@anuma/sdk 6.16.0 → 6.17.1

package/dist/expo/index.cjs

@@ -9749,6 +9749,7 @@ __export(index_exports, {
   isDegradedTopicSkip: () => isDegradedTopicSkip,
   isPiiRedactor: () => isPiiRedactor,
   isRerankerAvailable: () => isRerankerAvailable,
+  maskScopedEmbeddingCache: () => maskScopedEmbeddingCache,
   moveMemoriesToFolderOp: () => moveMemoriesToFolderOp,
   noopLogger: () => noopLogger,
   onKeyAvailable: () => onKeyAvailable,
@@ -13485,6 +13486,28 @@ var xhrTransport = (options) => {
   return { stream: createStream() };
 };
 
+// src/lib/db/chat/embeddingCache.ts
+var MaskScopedEmbeddingCache = class extends Map {
+  constructor(inner, masked) {
+    super();
+    this.inner = inner;
+    this.masked = masked;
+  }
+  key(text12) {
+    return `${this.masked ? "m" : "r"}:${text12}`;
+  }
+  get(text12) {
+    return this.inner.get(this.key(text12));
+  }
+  set(text12, value) {
+    this.inner.set(this.key(text12), value);
+    return this;
+  }
+};
+function maskScopedEmbeddingCache(cache, masked) {
+  return new MaskScopedEmbeddingCache(cache, masked);
+}
+
 // src/react/useCredits.ts
 var import_react = require("react");
 init_sdk_gen();
@@ -20182,12 +20205,14 @@ var DEFAULT_SERVER_TOOLS_MATCH_OPTIONS = {
 };
 var SERVER_TOOL_DEPENDENCY_SETS = [
   {
-    // search finds links; reading them is always the next step.
-    name: "jina-research",
+    // search finds links; reading them is always the next step. The reader now
+    // lives on the search server — Jina's read_url / parallel_read_url were
+    // removed in favour of anuma_scrape_url, which takes a batch of URLs — so
+    // this edge deliberately spans two MCP servers.
+    name: "web-research",
     members: [
       "AnumaJinaMCP-search_web",
-      "AnumaJinaMCP-read_url",
-      "AnumaJinaMCP-parallel_read_url",
+      "AnumaSearchMCP-anuma_scrape_url",
       "AnumaJinaMCP-parallel_search_web"
     ],
     anchors: ["AnumaJinaMCP-search_web"],
@@ -21091,6 +21116,7 @@ function useChatStorage(options) {
         summaryModel = DEFAULT_SUMMARY_MODEL,
         files,
         storedUserContent,
+        embeddingCache,
         onData: perRequestOnData,
         onThinking: perRequestOnThinking,
         memoryContext,
@@ -21130,12 +21156,18 @@ function useChatStorage(options) {
       };
       pendingResumeRef.current = null;
       const effectiveAssistantUniqueId = assistantUniqueId ?? (resumable ? `msg_${(0, import_uuid5.v7)()}` : void 0);
-      const embedToolText = (text12, mask, token) => {
-        const opts = { getToken: token, baseUrl, model: embeddingModel };
+      const embedToolText = (text12, mask, masked, token) => {
+        const opts = {
+          getToken: token,
+          baseUrl,
+          model: embeddingModel,
+          maskInput: mask,
+          cache: embeddingCache ? maskScopedEmbeddingCache(embeddingCache, masked) : void 0
+        };
         return shouldChunkMessage(text12, DEFAULT_CHUNK_SIZE) ? generateEmbeddings(
-          chunkText(text12).map((c) => mask(c.text)),
+          chunkText(text12).map((c) => c.text),
           opts
-        ) : generateEmbedding(mask(text12), opts);
+        ) : generateEmbedding(text12, opts);
       };
       if (walletAddress && signMessage && !hasEncryptionKey(walletAddress)) {
         try {
@@ -21175,6 +21207,7 @@ function useChatStorage(options) {
                 skipUserEmbedding = await embedToolText(
                   messageContent,
                   maskForCall2,
+                  Boolean(callPiiRedaction2),
                   getTokenRef.current
                 );
                 const toolNames = serverToolsFilter(skipUserEmbedding, allServerTools);
@@ -21195,6 +21228,7 @@ function useChatStorage(options) {
                 skipUserEmbedding = await embedToolText(
                   messageContent,
                   maskForCall2,
+                  Boolean(callPiiRedaction2),
                   getTokenRef.current
                 );
               } catch {
@@ -21343,7 +21377,14 @@ function useChatStorage(options) {
       };
       const embedForToolsSettled = async (token) => {
         try {
-          return { embedding: await embedToolText(contentForStorage, maskForCall, token) };
+          return {
+            embedding: await embedToolText(
+              contentForStorage,
+              maskForCall,
+              Boolean(callPiiRedaction),
+              token
+            )
+          };
         } catch (error) {
           return { error };
         }
@@ -24825,6 +24866,7 @@ function createNotionTools(getAccessToken, requestNotionAccess) {
   isDegradedTopicSkip,
   isPiiRedactor,
   isRerankerAvailable,
+  maskScopedEmbeddingCache,
   moveMemoriesToFolderOp,
   noopLogger,
   onKeyAvailable,

package/dist/expo/index.d.mts

@@ -2755,6 +2755,38 @@ declare function resumeStream(options: ResumeStreamOptions): Promise<ResumeStrea
  */
 declare const xhrTransport: StreamingTransport;
 
+/**
+ * Namespacing for a caller-shared embedding cache.
+ *
+ * Lives here, next to the `embeddingCache` send-arg it serves, rather than in either platform hook:
+ * react and expo both thread the option through their own send paths, and a copy per platform is how
+ * the two drift.
+ */
+/**
+ * Build the same namespaced view over `cache` that a send with this masking decision uses.
+ *
+ * REQUIRED FOR SHARING, and the reason this is exported at all. `embeddingCache`'s whole purpose is
+ * that a caller who also needs the prompt vector passes one `Map` to both `sendMessage` and its own
+ * `generateEmbedding` call, so the turn embeds once. But the send namespaces its entries by masking
+ * decision (see {@link MaskScopedEmbeddingCache}) while `generateEmbedding` keys on the text alone —
+ * so a caller handing the RAW `Map` to its own call writes `"hello"` where the send looks for
+ * `"r:hello"`, and neither side ever hits. The dedupe silently does not happen.
+ *
+ * So: pass the plain `Map` as `embeddingCache`, and wrap it with this for your own call.
+ *
+ * ```ts
+ * const shared = new Map<string, Float32Array>();
+ * // your own ranking embed, same masking decision as the send:
+ * await generateEmbedding(text, { getToken, cache: maskScopedEmbeddingCache(shared, masked) });
+ * await sendMessage({ ..., embeddingCache: shared, piiRedaction: masked });
+ * ```
+ *
+ * `masked` must match what the send resolves for the same turn — i.e. whether PII redaction is on.
+ * Get it wrong and you simply miss the cache; you cannot be served the other decision's vector,
+ * which is the property the namespacing exists to guarantee.
+ */
+declare function maskScopedEmbeddingCache(cache: Map<string, Float32Array>, masked: boolean): Map<string, Float32Array>;
+
 /**
  * @inline
  */
@@ -4071,6 +4103,37 @@ interface BaseSendMessageWithStorageArgs {
      * user content (matching a textless turn), NOT a request to fall back.
      */
     storedUserContent?: string;
+    /**
+     * Optional embedding cache, shared with the caller.
+     *
+     * Every send that needs tool selection embeds the user text once (see
+     * `storedUserContent`). A caller that ALSO needs that vector — to rank tools itself, say, or to
+     * hand a server a prompt-aware shortlist — otherwise pays for a second, identical embedding of
+     * the same text in the same turn. Pass a `Map` here and into your own
+     * `generateEmbedding`/`generateEmbeddings` call and whichever runs first fills it; the other is a
+     * cache hit, so the turn embeds once.
+     *
+     * Keyed on the text **as passed in** (before `maskInput` is applied to the request body), prefixed
+     * with a marker for this send's masking decision — `"r:"` raw, `"m:"` masked — so a mismatched
+     * masking decision cannot silently serve the wrong vector: masked and unmasked occupy different
+     * entries. `generateEmbedding`'s own contract is unchanged; it still keys on the text alone, and
+     * the prefixing is a view this send wraps around the `Map` you hand it.
+     *
+     * WHICH MEANS SHARING TAKES ONE MORE STEP, and skipping it costs you the dedupe silently: pass
+     * the plain `Map` here, and wrap it with `maskScopedEmbeddingCache(map, masked)` for your OWN
+     * `generateEmbedding` call, so both sides look under the same key. Hand the raw `Map` to both and
+     * your call writes `"hello"` while this one reads `"r:hello"` — no hit, and the second embedding
+     * you were trying to avoid still happens.
+     *
+     * **Long messages are not deduped.** Past `DEFAULT_CHUNK_SIZE` (400 chars) the send embeds one
+     * vector per chunk and keys each entry on its chunk text, so a caller that embedded the whole
+     * prompt in one call finds nothing and both sides still pay. Making that hit would mean the caller
+     * reproducing the SDK's chunking, which is a worse contract than admitting the gap: short prompts
+     * — the overwhelming majority — dedupe, long ones do not.
+     *
+     * Omit it and nothing changes: every send embeds independently, exactly as before.
+     */
+    embeddingCache?: Map<string, Float32Array>;
     /**
      * Per-request callback invoked with each streamed response chunk.
      * Overrides the hook-level `onData` callback for this request only.
@@ -10591,4 +10654,4 @@ declare function refreshNotionAccessToken(params: NotionRefreshTokenParams): Pro
  */
 declare function createNotionTools(getAccessToken: () => string | null, requestNotionAccess: () => Promise<string>): ToolConfig[];
 
-export { type AutoExtractMessage, type AutoExtractor, type Budget, CHUNKS_DISCARDED_ORIGIN, type CachedChunkVectors, type CachedServerTools, Conversation as ChatConversation, Message as ChatMessage, type ChatRole, type ChunkVectorCache, type ChunkingOptions, type ConsolidationFallbackReason, type ConversationMemoryInput, ConversationMemory as ConversationMemoryModel, type ConversationMemoryOperationsContext, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateDecaySweeperOptions, type CreateMessageOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_CACHE_SIZE, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DECAY_POLICY, DEFAULT_MAX_CLASSIFIER_CALLS_PER_SWEEP, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_PROFILE_FACETS, DEFAULT_PROFILE_FACT_TYPE_WEIGHTS, DEFAULT_PROFILE_PROOF_ALPHA, DEFAULT_PROFILE_TREND_MULTIPLIERS, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type DecayCandidateRaw, type DecayClassifier, type DecayInput, type DecayPolicy, type DecaySweepResult, type DecaySweeper, type DecayVerdict, ENTITY_FANOUT, type EmbeddedWalletSignerFn, EncryptionKeyMissingError, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type ExtractionCursorStore, type FactType, type FileMetadata, type FlushResult, type GetConversationsPageOptions, type GetMessagesPageOptions, type GraphTraversalOptions, HARD_DELETE_WINDOW_MS, INFERENCE_ID_HEADER, type InjectionClassifierOptions, type InjectionReason, type LazyStoredConversation, type LlmDecayClassifierOptions, type LlmNeighborRefinerOptions, type Logger, LoggerProvider, type LoggerProviderProps, MAX_HOPS, MEDIUM_TTL_MS, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryQuarantinedEvent, type MemoryToVerify, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, type MemoryVerification, type MessageRedactionResult, type MessageSkeleton, NEVER_TTL_MS, NODE_BUDGET, NOTION_OAUTH_CONFIG, type NeighborRefiner, type NotionAuthUrlParams, type NotionClientRegistration, type NotionExchangeCodeParams, type NotionOAuthEndpoints, type NotionPKCEChallenge, type NotionRefreshTokenParams, type NotionTokenResponse, type NowSource, type ObservationTrend, type ObservationTrendInput, PAST_EVENT_GRACE_MS, PII_PATTERNS, PROFILE_DOC_VERSION, type PhotoIngestResult, type PhotoMediaRef, type PiiCategory, type PiiMatch, type PiiPattern, PiiRedactor, type PiiRedactorOptions, type PlatformStorage, type PortalLlmAuth, type PortalLlmFailure, type PortalLlmFailureReason, type ProfileConfigFingerprint, type ProfileDoc, type ProfileFacet, type ProfileFacetKey, type ProfileSalienceInput, type ProfileSection, type PromptPreProcessor, type PromptPreProcessorContext, type PublishedPhotoMemory, type QuantizedEmbedding, type QuarantinedMemoryInfo, QueueManager, type QueueStatus, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankableVaultMemory, type RankedMemory, type RankedProfileCandidate, type RecallContext, type RecallDegradation, type RecallDiagnostics, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type RedactionResult, type ReflectOptions, type ReflectResult, type RequestEncryptionKeyOptions, RerankerUnavailableError, type ResumeStreamOptions, type ResumeStreamResult, type ResumeStreamWithStorageResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, SDK_SCHEMA_VERSION, SHORT_TTL_MS, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type ScoreProfileSalienceOptions, type ScreenResult, type ScreenedCandidate, type SearchSource, type SendMessageWithStorageArgs, type SendMessageWithStorageDetachedResult, type SendMessageWithStorageResult, type ServerToolsOptions, type ServerToolsResponse, type SignMessageFn, type StockPriceClassification, type StockPricePreProcessorOptions, type ChatCompletionUsage as StoredChatCompletionUsage, type StoredConversation, type StoredConversationMemory, type StoredMessage, type StoredMessageWithSimilarity, type StoredVaultFolder, VaultFolder as StoredVaultFolderModel, type StoredVaultMemory, VaultMemory as StoredVaultMemoryModel, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type SynthesizeProfileOptions, TOOL_RESULT_ORIGIN, TOPICS_EXTRACTION_VERSION, TOPIC_EXTRACTION_BATCH_SIZE, TREND_RECENT_WINDOW_DAYS, TREND_STALE_WINDOW_DAYS, type TextChunk, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TopicSkipReason, type TurnCompleteEvent, type TurnSkippedEvent, type UncheckedReason, type UnverifiableReason, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, type UseChatStorageOptions, type UseChatStorageResult, type UseCreditsOptions, type UseCreditsResult, type UseModelsOptions, type UseModelsResult, VAULT_SIZE_HOP_CAP, type VaultEmbeddingCache, type VaultFolderOperationsContext, type VaultMemoryOperationsContext, type VaultMemoryVisibility, type VaultSaveOperation, type VaultSearchResult, type VerificationSources, type VerifyMemoriesForPublishOptions, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, addConversationMemoriesOp, archiveVaultMemoryOp, backfillMemoryTopicsOp, buildNotionAuthUrl, capHopsForDensity, chatStorageMigrations, chatStorageSchema, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyDecay, classifyInjectionCandidates, classifyObservationTrend, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, clearAllEncryptionKeys, clearAllEncryptionState, clearConversationMemoriesOp, clearEncryptionKey, clearLazyTitleCache, clearServerToolsCache, consoleLogger, conversationMemoryToStored, cosineInt8, createAutoExtractor, createChunkVectorCache, createCryptoPricePreProcessor, createDecaySweeper, createLlmDecayClassifier, createLlmNeighborRefiner, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageSourceResolver, createNotionTools, createPlatformCursorStore, createRecallTool, createStockPricePreProcessor, createStreamingDeAnonymizer, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decodeChunkVector, decryptConversationTitle, deleteAllVaultMemoriesForUserOp, deleteVaultFolderOp, deleteVaultMemoryOp, dequantizeEmbedding, discoverNotionOAuthEndpoints, eagerEmbedContent, embedAllMessages, embedMessage, encodeChunkVector, ensureDefaultFoldersOp, exchangeNotionCode, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, generateConversationId, generateEmbedding, generateEmbeddings, generateNotionPKCE, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getCachedServerTools, getConversationMemoriesOp, getConversationsByProjectLazyOp, getConversationsLazyOp, getConversationsPageOp, getDecayCandidatesRawOp, getLogger, getMemoriesNeedingTopicExtractionOp, getServerTools, getUnfiledVaultMemoriesOp, getVaultFolderMemoryCountOp, getVaultMemoriesByIdsOp, getVaultMemoryOp, getVaultRankingProjectionsOp, hardDeleteDecayedOp, hasEncryptionKey, ingestPublishedPhotoMemoriesOp, injectionSignatureCatalog, isDegradedTopicSkip, isPiiRedactor, isRerankerAvailable, moveMemoriesToFolderOp, noopLogger, onKeyAvailable, parseMedia, preEmbedVaultMemories, quantizeEmbedding, queueManager, rankProfileCandidates, recall, reflect, refreshEncryptionKeyIfMatches, refreshNotionAccessToken, registerNotionClient, relinkMemoryTopicsOp, requestEncryptionKey, resolvePiiRedactor, restoreVaultMemoryOp, resumeStream, retain, scoreProfileSalience, screenCandidatesForInjection, sdkMigrations, sdkModelClasses, sdkSchema, searchVaultMemories, seedEncryptionKeys, setLogger, setMemoryVisibilityOp, shouldChunkMessage, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summarizeObservationTrends, supersedeVaultMemoryOp, synthesizeProfile, traverseGraphLane, ttlForType, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertMessageOp, useChat, useChatStorage, useCredits, useEncryption, useModels, verifyMemoriesForPublish, xhrTransport };
+export { type AutoExtractMessage, type AutoExtractor, type Budget, CHUNKS_DISCARDED_ORIGIN, type CachedChunkVectors, type CachedServerTools, Conversation as ChatConversation, Message as ChatMessage, type ChatRole, type ChunkVectorCache, type ChunkingOptions, type ConsolidationFallbackReason, type ConversationMemoryInput, ConversationMemory as ConversationMemoryModel, type ConversationMemoryOperationsContext, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateDecaySweeperOptions, type CreateMessageOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_CACHE_SIZE, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DECAY_POLICY, DEFAULT_MAX_CLASSIFIER_CALLS_PER_SWEEP, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_PROFILE_FACETS, DEFAULT_PROFILE_FACT_TYPE_WEIGHTS, DEFAULT_PROFILE_PROOF_ALPHA, DEFAULT_PROFILE_TREND_MULTIPLIERS, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type DecayCandidateRaw, type DecayClassifier, type DecayInput, type DecayPolicy, type DecaySweepResult, type DecaySweeper, type DecayVerdict, ENTITY_FANOUT, type EmbeddedWalletSignerFn, EncryptionKeyMissingError, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type ExtractionCursorStore, type FactType, type FileMetadata, type FlushResult, type GetConversationsPageOptions, type GetMessagesPageOptions, type GraphTraversalOptions, HARD_DELETE_WINDOW_MS, INFERENCE_ID_HEADER, type InjectionClassifierOptions, type InjectionReason, type LazyStoredConversation, type LlmDecayClassifierOptions, type LlmNeighborRefinerOptions, type Logger, LoggerProvider, type LoggerProviderProps, MAX_HOPS, MEDIUM_TTL_MS, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryQuarantinedEvent, type MemoryToVerify, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, type MemoryVerification, type MessageRedactionResult, type MessageSkeleton, NEVER_TTL_MS, NODE_BUDGET, NOTION_OAUTH_CONFIG, type NeighborRefiner, type NotionAuthUrlParams, type NotionClientRegistration, type NotionExchangeCodeParams, type NotionOAuthEndpoints, type NotionPKCEChallenge, type NotionRefreshTokenParams, type NotionTokenResponse, type NowSource, type ObservationTrend, type ObservationTrendInput, PAST_EVENT_GRACE_MS, PII_PATTERNS, PROFILE_DOC_VERSION, type PhotoIngestResult, type PhotoMediaRef, type PiiCategory, type PiiMatch, type PiiPattern, PiiRedactor, type PiiRedactorOptions, type PlatformStorage, type PortalLlmAuth, type PortalLlmFailure, type PortalLlmFailureReason, type ProfileConfigFingerprint, type ProfileDoc, type ProfileFacet, type ProfileFacetKey, type ProfileSalienceInput, type ProfileSection, type PromptPreProcessor, type PromptPreProcessorContext, type PublishedPhotoMemory, type QuantizedEmbedding, type QuarantinedMemoryInfo, QueueManager, type QueueStatus, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankableVaultMemory, type RankedMemory, type RankedProfileCandidate, type RecallContext, type RecallDegradation, type RecallDiagnostics, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type RedactionResult, type ReflectOptions, type ReflectResult, type RequestEncryptionKeyOptions, RerankerUnavailableError, type ResumeStreamOptions, type ResumeStreamResult, type ResumeStreamWithStorageResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, SDK_SCHEMA_VERSION, SHORT_TTL_MS, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type ScoreProfileSalienceOptions, type ScreenResult, type ScreenedCandidate, type SearchSource, type SendMessageWithStorageArgs, type SendMessageWithStorageDetachedResult, type SendMessageWithStorageResult, type ServerToolsOptions, type ServerToolsResponse, type SignMessageFn, type StockPriceClassification, type StockPricePreProcessorOptions, type ChatCompletionUsage as StoredChatCompletionUsage, type StoredConversation, type StoredConversationMemory, type StoredMessage, type StoredMessageWithSimilarity, type StoredVaultFolder, VaultFolder as StoredVaultFolderModel, type StoredVaultMemory, VaultMemory as StoredVaultMemoryModel, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type SynthesizeProfileOptions, TOOL_RESULT_ORIGIN, TOPICS_EXTRACTION_VERSION, TOPIC_EXTRACTION_BATCH_SIZE, TREND_RECENT_WINDOW_DAYS, TREND_STALE_WINDOW_DAYS, type TextChunk, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TopicSkipReason, type TurnCompleteEvent, type TurnSkippedEvent, type UncheckedReason, type UnverifiableReason, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, type UseChatStorageOptions, type UseChatStorageResult, type UseCreditsOptions, type UseCreditsResult, type UseModelsOptions, type UseModelsResult, VAULT_SIZE_HOP_CAP, type VaultEmbeddingCache, type VaultFolderOperationsContext, type VaultMemoryOperationsContext, type VaultMemoryVisibility, type VaultSaveOperation, type VaultSearchResult, type VerificationSources, type VerifyMemoriesForPublishOptions, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, addConversationMemoriesOp, archiveVaultMemoryOp, backfillMemoryTopicsOp, buildNotionAuthUrl, capHopsForDensity, chatStorageMigrations, chatStorageSchema, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyDecay, classifyInjectionCandidates, classifyObservationTrend, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, clearAllEncryptionKeys, clearAllEncryptionState, clearConversationMemoriesOp, clearEncryptionKey, clearLazyTitleCache, clearServerToolsCache, consoleLogger, conversationMemoryToStored, cosineInt8, createAutoExtractor, createChunkVectorCache, createCryptoPricePreProcessor, createDecaySweeper, createLlmDecayClassifier, createLlmNeighborRefiner, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageSourceResolver, createNotionTools, createPlatformCursorStore, createRecallTool, createStockPricePreProcessor, createStreamingDeAnonymizer, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decodeChunkVector, decryptConversationTitle, deleteAllVaultMemoriesForUserOp, deleteVaultFolderOp, deleteVaultMemoryOp, dequantizeEmbedding, discoverNotionOAuthEndpoints, eagerEmbedContent, embedAllMessages, embedMessage, encodeChunkVector, ensureDefaultFoldersOp, exchangeNotionCode, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, generateConversationId, generateEmbedding, generateEmbeddings, generateNotionPKCE, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getCachedServerTools, getConversationMemoriesOp, getConversationsByProjectLazyOp, getConversationsLazyOp, getConversationsPageOp, getDecayCandidatesRawOp, getLogger, getMemoriesNeedingTopicExtractionOp, getServerTools, getUnfiledVaultMemoriesOp, getVaultFolderMemoryCountOp, getVaultMemoriesByIdsOp, getVaultMemoryOp, getVaultRankingProjectionsOp, hardDeleteDecayedOp, hasEncryptionKey, ingestPublishedPhotoMemoriesOp, injectionSignatureCatalog, isDegradedTopicSkip, isPiiRedactor, isRerankerAvailable, maskScopedEmbeddingCache, moveMemoriesToFolderOp, noopLogger, onKeyAvailable, parseMedia, preEmbedVaultMemories, quantizeEmbedding, queueManager, rankProfileCandidates, recall, reflect, refreshEncryptionKeyIfMatches, refreshNotionAccessToken, registerNotionClient, relinkMemoryTopicsOp, requestEncryptionKey, resolvePiiRedactor, restoreVaultMemoryOp, resumeStream, retain, scoreProfileSalience, screenCandidatesForInjection, sdkMigrations, sdkModelClasses, sdkSchema, searchVaultMemories, seedEncryptionKeys, setLogger, setMemoryVisibilityOp, shouldChunkMessage, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summarizeObservationTrends, supersedeVaultMemoryOp, synthesizeProfile, traverseGraphLane, ttlForType, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertMessageOp, useChat, useChatStorage, useCredits, useEncryption, useModels, verifyMemoriesForPublish, xhrTransport };

package/dist/expo/index.d.ts

@@ -2755,6 +2755,38 @@ declare function resumeStream(options: ResumeStreamOptions): Promise<ResumeStrea
  */
 declare const xhrTransport: StreamingTransport;
 
+/**
+ * Namespacing for a caller-shared embedding cache.
+ *
+ * Lives here, next to the `embeddingCache` send-arg it serves, rather than in either platform hook:
+ * react and expo both thread the option through their own send paths, and a copy per platform is how
+ * the two drift.
+ */
+/**
+ * Build the same namespaced view over `cache` that a send with this masking decision uses.
+ *
+ * REQUIRED FOR SHARING, and the reason this is exported at all. `embeddingCache`'s whole purpose is
+ * that a caller who also needs the prompt vector passes one `Map` to both `sendMessage` and its own
+ * `generateEmbedding` call, so the turn embeds once. But the send namespaces its entries by masking
+ * decision (see {@link MaskScopedEmbeddingCache}) while `generateEmbedding` keys on the text alone —
+ * so a caller handing the RAW `Map` to its own call writes `"hello"` where the send looks for
+ * `"r:hello"`, and neither side ever hits. The dedupe silently does not happen.
+ *
+ * So: pass the plain `Map` as `embeddingCache`, and wrap it with this for your own call.
+ *
+ * ```ts
+ * const shared = new Map<string, Float32Array>();
+ * // your own ranking embed, same masking decision as the send:
+ * await generateEmbedding(text, { getToken, cache: maskScopedEmbeddingCache(shared, masked) });
+ * await sendMessage({ ..., embeddingCache: shared, piiRedaction: masked });
+ * ```
+ *
+ * `masked` must match what the send resolves for the same turn — i.e. whether PII redaction is on.
+ * Get it wrong and you simply miss the cache; you cannot be served the other decision's vector,
+ * which is the property the namespacing exists to guarantee.
+ */
+declare function maskScopedEmbeddingCache(cache: Map<string, Float32Array>, masked: boolean): Map<string, Float32Array>;
+
 /**
  * @inline
  */
@@ -4071,6 +4103,37 @@ interface BaseSendMessageWithStorageArgs {
      * user content (matching a textless turn), NOT a request to fall back.
      */
     storedUserContent?: string;
+    /**
+     * Optional embedding cache, shared with the caller.
+     *
+     * Every send that needs tool selection embeds the user text once (see
+     * `storedUserContent`). A caller that ALSO needs that vector — to rank tools itself, say, or to
+     * hand a server a prompt-aware shortlist — otherwise pays for a second, identical embedding of
+     * the same text in the same turn. Pass a `Map` here and into your own
+     * `generateEmbedding`/`generateEmbeddings` call and whichever runs first fills it; the other is a
+     * cache hit, so the turn embeds once.
+     *
+     * Keyed on the text **as passed in** (before `maskInput` is applied to the request body), prefixed
+     * with a marker for this send's masking decision — `"r:"` raw, `"m:"` masked — so a mismatched
+     * masking decision cannot silently serve the wrong vector: masked and unmasked occupy different
+     * entries. `generateEmbedding`'s own contract is unchanged; it still keys on the text alone, and
+     * the prefixing is a view this send wraps around the `Map` you hand it.
+     *
+     * WHICH MEANS SHARING TAKES ONE MORE STEP, and skipping it costs you the dedupe silently: pass
+     * the plain `Map` here, and wrap it with `maskScopedEmbeddingCache(map, masked)` for your OWN
+     * `generateEmbedding` call, so both sides look under the same key. Hand the raw `Map` to both and
+     * your call writes `"hello"` while this one reads `"r:hello"` — no hit, and the second embedding
+     * you were trying to avoid still happens.
+     *
+     * **Long messages are not deduped.** Past `DEFAULT_CHUNK_SIZE` (400 chars) the send embeds one
+     * vector per chunk and keys each entry on its chunk text, so a caller that embedded the whole
+     * prompt in one call finds nothing and both sides still pay. Making that hit would mean the caller
+     * reproducing the SDK's chunking, which is a worse contract than admitting the gap: short prompts
+     * — the overwhelming majority — dedupe, long ones do not.
+     *
+     * Omit it and nothing changes: every send embeds independently, exactly as before.
+     */
+    embeddingCache?: Map<string, Float32Array>;
     /**
      * Per-request callback invoked with each streamed response chunk.
      * Overrides the hook-level `onData` callback for this request only.
@@ -10591,4 +10654,4 @@ declare function refreshNotionAccessToken(params: NotionRefreshTokenParams): Pro
  */
 declare function createNotionTools(getAccessToken: () => string | null, requestNotionAccess: () => Promise<string>): ToolConfig[];
 
-export { type AutoExtractMessage, type AutoExtractor, type Budget, CHUNKS_DISCARDED_ORIGIN, type CachedChunkVectors, type CachedServerTools, Conversation as ChatConversation, Message as ChatMessage, type ChatRole, type ChunkVectorCache, type ChunkingOptions, type ConsolidationFallbackReason, type ConversationMemoryInput, ConversationMemory as ConversationMemoryModel, type ConversationMemoryOperationsContext, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateDecaySweeperOptions, type CreateMessageOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_CACHE_SIZE, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DECAY_POLICY, DEFAULT_MAX_CLASSIFIER_CALLS_PER_SWEEP, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_PROFILE_FACETS, DEFAULT_PROFILE_FACT_TYPE_WEIGHTS, DEFAULT_PROFILE_PROOF_ALPHA, DEFAULT_PROFILE_TREND_MULTIPLIERS, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type DecayCandidateRaw, type DecayClassifier, type DecayInput, type DecayPolicy, type DecaySweepResult, type DecaySweeper, type DecayVerdict, ENTITY_FANOUT, type EmbeddedWalletSignerFn, EncryptionKeyMissingError, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type ExtractionCursorStore, type FactType, type FileMetadata, type FlushResult, type GetConversationsPageOptions, type GetMessagesPageOptions, type GraphTraversalOptions, HARD_DELETE_WINDOW_MS, INFERENCE_ID_HEADER, type InjectionClassifierOptions, type InjectionReason, type LazyStoredConversation, type LlmDecayClassifierOptions, type LlmNeighborRefinerOptions, type Logger, LoggerProvider, type LoggerProviderProps, MAX_HOPS, MEDIUM_TTL_MS, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryQuarantinedEvent, type MemoryToVerify, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, type MemoryVerification, type MessageRedactionResult, type MessageSkeleton, NEVER_TTL_MS, NODE_BUDGET, NOTION_OAUTH_CONFIG, type NeighborRefiner, type NotionAuthUrlParams, type NotionClientRegistration, type NotionExchangeCodeParams, type NotionOAuthEndpoints, type NotionPKCEChallenge, type NotionRefreshTokenParams, type NotionTokenResponse, type NowSource, type ObservationTrend, type ObservationTrendInput, PAST_EVENT_GRACE_MS, PII_PATTERNS, PROFILE_DOC_VERSION, type PhotoIngestResult, type PhotoMediaRef, type PiiCategory, type PiiMatch, type PiiPattern, PiiRedactor, type PiiRedactorOptions, type PlatformStorage, type PortalLlmAuth, type PortalLlmFailure, type PortalLlmFailureReason, type ProfileConfigFingerprint, type ProfileDoc, type ProfileFacet, type ProfileFacetKey, type ProfileSalienceInput, type ProfileSection, type PromptPreProcessor, type PromptPreProcessorContext, type PublishedPhotoMemory, type QuantizedEmbedding, type QuarantinedMemoryInfo, QueueManager, type QueueStatus, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankableVaultMemory, type RankedMemory, type RankedProfileCandidate, type RecallContext, type RecallDegradation, type RecallDiagnostics, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type RedactionResult, type ReflectOptions, type ReflectResult, type RequestEncryptionKeyOptions, RerankerUnavailableError, type ResumeStreamOptions, type ResumeStreamResult, type ResumeStreamWithStorageResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, SDK_SCHEMA_VERSION, SHORT_TTL_MS, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type ScoreProfileSalienceOptions, type ScreenResult, type ScreenedCandidate, type SearchSource, type SendMessageWithStorageArgs, type SendMessageWithStorageDetachedResult, type SendMessageWithStorageResult, type ServerToolsOptions, type ServerToolsResponse, type SignMessageFn, type StockPriceClassification, type StockPricePreProcessorOptions, type ChatCompletionUsage as StoredChatCompletionUsage, type StoredConversation, type StoredConversationMemory, type StoredMessage, type StoredMessageWithSimilarity, type StoredVaultFolder, VaultFolder as StoredVaultFolderModel, type StoredVaultMemory, VaultMemory as StoredVaultMemoryModel, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type SynthesizeProfileOptions, TOOL_RESULT_ORIGIN, TOPICS_EXTRACTION_VERSION, TOPIC_EXTRACTION_BATCH_SIZE, TREND_RECENT_WINDOW_DAYS, TREND_STALE_WINDOW_DAYS, type TextChunk, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TopicSkipReason, type TurnCompleteEvent, type TurnSkippedEvent, type UncheckedReason, type UnverifiableReason, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, type UseChatStorageOptions, type UseChatStorageResult, type UseCreditsOptions, type UseCreditsResult, type UseModelsOptions, type UseModelsResult, VAULT_SIZE_HOP_CAP, type VaultEmbeddingCache, type VaultFolderOperationsContext, type VaultMemoryOperationsContext, type VaultMemoryVisibility, type VaultSaveOperation, type VaultSearchResult, type VerificationSources, type VerifyMemoriesForPublishOptions, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, addConversationMemoriesOp, archiveVaultMemoryOp, backfillMemoryTopicsOp, buildNotionAuthUrl, capHopsForDensity, chatStorageMigrations, chatStorageSchema, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyDecay, classifyInjectionCandidates, classifyObservationTrend, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, clearAllEncryptionKeys, clearAllEncryptionState, clearConversationMemoriesOp, clearEncryptionKey, clearLazyTitleCache, clearServerToolsCache, consoleLogger, conversationMemoryToStored, cosineInt8, createAutoExtractor, createChunkVectorCache, createCryptoPricePreProcessor, createDecaySweeper, createLlmDecayClassifier, createLlmNeighborRefiner, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageSourceResolver, createNotionTools, createPlatformCursorStore, createRecallTool, createStockPricePreProcessor, createStreamingDeAnonymizer, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decodeChunkVector, decryptConversationTitle, deleteAllVaultMemoriesForUserOp, deleteVaultFolderOp, deleteVaultMemoryOp, dequantizeEmbedding, discoverNotionOAuthEndpoints, eagerEmbedContent, embedAllMessages, embedMessage, encodeChunkVector, ensureDefaultFoldersOp, exchangeNotionCode, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, generateConversationId, generateEmbedding, generateEmbeddings, generateNotionPKCE, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getCachedServerTools, getConversationMemoriesOp, getConversationsByProjectLazyOp, getConversationsLazyOp, getConversationsPageOp, getDecayCandidatesRawOp, getLogger, getMemoriesNeedingTopicExtractionOp, getServerTools, getUnfiledVaultMemoriesOp, getVaultFolderMemoryCountOp, getVaultMemoriesByIdsOp, getVaultMemoryOp, getVaultRankingProjectionsOp, hardDeleteDecayedOp, hasEncryptionKey, ingestPublishedPhotoMemoriesOp, injectionSignatureCatalog, isDegradedTopicSkip, isPiiRedactor, isRerankerAvailable, moveMemoriesToFolderOp, noopLogger, onKeyAvailable, parseMedia, preEmbedVaultMemories, quantizeEmbedding, queueManager, rankProfileCandidates, recall, reflect, refreshEncryptionKeyIfMatches, refreshNotionAccessToken, registerNotionClient, relinkMemoryTopicsOp, requestEncryptionKey, resolvePiiRedactor, restoreVaultMemoryOp, resumeStream, retain, scoreProfileSalience, screenCandidatesForInjection, sdkMigrations, sdkModelClasses, sdkSchema, searchVaultMemories, seedEncryptionKeys, setLogger, setMemoryVisibilityOp, shouldChunkMessage, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summarizeObservationTrends, supersedeVaultMemoryOp, synthesizeProfile, traverseGraphLane, ttlForType, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertMessageOp, useChat, useChatStorage, useCredits, useEncryption, useModels, verifyMemoriesForPublish, xhrTransport };
+export { type AutoExtractMessage, type AutoExtractor, type Budget, CHUNKS_DISCARDED_ORIGIN, type CachedChunkVectors, type CachedServerTools, Conversation as ChatConversation, Message as ChatMessage, type ChatRole, type ChunkVectorCache, type ChunkingOptions, type ConsolidationFallbackReason, type ConversationMemoryInput, ConversationMemory as ConversationMemoryModel, type ConversationMemoryOperationsContext, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateDecaySweeperOptions, type CreateMessageOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_CACHE_SIZE, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DECAY_POLICY, DEFAULT_MAX_CLASSIFIER_CALLS_PER_SWEEP, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_PROFILE_FACETS, DEFAULT_PROFILE_FACT_TYPE_WEIGHTS, DEFAULT_PROFILE_PROOF_ALPHA, DEFAULT_PROFILE_TREND_MULTIPLIERS, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type DecayCandidateRaw, type DecayClassifier, type DecayInput, type DecayPolicy, type DecaySweepResult, type DecaySweeper, type DecayVerdict, ENTITY_FANOUT, type EmbeddedWalletSignerFn, EncryptionKeyMissingError, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type ExtractionCursorStore, type FactType, type FileMetadata, type FlushResult, type GetConversationsPageOptions, type GetMessagesPageOptions, type GraphTraversalOptions, HARD_DELETE_WINDOW_MS, INFERENCE_ID_HEADER, type InjectionClassifierOptions, type InjectionReason, type LazyStoredConversation, type LlmDecayClassifierOptions, type LlmNeighborRefinerOptions, type Logger, LoggerProvider, type LoggerProviderProps, MAX_HOPS, MEDIUM_TTL_MS, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryQuarantinedEvent, type MemoryToVerify, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, type MemoryVerification, type MessageRedactionResult, type MessageSkeleton, NEVER_TTL_MS, NODE_BUDGET, NOTION_OAUTH_CONFIG, type NeighborRefiner, type NotionAuthUrlParams, type NotionClientRegistration, type NotionExchangeCodeParams, type NotionOAuthEndpoints, type NotionPKCEChallenge, type NotionRefreshTokenParams, type NotionTokenResponse, type NowSource, type ObservationTrend, type ObservationTrendInput, PAST_EVENT_GRACE_MS, PII_PATTERNS, PROFILE_DOC_VERSION, type PhotoIngestResult, type PhotoMediaRef, type PiiCategory, type PiiMatch, type PiiPattern, PiiRedactor, type PiiRedactorOptions, type PlatformStorage, type PortalLlmAuth, type PortalLlmFailure, type PortalLlmFailureReason, type ProfileConfigFingerprint, type ProfileDoc, type ProfileFacet, type ProfileFacetKey, type ProfileSalienceInput, type ProfileSection, type PromptPreProcessor, type PromptPreProcessorContext, type PublishedPhotoMemory, type QuantizedEmbedding, type QuarantinedMemoryInfo, QueueManager, type QueueStatus, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankableVaultMemory, type RankedMemory, type RankedProfileCandidate, type RecallContext, type RecallDegradation, type RecallDiagnostics, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type RedactionResult, type ReflectOptions, type ReflectResult, type RequestEncryptionKeyOptions, RerankerUnavailableError, type ResumeStreamOptions, type ResumeStreamResult, type ResumeStreamWithStorageResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, SDK_SCHEMA_VERSION, SHORT_TTL_MS, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type ScoreProfileSalienceOptions, type ScreenResult, type ScreenedCandidate, type SearchSource, type SendMessageWithStorageArgs, type SendMessageWithStorageDetachedResult, type SendMessageWithStorageResult, type ServerToolsOptions, type ServerToolsResponse, type SignMessageFn, type StockPriceClassification, type StockPricePreProcessorOptions, type ChatCompletionUsage as StoredChatCompletionUsage, type StoredConversation, type StoredConversationMemory, type StoredMessage, type StoredMessageWithSimilarity, type StoredVaultFolder, VaultFolder as StoredVaultFolderModel, type StoredVaultMemory, VaultMemory as StoredVaultMemoryModel, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type SynthesizeProfileOptions, TOOL_RESULT_ORIGIN, TOPICS_EXTRACTION_VERSION, TOPIC_EXTRACTION_BATCH_SIZE, TREND_RECENT_WINDOW_DAYS, TREND_STALE_WINDOW_DAYS, type TextChunk, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TopicSkipReason, type TurnCompleteEvent, type TurnSkippedEvent, type UncheckedReason, type UnverifiableReason, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, type UseChatStorageOptions, type UseChatStorageResult, type UseCreditsOptions, type UseCreditsResult, type UseModelsOptions, type UseModelsResult, VAULT_SIZE_HOP_CAP, type VaultEmbeddingCache, type VaultFolderOperationsContext, type VaultMemoryOperationsContext, type VaultMemoryVisibility, type VaultSaveOperation, type VaultSearchResult, type VerificationSources, type VerifyMemoriesForPublishOptions, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, addConversationMemoriesOp, archiveVaultMemoryOp, backfillMemoryTopicsOp, buildNotionAuthUrl, capHopsForDensity, chatStorageMigrations, chatStorageSchema, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyDecay, classifyInjectionCandidates, classifyObservationTrend, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, clearAllEncryptionKeys, clearAllEncryptionState, clearConversationMemoriesOp, clearEncryptionKey, clearLazyTitleCache, clearServerToolsCache, consoleLogger, conversationMemoryToStored, cosineInt8, createAutoExtractor, createChunkVectorCache, createCryptoPricePreProcessor, createDecaySweeper, createLlmDecayClassifier, createLlmNeighborRefiner, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageSourceResolver, createNotionTools, createPlatformCursorStore, createRecallTool, createStockPricePreProcessor, createStreamingDeAnonymizer, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decodeChunkVector, decryptConversationTitle, deleteAllVaultMemoriesForUserOp, deleteVaultFolderOp, deleteVaultMemoryOp, dequantizeEmbedding, discoverNotionOAuthEndpoints, eagerEmbedContent, embedAllMessages, embedMessage, encodeChunkVector, ensureDefaultFoldersOp, exchangeNotionCode, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, generateConversationId, generateEmbedding, generateEmbeddings, generateNotionPKCE, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getCachedServerTools, getConversationMemoriesOp, getConversationsByProjectLazyOp, getConversationsLazyOp, getConversationsPageOp, getDecayCandidatesRawOp, getLogger, getMemoriesNeedingTopicExtractionOp, getServerTools, getUnfiledVaultMemoriesOp, getVaultFolderMemoryCountOp, getVaultMemoriesByIdsOp, getVaultMemoryOp, getVaultRankingProjectionsOp, hardDeleteDecayedOp, hasEncryptionKey, ingestPublishedPhotoMemoriesOp, injectionSignatureCatalog, isDegradedTopicSkip, isPiiRedactor, isRerankerAvailable, maskScopedEmbeddingCache, moveMemoriesToFolderOp, noopLogger, onKeyAvailable, parseMedia, preEmbedVaultMemories, quantizeEmbedding, queueManager, rankProfileCandidates, recall, reflect, refreshEncryptionKeyIfMatches, refreshNotionAccessToken, registerNotionClient, relinkMemoryTopicsOp, requestEncryptionKey, resolvePiiRedactor, restoreVaultMemoryOp, resumeStream, retain, scoreProfileSalience, screenCandidatesForInjection, sdkMigrations, sdkModelClasses, sdkSchema, searchVaultMemories, seedEncryptionKeys, setLogger, setMemoryVisibilityOp, shouldChunkMessage, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summarizeObservationTrends, supersedeVaultMemoryOp, synthesizeProfile, traverseGraphLane, ttlForType, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertMessageOp, useChat, useChatStorage, useCredits, useEncryption, useModels, verifyMemoriesForPublish, xhrTransport };

package/dist/expo/index.mjs

@@ -3818,6 +3818,28 @@ var xhrTransport = (options) => {
   return { stream: createStream() };
 };
 
+// src/lib/db/chat/embeddingCache.ts
+var MaskScopedEmbeddingCache = class extends Map {
+  constructor(inner, masked) {
+    super();
+    this.inner = inner;
+    this.masked = masked;
+  }
+  key(text12) {
+    return `${this.masked ? "m" : "r"}:${text12}`;
+  }
+  get(text12) {
+    return this.inner.get(this.key(text12));
+  }
+  set(text12, value) {
+    this.inner.set(this.key(text12), value);
+    return this;
+  }
+};
+function maskScopedEmbeddingCache(cache, masked) {
+  return new MaskScopedEmbeddingCache(cache, masked);
+}
+
 // src/react/useCredits.ts
 import { useCallback, useEffect, useRef, useState } from "react";
 function useCredits(options = {}) {
@@ -10383,12 +10405,14 @@ var DEFAULT_SERVER_TOOLS_MATCH_OPTIONS = {
 };
 var SERVER_TOOL_DEPENDENCY_SETS = [
   {
-    // search finds links; reading them is always the next step.
-    name: "jina-research",
+    // search finds links; reading them is always the next step. The reader now
+    // lives on the search server — Jina's read_url / parallel_read_url were
+    // removed in favour of anuma_scrape_url, which takes a batch of URLs — so
+    // this edge deliberately spans two MCP servers.
+    name: "web-research",
     members: [
       "AnumaJinaMCP-search_web",
-      "AnumaJinaMCP-read_url",
-      "AnumaJinaMCP-parallel_read_url",
+      "AnumaSearchMCP-anuma_scrape_url",
       "AnumaJinaMCP-parallel_search_web"
     ],
     anchors: ["AnumaJinaMCP-search_web"],
@@ -11291,6 +11315,7 @@ function useChatStorage(options) {
         summaryModel = DEFAULT_SUMMARY_MODEL,
         files,
         storedUserContent,
+        embeddingCache,
         onData: perRequestOnData,
         onThinking: perRequestOnThinking,
         memoryContext,
@@ -11330,12 +11355,18 @@ function useChatStorage(options) {
       };
       pendingResumeRef.current = null;
       const effectiveAssistantUniqueId = assistantUniqueId ?? (resumable ? `msg_${uuidv73()}` : void 0);
-      const embedToolText = (text12, mask, token) => {
-        const opts = { getToken: token, baseUrl, model: embeddingModel };
+      const embedToolText = (text12, mask, masked, token) => {
+        const opts = {
+          getToken: token,
+          baseUrl,
+          model: embeddingModel,
+          maskInput: mask,
+          cache: embeddingCache ? maskScopedEmbeddingCache(embeddingCache, masked) : void 0
+        };
         return shouldChunkMessage(text12, DEFAULT_CHUNK_SIZE) ? generateEmbeddings(
-          chunkText(text12).map((c) => mask(c.text)),
+          chunkText(text12).map((c) => c.text),
           opts
-        ) : generateEmbedding(mask(text12), opts);
+        ) : generateEmbedding(text12, opts);
       };
       if (walletAddress && signMessage && !hasEncryptionKey(walletAddress)) {
         try {
@@ -11375,6 +11406,7 @@ function useChatStorage(options) {
                 skipUserEmbedding = await embedToolText(
                   messageContent,
                   maskForCall2,
+                  Boolean(callPiiRedaction2),
                   getTokenRef.current
                 );
                 const toolNames = serverToolsFilter(skipUserEmbedding, allServerTools);
@@ -11395,6 +11427,7 @@ function useChatStorage(options) {
                 skipUserEmbedding = await embedToolText(
                   messageContent,
                   maskForCall2,
+                  Boolean(callPiiRedaction2),
                   getTokenRef.current
                 );
               } catch {
@@ -11543,7 +11576,14 @@ function useChatStorage(options) {
       };
       const embedForToolsSettled = async (token) => {
         try {
-          return { embedding: await embedToolText(contentForStorage, maskForCall, token) };
+          return {
+            embedding: await embedToolText(
+              contentForStorage,
+              maskForCall,
+              Boolean(callPiiRedaction),
+              token
+            )
+          };
         } catch (error) {
           return { error };
         }
@@ -15017,6 +15057,7 @@ export {
   isDegradedTopicSkip,
   isPiiRedactor,
   isRerankerAvailable,
+  maskScopedEmbeddingCache,
   moveMemoriesToFolderOp,
   noopLogger,
   onKeyAvailable,

package/dist/index.cjs

@@ -2332,12 +2332,14 @@ var DEFAULT_SERVER_TOOLS_MATCH_OPTIONS = {
 };
 var SERVER_TOOL_DEPENDENCY_SETS = [
   {
-    // search finds links; reading them is always the next step.
-    name: "jina-research",
+    // search finds links; reading them is always the next step. The reader now
+    // lives on the search server — Jina's read_url / parallel_read_url were
+    // removed in favour of anuma_scrape_url, which takes a batch of URLs — so
+    // this edge deliberately spans two MCP servers.
+    name: "web-research",
     members: [
       "AnumaJinaMCP-search_web",
-      "AnumaJinaMCP-read_url",
-      "AnumaJinaMCP-parallel_read_url",
+      "AnumaSearchMCP-anuma_scrape_url",
       "AnumaJinaMCP-parallel_search_web"
     ],
     anchors: ["AnumaJinaMCP-search_web"],

package/dist/index.d.mts

@@ -7808,7 +7808,7 @@ declare function applyToolSets(matchedNames: Set<string>, availableNames: Set<st
  * precision.
  *
  * Use this for server-side toolkit suites where the LLM needs the full
- * call chain (e.g. search_web → read_url / parallel_read_url, or
+ * call chain (e.g. search_web → anuma_scrape_url, or
  * geocoding before the OpenMeteo data tools). Differs from `applyToolSets`,
  * which replaces non-set matches when a set activates.
  *
@@ -7890,7 +7890,7 @@ interface CreateServerToolsFilterOptions {
  *   toolSets: [
  *     {
  *       name: "research",
- *       members: ["AnumaJinaMCP-search_web", "AnumaJinaMCP-read_url", ...],
+ *       members: ["AnumaJinaMCP-search_web", "AnumaSearchMCP-anuma_scrape_url", ...],
  *       anchors: ["AnumaJinaMCP-search_web"],
  *       anchorMinSimilarity: 0.7,
  *     },

package/dist/index.d.ts

@@ -7808,7 +7808,7 @@ declare function applyToolSets(matchedNames: Set<string>, availableNames: Set<st
  * precision.
  *
  * Use this for server-side toolkit suites where the LLM needs the full
- * call chain (e.g. search_web → read_url / parallel_read_url, or
+ * call chain (e.g. search_web → anuma_scrape_url, or
  * geocoding before the OpenMeteo data tools). Differs from `applyToolSets`,
  * which replaces non-set matches when a set activates.
  *
@@ -7890,7 +7890,7 @@ interface CreateServerToolsFilterOptions {
  *   toolSets: [
  *     {
  *       name: "research",
- *       members: ["AnumaJinaMCP-search_web", "AnumaJinaMCP-read_url", ...],
+ *       members: ["AnumaJinaMCP-search_web", "AnumaSearchMCP-anuma_scrape_url", ...],
  *       anchors: ["AnumaJinaMCP-search_web"],
  *       anchorMinSimilarity: 0.7,
  *     },

package/dist/index.mjs

@@ -2130,12 +2130,14 @@ var DEFAULT_SERVER_TOOLS_MATCH_OPTIONS = {
 };
 var SERVER_TOOL_DEPENDENCY_SETS = [
   {
-    // search finds links; reading them is always the next step.
-    name: "jina-research",
+    // search finds links; reading them is always the next step. The reader now
+    // lives on the search server — Jina's read_url / parallel_read_url were
+    // removed in favour of anuma_scrape_url, which takes a batch of URLs — so
+    // this edge deliberately spans two MCP servers.
+    name: "web-research",
     members: [
       "AnumaJinaMCP-search_web",
-      "AnumaJinaMCP-read_url",
-      "AnumaJinaMCP-parallel_read_url",
+      "AnumaSearchMCP-anuma_scrape_url",
       "AnumaJinaMCP-parallel_search_web"
     ],
     anchors: ["AnumaJinaMCP-search_web"],