updatedAt, memory.content);
         }
       }
     }
     embeddedItems = memories.map((m) => ({
       id: m.uniqueId,
       content: m.content,
-      embedding: cache.get(m.uniqueId) ?? [],
+      embedding: resolvedVectors.get(m.uniqueId) ?? [],
       createdAt: m.createdAt,
       updatedAt: m.updatedAt,
       proofCount: m.proofCount,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "10.5.1",
+  "version": "10.5.3",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",