@anuma/sdk 2.3.1 → 2.4.1

package/dist/server/index.d.ts

@@ -3873,6 +3873,37 @@ declare function searchMessagesOp(ctx: StorageOperationsContext, queryVector: nu
     minSimilarity?: number;
     conversationId?: string;
 }): Promise<StoredMessageWithSimilarity[]>;
+/**
+ * Decrypted, model-native chunk vectors for one message, cached across recall
+ * calls so the chunk lane doesn't re-decrypt + re-JSON.parse every message's
+ * embeddings on every query. Keyed by message id; `version` is the message's
+ * `updated_at` at cache time, so a re-embed (which bumps `updated_at`) misses
+ * and repopulates. Vectors are stored Float32 — model-native precision, half
+ * the RAM of float64 `number[]` — mirroring the vault embedding cache (#705).
+ * Content text is deliberately NOT cached (see searchChunksOp): it isn't
+ * needed for scoring, and keeping decrypted content out of the cache preserves
+ * the SDK's data-at-rest posture.
+ */
+interface CachedChunkVectors {
+    /** Message `updated_at` (ms) at cache time — the invalidation token. */
+    version: number;
+    /**
+     * Per-chunk embedding vectors, index-aligned with the message's decrypted
+     * `chunks` array (empty-vector chunks kept as zero-length placeholders so
+     * indices stay aligned). Empty when the message has only a whole-message
+     * `fallback` vector.
+     */
+    chunks: Float32Array[];
+    /** Whole-message vector when the message was embedded without chunking. */
+    fallback?: Float32Array;
+}
+/**
+ * Cache consumed by {@link searchChunksOp}, keyed by message id. A plain `Map`
+ * (satisfied by the LRU from `createChunkVectorCache`) — mirrors
+ * `VaultEmbeddingCache` so consumers get the same `clear()`/`delete()`
+ * ergonomics on an encryption-key reset.
+ */
+type ChunkVectorCache = Map<string, CachedChunkVectors>;
 /**
  * Search through message chunks for fine-grained semantic search.
  * Returns the matching chunk text along with the parent message.
@@ -3890,6 +3921,13 @@ declare function searchChunksOp(ctx: StorageOperationsContext, queryVector: numb
      * messages are re-embedded out-of-band by `chunkAndEmbedAllMessages`.
      */
     embeddingModel?: string;
+    /**
+     * Optional decrypted-vector cache. When provided, the per-query decrypt +
+     * JSON.parse of every message's chunk vectors is skipped on warm entries
+     * (validated by `updated_at`), which is the dominant cost of the chunk
+     * lane. Omit for the legacy always-decrypt behavior.
+     */
+    chunkCache?: ChunkVectorCache;
 }): Promise<ChunkSearchResult[]>;
 /**
  * Get all files from all conversations, sorted by creation date (newest first).
@@ -5710,6 +5748,14 @@ interface RecallOptions {
      * resolves windows in 2026 and never overlaps stored event_time.
      */
     now?: number;
+    /**
+     * Best-effort observability hook. Called once per `recall()` with per-lane
+     * timings, lane counts, and soft-degradation signals — the raw material for
+     * tuning latency/quality and for wiring recall telemetry to PostHog. Invoked
+     * synchronously just before `recall()` returns; a throwing callback is
+     * swallowed (diagnostics must never break retrieval). Off unless provided.
+     */
+    onDiagnostics?: (diagnostics: RecallDiagnostics) => void;
     /** Number of candidates fed to the cross-encoder rerank stage. Default: 30. */
     rerankTopN?: number;
     /** Multiplicative cross-encoder blend weight. Default: 0.1. */
@@ -5740,6 +5786,12 @@ interface RecallContext {
     embeddingOptions: EmbeddingOptions;
     /** Vault embedding LRU cache. */
     vaultCache?: VaultEmbeddingCache;
+    /**
+     * Optional chunk-vector LRU cache. When provided, the chunk lane skips the
+     * per-query decrypt + JSON.parse of every message's chunk vectors on warm
+     * entries. Build via `createChunkVectorCache`. Omit for legacy behavior.
+     */
+    chunkCache?: ChunkVectorCache;
     /**
      * Optional — when provided, recall extracts entities from the query
      * and adds a graph lane to the RRF fusion (memories sharing entities
@@ -5759,6 +5811,48 @@ interface RecallResult {
     /** Diagnostic: total memories in the vault when fact lane was queried. */
     vaultSize?: number;
 }
+/** Soft-degradation signals surfaced via {@link RecallDiagnostics.degraded}. */
+type RecallDegradation = 
+/** Rerank was requested (budget mid/high) but the cross-encoder didn't run
+ *  this call — unavailable (e.g. React Native) or a transient failure. */
+"rerank-unavailable"
+/** `budget: 'high'` requested but no `decomposeOptions`, so query
+ *  decomposition was skipped and the budget downgraded to mid. */
+ | "decompose-unavailable";
+/**
+ * Per-call recall observability payload (see {@link RecallOptions.onDiagnostics}).
+ * All timings are wall-clock milliseconds. Lane counts are post-dedupe,
+ * pre-fusion. Intended to be forwarded to a metrics sink (e.g. PostHog).
+ */
+interface RecallDiagnostics {
+    /** Budget actually executed (may have downgraded from the requested one). */
+    usedBudget: Budget;
+    /** Whether the cross-encoder actually reranked the fact lane this call. */
+    reranked: boolean;
+    /** Total candidates considered before truncation. */
+    candidateCount: number;
+    /** Total vault size when the fact lane ran (absent if it didn't). */
+    vaultSize?: number;
+    /** Facts the fact lane returned (post-dedupe, pre-fusion). */
+    factCount: number;
+    /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */
+    chunkCount: number;
+    /** Wall-clock phase timings (ms). */
+    timings: {
+        /** Whole `recall()` call. */
+        total: number;
+        /** Parallel query-embed + graph/temporal side-lane build. */
+        prep: number;
+        /** Vault fact-lane search (`searchVaultMemoriesWithSize`). */
+        factLane: number;
+        /** Chunk-lane search (`searchChunksOp`). */
+        chunkLane: number;
+        /** Cross-lane RRF fusion + provenance dedup after both lanes. */
+        fuse: number;
+    };
+    /** Soft-degradation signals that fired this call (empty when clean). */
+    degraded: RecallDegradation[];
+}
 type RetainAction = "create" | "merge" | "update" | "skip" | "suppressed";
 type RetainSource = "manual" | "auto-extracted" | "capsule";
 /** Why the consolidator fell back to "create" instead of a real decision. */
@@ -6109,15 +6203,32 @@ declare function extractAndRetain(messages: AutoExtractMessage[], retainCtx: Ret
  * newest (superset) message list re-covers any superseded turn. Turns for
  * different conversations queue independently and never displace each other.
  *
- * Watermark + window: the worker keeps an in-memory, per-conversation
- * watermark of the last message it extracted through. Each run sends every
- * message *after* that watermark (plus a small overlap for coreference,
- * capped at `maxWindowSize`) rather than a fixed trailing slice — so facts
- * stated in turns that arrived during a busy window are still examined instead
- * of scrolling out of a last-N window. The watermark is session-scoped (in
- * memory); it does NOT survive process death, so an extraction killed mid-flight
- * by app backgrounding is not replayed in a later session. Durable cross-process
- * replay would require a persisted watermark (out of scope here).
+ * Watermark + window: the worker keeps a per-conversation watermark of the
+ * last message it extracted through. Each run sends every message *after* that
+ * watermark (plus a small overlap for coreference, capped at `maxWindowSize`)
+ * rather than a fixed trailing slice — so facts stated in turns that arrived
+ * during a busy window are still examined instead of scrolling out of a last-N
+ * window.
+ *
+ * Durability: the in-memory watermark alone does NOT survive process death, so
+ * a session killed after messages accumulate but before extraction would, on
+ * the next launch, fall back to a trailing-window guess and could skip the
+ * un-extracted tail. Pass a {@link ExtractionCursorStore} (`cursorStore`) to
+ * persist the watermark per conversation — it is hydrated synchronously the
+ * first time a conversation is touched, and written through on every *contiguous*
+ * advance (never on a trailing-slice guess, which could jump the durable cursor
+ * past an un-extracted gap), so a later session resumes exactly after the last
+ * extracted message. The primary guarantee is single-session resume across
+ * process death.
+ *
+ * Concurrent sessions sharing one store (web `localStorage`, mobile MMKV) are
+ * best-effort, NOT exclusive: two sessions can each hydrate the same cursor
+ * before either advances it and briefly re-extract the same turn (retain()'s
+ * dedup/consolidation absorbs the duplicate — wasteful, not corrupting). The
+ * durable write is last-writer-wins with a regression guard (it won't move the
+ * cursor backwards when the currently-stored id is ahead within the same turn's
+ * messages), but it does not serialize writers. Omit `cursorStore` for the
+ * legacy in-memory-only behavior.
  *
  * Memory Studio panel subscribes by passing `onMemoryExtracted` and/or
  * `onTurnComplete` callbacks. The Studio uses `onMemoryExtracted` to
@@ -6165,6 +6276,37 @@ interface TurnCompleteEvent {
     outcome: ExtractOutcome;
 }
 /** @public */
+/**
+ * Durable per-conversation extraction cursor. Synchronous by contract (both
+ * SDK platform stores — web `localStorage`, mobile MMKV — are sync), so the
+ * worker can hydrate the watermark inline without changing its fire-and-forget
+ * control flow. Implementations should be best-effort; the worker guards every
+ * call, so a throwing store degrades to in-memory-only rather than breaking
+ * extraction.
+ *
+ * @public
+ */
+interface ExtractionCursorStore {
+    /** Last message id extracted through for `conversationId`, or undefined. */
+    get(conversationId: string): string | undefined;
+    /** Persist the last-extracted message id for `conversationId`. */
+    set(conversationId: string, messageId: string): void;
+}
+/** Minimal synchronous key/value surface — satisfied by the SDK's
+ * `PlatformStorage` (web `localStorage`, mobile MMKV). */
+interface SyncKeyValueStore {
+    getItem(key: string): string | null;
+    setItem(key: string, value: string): void;
+}
+/**
+ * Build an {@link ExtractionCursorStore} over a synchronous key/value store
+ * (e.g. the SDK's `PlatformStorage`). Keys are namespaced by `keyPrefix`.
+ * Reads/writes are guarded so a store that throws (quota, private mode) can't
+ * break extraction — the watermark just falls back to in-memory.
+ *
+ * @public
+ */
+declare function createPlatformCursorStore(storage: SyncKeyValueStore, keyPrefix?: string): ExtractionCursorStore;
 interface CreateAutoExtractorOptions {
     retainCtx: RetainContext;
     extract: ExtractFactsOptions;
@@ -6195,6 +6337,15 @@ interface CreateAutoExtractorOptions {
      * sessions in RAM-constrained hosts. Default 200.
      */
     maxTrackedConversations?: number;
+    /**
+     * Durable per-conversation watermark store. When provided, the watermark is
+     * hydrated from it the first time each conversation is touched and written
+     * through on every advance, so extraction resumes exactly after the last
+     * extracted message across process restarts (and, with a process-shared
+     * store, across concurrent sessions). Build via {@link createPlatformCursorStore}.
+     * Omit for in-memory-only (legacy) behavior.
+     */
+    cursorStore?: ExtractionCursorStore;
     /**
      * Entity / memory_entity write context — when provided, each retained
      * candidate's `entities[]` is persisted via `linkMemoryEntitiesOp`,
@@ -6270,6 +6421,24 @@ interface AutoExtractor {
  */
 declare function createAutoExtractor(options: CreateAutoExtractorOptions): AutoExtractor;
 
+/**
+ * Default entry cap for the chunk-vector cache. Entries are keyed by message
+ * id, so this bounds how many messages' decrypted chunk vectors stay resident.
+ * Matches the vault cache default (#705).
+ */
+declare const DEFAULT_CHUNK_CACHE_SIZE = 5000;
+/**
+ * Create a {@link ChunkVectorCache} backed by an LRU. Pass the returned cache
+ * into `recall()` (via `RecallContext.chunkCache`) so the chunk lane skips the
+ * per-query decrypt + JSON.parse of every message's chunk vectors on warm
+ * entries. Stale entries self-invalidate on `updated_at` mismatch, so no
+ * explicit invalidation is required beyond clearing it on an encryption-key
+ * reset (wallet switch), mirroring the vault embedding cache.
+ *
+ * @public
+ */
+declare function createChunkVectorCache(maxSize?: number): ChunkVectorCache;
+
 /**
  * Unified Recall API — single retrieval surface above vault + engine.
  *
@@ -6422,6 +6591,46 @@ interface ReflectResult {
  */
 declare function reflect(query: string, ctx: RecallContext, options: ReflectOptions): Promise<ReflectResult>;
 
+/**
+ * Cross-encoder reranker — local inference via @huggingface/transformers.
+ *
+ * Mirrors Hindsight's default reranker (`cross-encoder/ms-marco-MiniLM-L-6-v2`)
+ * for parity with their reference pipeline. Runs fully on-device — no API
+ * key, no network call, works offline. Model file ~25MB; lazy-loaded and
+ * cached for the lifetime of the process.
+ *
+ * Implementation note: transformers.js v3 doesn't expose a "text-ranking"
+ * pipeline, so we drive the cross-encoder directly via the tokenizer +
+ * sequence-classification model. The model outputs a single relevance
+ * logit per (query, doc) pair; we sigmoid it into [0, 1].
+ */
+/**
+ * Thrown when the cross-encoder cannot run because its optional peer
+ * dependency (`@huggingface/transformers`) isn't installed — the expected
+ * state on React Native, where the package isn't part of the bundle. Callers
+ * treat this as "reranker unavailable, degrade to the fused ranking" rather
+ * than a transient error worth warning about on every recall.
+ *
+ * @public
+ */
+declare class RerankerUnavailableError extends Error {
+    /** The underlying import failure, kept for debugging. */
+    readonly reason: unknown;
+    constructor(reason: unknown);
+}
+/**
+ * Whether the cross-encoder reranker can run in this environment.
+ *
+ * - `true` — the model loaded successfully at least once.
+ * - `false` — the optional `@huggingface/transformers` dependency is missing
+ *   (e.g. React Native); reranking is permanently degraded to the fused
+ *   ranking and no further load attempts are made.
+ * - `undefined` — no rerank has been attempted yet, so availability is unknown.
+ *
+ * @public
+ */
+declare function isRerankerAvailable(): boolean | undefined;
+
 /**
  * Standalone topic (entity) extraction for EXISTING memories.
  *
@@ -7522,4 +7731,4 @@ declare function sanitizeQuotes(s: string): string;
 declare function renderTemplate(value: string | null | undefined, template: string): string | null;
 declare function buildSystemPrompt(ctx: PromptContext, options?: BuildPromptOptions): BuildPromptResult;
 
-export { type AnalyzeStyleOptions, type AnalyzeStyleResult, type ApiResponse, type ApiType, type AutoExecutedToolResult, type AutoExtractMessage, type AutoExtractor, type Budget, type BuildPromptOptions, type BuildPromptResult, type CachedServerTools, type ChatCompletionUsage, type ChatRole, type ChunkSearchResult, type ChunkingOptions, type ConnectorGuidance, type ConnectorGuidanceInput, type ConsolidationFallbackReason, Conversation, ConversationSummary, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateMediaOptions, type CreateMessageOptions, type CreateModelPreferenceOptions, type CreateProjectOptions, type CreateUserPreferenceOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_ANALYSIS_PROMPT, DEFAULT_ANALYZE_AFTER_MESSAGES, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DOCUMENT_PATTERN, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_MAX_INPUT_MESSAGES, DEFAULT_MAX_PROFILE_LENGTH, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_MIN_MESSAGES, DEFAULT_PERSONALITY_SETTINGS, DEFAULT_PROMPT_TEMPLATES, DEFAULT_REFRESH_EVERY_MESSAGES, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, DEFAULT_SUMMARY_MIN_WINDOW_MESSAGES, DEFAULT_SUMMARY_MODEL, DEFAULT_SUMMARY_TOKEN_THRESHOLD, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type EmbeddedWalletSignerFn, ExcelProcessor, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type FactType, type FileMetadata, type FileProcessor, type FileTypeQuery, type FileWithData, type FlushResult, INFERENCE_ID_HEADER, type LazyStoredConversation, type LlmTokenUsage, MAX_MESSAGES_PER_SUMMARIZATION, Media, type MediaDimensions, type MediaFilterOptions, type MediaMetadata, type MediaOperationsContext, type MediaRole, type MediaType, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, Message, type MessageChunk, type MessageFeedback, type ModelCallEndEvent, type ModelCallStartEvent, ModelPreference, type OperationExecutor, type ParsedServerToolsResponse, PdfProcessor, type PersonalitySettings, type PersonalitySliders, type PersonalityStyle, type PgClientLike, type PgPoolLike, type PlatformStorage, type PortalLlmAuth, PostgreSQLAdapter, type PostgreSQLAdapterOptions, type PreprocessingOptions, type PreprocessingResult, type ProcessedFileResult, ProcessorRegistry, type ProfileUpdate, Project, type ProjectOperationsContext, type PromptContext, type PromptPreProcessor, type PromptPreProcessorContext, type PromptSection, type PromptTemplates, type QueueEncryptionContext, QueueManager, type QueueStatus, type QueuedOperation, type QueuedOperationType, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankedMemory, type RecallContext, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type ReflectOptions, type ReflectResult, type ResumeStreamOptions, type ResumeStreamResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, type RunEndEvent, type RunErrorEvent, type RunHooks, type RunStartEvent, type RunToolLoopOptions, type RunToolLoopResult, SDK_SCHEMA_VERSION, SERVER_TOOL_DEPENDENCY_SETS, SLIDER_CONFIG, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type SearchSource, type SelectServerToolsForPromptOptions, type ServerTool, type ServerToolCallEvent, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, type SettingsStorageOperationsContext, type SignMessageFn, type StepFinishEvent, type StockPriceClassification, type StockPricePreProcessorOptions, type StorageOperationsContext, type StoredConversation, type StoredConversationSummary, type StoredFileWithContext, type StoredMedia, type StoredMessage, type StoredMessageWithSimilarity, type StoredModelPreference, type StoredProject, type StoredUserPreference, type StoredVaultFolder, type StoredVaultMemory, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type StreamSmoothingConfig, type StreamingTransport, type StreamingTransportOptions, type StreamingTransportResult, type StyleAnalysisSchedule, TOOL_CATALOG, TOPIC_EXTRACTION_BATCH_SIZE, type TextChunk, TextProcessor, type ToolConfig, type ToolExecutor, type ToolMatchOptions, type ToolMatchResult, type ToolSet, type ToolUseEndEvent, type ToolUseStartEvent, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TurnCompleteEvent, type TurnSkippedEvent, type UpdateMediaOptions, type UpdateModelPreferenceOptions, type UpdateProjectOptions, type UpdateUserPreferenceOptions, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, UserPreference, type UserPreferencesStorageOperationsContext, type VaultEmbeddingCache, VaultFolder, type VaultFolderOperationsContext, VaultMemory, type VaultMemoryOperationsContext, type VaultSaveOperation, type VaultSearchResult, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, WordProcessor, ZipProcessor, type ZipProcessorOptions, activatedToolSetNames, analyzeStyle, buildConnectorGuidance, buildDeniedToolsRider, buildSystemPrompt, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, cleanupConversationSummary, clearAllEncryptionKeys, clearAllEncryptionState, clearAllKeyPairs, clearEncryptionKey, clearKeyPair, clearLazyTitleCache, clearMessagesOp, clearServerToolsCache, composeHooks, createAutoExtractor, createConversationOp, createCryptoPricePreProcessor, createMediaBatchOp, createMediaOp, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageOp, createProjectOp, createRecallTool, createServerToolsFilter, createStockPricePreProcessor, createSummaryContext, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decryptConversationTitle, decryptData, decryptDataBatch, decryptDataBytes, decryptDataWithKey, defaultServerToolsFilter, deleteAllVaultMemoriesForUserOp, deleteConversationOp, deleteConversationSummaryOp, deleteMediaByConversationOp, deleteMediaByMessageOp, deleteMediaOp, deleteModelPreferenceOp, deleteProjectOp, deleteUserPreferenceOp, deleteVaultFolderOp, deleteVaultMemoryOp, eagerEmbedContent, embedAllMessages, embedMessage, encryptData, encryptDataBatch, encryptDataWithKey, ensureDefaultFoldersOp, estimateMessagesTokens, estimateTokens, exportPublicKey, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, findMatchingTools, generateConversationId, generateEmbedding, generateEmbeddings, generateMediaId, generateProjectId, getAIGeneratedMediaOp, getAllFilesOp, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getAudioOp, getCachedServerTools, getConversationOp, getConversationSummaryOp, getConversationsByProjectLazyOp, getConversationsByProjectOp, getConversationsLazyOp, getConversationsOp, getDocumentsOp, getEncryptionKey, getImagesOp, getMediaByConversationOp, getMediaByIdOp, getMediaByIdsOp, getMediaByMessageOp, getMediaByModelOp, getMediaByRoleOp, getMediaBySourceUrlOp, getMediaByTypeOp, getMediaCountOp, getMediaCountsByTypeOp, getMediaOp, getMediaTypeFromMime, getMemoriesNeedingTopicExtractionOp, getMessagesOp, getModelPreferenceOp, getProjectConversationCountOp, getProjectConversationsOp, getProjectOp, getProjectsOp, getRecentMediaOp, getServerTools, getSupportedFileTypes, getToolsChecksum, getUnfiledVaultMemoriesOp, getUserPreferenceOp, getUserUploadedMediaOp, getVaultFolderMemoryCountOp, getVaultMemoryOp, getVideosOp, hardDeleteMediaOp, hasEncryptionKey, hasKeyPair, isSupportedFile, isSupportedMediaType, makeSyntheticStoredConversation, makeSyntheticStoredMessage, maybeSummarizeHistory, mediaToStored, migrateFromModelPreferencesOp, moveMemoriesToFolderOp, onKeyAvailable, preEmbedVaultMemories, preprocessFiles, progressiveSummarize, projectToStored, queueManager, recall, reflect, relinkMisclassifiedVideosOp, renderTemplate, requestEncryptionKey, requestKeyPair, resumeStream, retain, runToolLoop, sanitizeQuotes, schemaToCreateSQL, sdkMigrations, sdkModelClasses, sdkSchema, searchChunksOp, searchMediaOp, searchMessagesOp, searchVaultMemories, selectServerToolsForPrompt, serverPlatformStorage, setModelPreferenceOp, setUserPreferenceOp, shouldAnalyzeStyle, shouldChunkMessage, shouldRefreshTools, splitMessagesAtThreshold, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summaryToSystemMessage, toolSetSystemPrompts, updateConversationProjectOp, updateConversationTitleOp, updateMediaMessageIdBatchOp, updateMediaOp, updateMessageChunksOp, updateMessageEmbeddingOp, updateMessageErrorOp, updateMessageFeedbackOp, updateModelsOp, updatePersonalityOp, updateProfileOp, updateProjectNameOp, updateProjectOp, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertConversationSummaryOp, upsertMessageOp };
+export { type AnalyzeStyleOptions, type AnalyzeStyleResult, type ApiResponse, type ApiType, type AutoExecutedToolResult, type AutoExtractMessage, type AutoExtractor, type Budget, type BuildPromptOptions, type BuildPromptResult, type CachedChunkVectors, type CachedServerTools, type ChatCompletionUsage, type ChatRole, type ChunkSearchResult, type ChunkVectorCache, type ChunkingOptions, type ConnectorGuidance, type ConnectorGuidanceInput, type ConsolidationFallbackReason, Conversation, ConversationSummary, type CreateAutoExtractorOptions, type CreateConversationOptions, type CreateMediaOptions, type CreateMessageOptions, type CreateModelPreferenceOptions, type CreateProjectOptions, type CreateUserPreferenceOptions, type CreateVaultFolderOptions, type CreateVaultMemoryOptions, type CryptoPriceClassification, type CryptoPricePreProcessorOptions, DEFAULT_ANALYSIS_PROMPT, DEFAULT_ANALYZE_AFTER_MESSAGES, DEFAULT_CACHE_EXPIRATION_MS, DEFAULT_CHUNK_CACHE_SIZE, DEFAULT_CHUNK_OVERLAP, DEFAULT_CHUNK_SIZE, DEFAULT_DOCUMENT_PATTERN, DEFAULT_EXCLUDED_SERVER_TOOLS, DEFAULT_MAX_INPUT_MESSAGES, DEFAULT_MAX_PROFILE_LENGTH, DEFAULT_MIN_CHUNK_SIZE, DEFAULT_MIN_MESSAGES, DEFAULT_PERSONALITY_SETTINGS, DEFAULT_PROMPT_TEMPLATES, DEFAULT_REFRESH_EVERY_MESSAGES, DEFAULT_SERVER_TOOLS_MATCH_OPTIONS, DEFAULT_SUMMARY_MIN_WINDOW_MESSAGES, DEFAULT_SUMMARY_MODEL, DEFAULT_SUMMARY_TOKEN_THRESHOLD, DEFAULT_VAULT_CACHE_SIZE, DatabaseManager, type DatabaseManagerLogger, type DatabaseManagerOptions, type EmbeddedWalletSignerFn, ExcelProcessor, type ExtractFactsOptions, type ExtractOutcome, type ExtractedCandidate, type ExtractedEntity, type ExtractionCursorStore, type FactType, type FileMetadata, type FileProcessor, type FileTypeQuery, type FileWithData, type FlushResult, INFERENCE_ID_HEADER, type LazyStoredConversation, type LlmTokenUsage, MAX_MESSAGES_PER_SUMMARIZATION, Media, type MediaDimensions, type MediaFilterOptions, type MediaMetadata, type MediaOperationsContext, type MediaRole, type MediaType, type MemoriesNeedingTopicExtraction, type EmbeddingOptions as MemoryEngineEmbeddingOptions, type MemoryEngineResult, type MemoryEngineSearchOptions, type MemoryExtractedEvent, type MemoryKind, type MemoryVaultSearchOptions, type MemoryVaultToolOptions, Message, type MessageChunk, type MessageFeedback, type ModelCallEndEvent, type ModelCallStartEvent, ModelPreference, type OperationExecutor, type ParsedServerToolsResponse, PdfProcessor, type PersonalitySettings, type PersonalitySliders, type PersonalityStyle, type PgClientLike, type PgPoolLike, type PlatformStorage, type PortalLlmAuth, PostgreSQLAdapter, type PostgreSQLAdapterOptions, type PreprocessingOptions, type PreprocessingResult, type ProcessedFileResult, ProcessorRegistry, type ProfileUpdate, Project, type ProjectOperationsContext, type PromptContext, type PromptPreProcessor, type PromptPreProcessorContext, type PromptSection, type PromptTemplates, type QueueEncryptionContext, QueueManager, type QueueStatus, type QueuedOperation, type QueuedOperationType, RECALL_MAX_LIMIT, RECALL_TOOL_NAME, type RankedMemory, type RecallContext, type RecallDegradation, type RecallDiagnostics, type RecallOptions, type RecallResult, type RecallToolCallbacks, type RecallToolOptions, type RecencyOptions, type ReflectOptions, type ReflectResult, RerankerUnavailableError, type ResumeStreamOptions, type ResumeStreamResult, type RetainAction, type RetainContext, type RetainOptions, type RetainResult, type RetainSource, type RunEndEvent, type RunErrorEvent, type RunHooks, type RunStartEvent, type RunToolLoopOptions, type RunToolLoopResult, SDK_SCHEMA_VERSION, SERVER_TOOL_DEPENDENCY_SETS, SLIDER_CONFIG, STREAM_RESUMABLE_HEADER, type ScoreBreakdown, type SearchSource, type SelectServerToolsForPromptOptions, type ServerTool, type ServerToolCallEvent, type ServerToolsFilterFunction, type ServerToolsOptions, type ServerToolsResponse, type SettingsStorageOperationsContext, type SignMessageFn, type StepFinishEvent, type StockPriceClassification, type StockPricePreProcessorOptions, type StorageOperationsContext, type StoredConversation, type StoredConversationSummary, type StoredFileWithContext, type StoredMedia, type StoredMessage, type StoredMessageWithSimilarity, type StoredModelPreference, type StoredProject, type StoredUserPreference, type StoredVaultFolder, type StoredVaultMemory, StreamExpiredError, type StreamMetaEvent, type StreamResumeHandle, type StreamSmoothingConfig, type StreamingTransport, type StreamingTransportOptions, type StreamingTransportResult, type StyleAnalysisSchedule, TOOL_CATALOG, TOPIC_EXTRACTION_BATCH_SIZE, type TextChunk, TextProcessor, type ToolConfig, type ToolExecutor, type ToolMatchOptions, type ToolMatchResult, type ToolSet, type ToolUseEndEvent, type ToolUseStartEvent, type TopicExtractOptions, type TopicExtractionInput, type TopicExtractionRunResult, type TurnCompleteEvent, type TurnSkippedEvent, type UpdateMediaOptions, type UpdateModelPreferenceOptions, type UpdateProjectOptions, type UpdateUserPreferenceOptions, type UpdateVaultFolderOptions, type UpdateVaultMemoryOptions, UserPreference, type UserPreferencesStorageOperationsContext, type VaultEmbeddingCache, VaultFolder, type VaultFolderOperationsContext, VaultMemory, type VaultMemoryOperationsContext, type VaultSaveOperation, type VaultSearchResult, WalletPoller, type WeatherClassification, type WeatherPreProcessorOptions, type WebSearchClassification, type WebSearchPreProcessorOptions, WordProcessor, ZipProcessor, type ZipProcessorOptions, activatedToolSetNames, analyzeStyle, buildConnectorGuidance, buildDeniedToolsRider, buildSystemPrompt, chunkAndEmbedAllMessages, chunkAndEmbedMessage, chunkText, classifyCryptoPrice, classifyCryptoPriceBatch, classifyStockPrice, classifyStockPriceBatch, classifyWeather, classifyWeatherBatch, classifyWebSearch, classifyWebSearchBatch, cleanupConversationSummary, clearAllEncryptionKeys, clearAllEncryptionState, clearAllKeyPairs, clearEncryptionKey, clearKeyPair, clearLazyTitleCache, clearMessagesOp, clearServerToolsCache, composeHooks, createAutoExtractor, createChunkVectorCache, createConversationOp, createCryptoPricePreProcessor, createMediaBatchOp, createMediaOp, createMemoryEngineTool, createMemoryVaultSearchTool, createMemoryVaultTool, createMessageOp, createPlatformCursorStore, createProjectOp, createRecallTool, createServerToolsFilter, createStockPricePreProcessor, createSummaryContext, createVaultEmbeddingCache, createVaultFolderOp, createVaultMemoriesBatchOp, createVaultMemoryOp, createWeatherPreProcessor, createWebSearchPreProcessor, decryptConversationTitle, decryptData, decryptDataBatch, decryptDataBytes, decryptDataWithKey, defaultServerToolsFilter, deleteAllVaultMemoriesForUserOp, deleteConversationOp, deleteConversationSummaryOp, deleteMediaByConversationOp, deleteMediaByMessageOp, deleteMediaOp, deleteModelPreferenceOp, deleteProjectOp, deleteUserPreferenceOp, deleteVaultFolderOp, deleteVaultMemoryOp, eagerEmbedContent, embedAllMessages, embedMessage, encryptData, encryptDataBatch, encryptDataWithKey, ensureDefaultFoldersOp, estimateMessagesTokens, estimateTokens, exportPublicKey, extractAndLinkEntitiesForMemoriesOp, extractAndRetain, extractEntitiesForMemories, extractFacts, findMatchingTools, generateConversationId, generateEmbedding, generateEmbeddings, generateMediaId, generateProjectId, getAIGeneratedMediaOp, getAllFilesOp, getAllVaultFoldersOp, getAllVaultMemoriesOp, getAllVaultMemoryContentsOp, getAudioOp, getCachedServerTools, getConversationOp, getConversationSummaryOp, getConversationsByProjectLazyOp, getConversationsByProjectOp, getConversationsLazyOp, getConversationsOp, getDocumentsOp, getEncryptionKey, getImagesOp, getMediaByConversationOp, getMediaByIdOp, getMediaByIdsOp, getMediaByMessageOp, getMediaByModelOp, getMediaByRoleOp, getMediaBySourceUrlOp, getMediaByTypeOp, getMediaCountOp, getMediaCountsByTypeOp, getMediaOp, getMediaTypeFromMime, getMemoriesNeedingTopicExtractionOp, getMessagesOp, getModelPreferenceOp, getProjectConversationCountOp, getProjectConversationsOp, getProjectOp, getProjectsOp, getRecentMediaOp, getServerTools, getSupportedFileTypes, getToolsChecksum, getUnfiledVaultMemoriesOp, getUserPreferenceOp, getUserUploadedMediaOp, getVaultFolderMemoryCountOp, getVaultMemoryOp, getVideosOp, hardDeleteMediaOp, hasEncryptionKey, hasKeyPair, isRerankerAvailable, isSupportedFile, isSupportedMediaType, makeSyntheticStoredConversation, makeSyntheticStoredMessage, maybeSummarizeHistory, mediaToStored, migrateFromModelPreferencesOp, moveMemoriesToFolderOp, onKeyAvailable, preEmbedVaultMemories, preprocessFiles, progressiveSummarize, projectToStored, queueManager, recall, reflect, relinkMisclassifiedVideosOp, renderTemplate, requestEncryptionKey, requestKeyPair, resumeStream, retain, runToolLoop, sanitizeQuotes, schemaToCreateSQL, sdkMigrations, sdkModelClasses, sdkSchema, searchChunksOp, searchMediaOp, searchMessagesOp, searchVaultMemories, selectServerToolsForPrompt, serverPlatformStorage, setModelPreferenceOp, setUserPreferenceOp, shouldAnalyzeStyle, shouldChunkMessage, shouldRefreshTools, splitMessagesAtThreshold, stampTopicsExtractedAtOp, streamCancelPath, streamReplayPath, summaryToSystemMessage, toolSetSystemPrompts, updateConversationProjectOp, updateConversationTitleOp, updateMediaMessageIdBatchOp, updateMediaOp, updateMessageChunksOp, updateMessageEmbeddingOp, updateMessageErrorOp, updateMessageFeedbackOp, updateModelsOp, updatePersonalityOp, updateProfileOp, updateProjectNameOp, updateProjectOp, updateVaultFolderContextOp, updateVaultFolderOp, updateVaultMemoryEmbeddingOp, updateVaultMemoryOp, upsertConversationSummaryOp, upsertMessageOp };

package/dist/server/index.mjs

@@ -5,6 +5,8 @@ import {
   DEFAULT_MIN_CHUNK_SIZE,
   DEFAULT_VAULT_CACHE_SIZE,
   ENTITY_KINDS,
+  LRUCache,
+  RerankerUnavailableError,
   chunkAndEmbedAllMessages,
   chunkAndEmbedMessage,
   chunkText,
@@ -60,6 +62,7 @@ import {
   hasEncryptionKey,
   hasKeyPair,
   isEncrypted,
+  isRerankerAvailable,
   linkMemoryEntitiesOp,
   makeSyntheticStoredConversation,
   makeSyntheticStoredMessage,
@@ -99,7 +102,7 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-3AWYZRTS.mjs";
+} from "./chunk-3MLG7SUH.mjs";
 import {
   createStreamingDeAnonymizer,
   resolvePiiRedactor
@@ -5187,6 +5190,23 @@ function parseLocalCalendarDay(year, month, day) {
 }
 
 // src/lib/memory/autoExtractWorker.ts
+function createPlatformCursorStore(storage, keyPrefix = "anuma:mem:xcursor:") {
+  return {
+    get(conversationId) {
+      try {
+        return storage.getItem(keyPrefix + conversationId) ?? void 0;
+      } catch {
+        return void 0;
+      }
+    },
+    set(conversationId, messageId) {
+      try {
+        storage.setItem(keyPrefix + conversationId, messageId);
+      } catch {
+      }
+    }
+  };
+}
 var DEFAULT_WINDOW_SIZE = 6;
 var DEFAULT_MAX_WINDOW_SIZE = 20;
 var CONTEXT_OVERLAP = 2;
@@ -5206,12 +5226,24 @@ function createAutoExtractor(options) {
   let inflight = 0;
   let disposed = false;
   const conversations = /* @__PURE__ */ new Map();
+  const cursorStore = options.cursorStore;
+  const hydrate = (state, conversationId) => {
+    if (state.hydrated) return;
+    state.hydrated = true;
+    if (!cursorStore || conversationId === void 0 || state.watermark !== void 0) return;
+    try {
+      const persisted = cursorStore.get(conversationId);
+      if (persisted) state.watermark = persisted;
+    } catch {
+    }
+  };
   const stateFor = (conversationId) => {
     let state = conversations.get(conversationId);
     if (!state) {
       if (conversations.size >= maxTrackedConversations) evictOldestIdle(conversationId);
       state = {};
       conversations.set(conversationId, state);
+      hydrate(state, conversationId);
     }
     return state;
   };
@@ -5224,10 +5256,12 @@ function createAutoExtractor(options) {
     }
   };
   function computeWindow(messages, conversationId) {
-    const lastId = conversations.get(conversationId)?.watermark;
+    const lastId = stateFor(conversationId).watermark;
     const idx = lastId === void 0 ? -1 : messages.findIndex((m) => m.id === lastId);
-    if (idx === -1) return messages.slice(-windowSize);
-    if (idx >= messages.length - 1) return [];
+    if (idx === -1) {
+      return { window: messages.slice(-windowSize), contiguous: messages.length <= windowSize };
+    }
+    if (idx >= messages.length - 1) return { window: [], contiguous: true };
     const start = Math.max(0, idx + 1 - CONTEXT_OVERLAP);
     let window2 = messages.slice(start);
     if (window2.length > maxWindowSize) {
@@ -5236,8 +5270,21 @@ function createAutoExtractor(options) {
       );
       window2 = window2.slice(0, maxWindowSize);
     }
-    return window2;
+    return { window: window2, contiguous: true };
   }
+  const persistCursor = (conversationId, messages, advancedToId, contiguous) => {
+    if (!cursorStore || conversationId === void 0 || !contiguous) return;
+    try {
+      const stored = cursorStore.get(conversationId);
+      if (stored && stored !== advancedToId) {
+        const storedIdx = messages.findIndex((m) => m.id === stored);
+        const newIdx = messages.findIndex((m) => m.id === advancedToId);
+        if (storedIdx !== -1 && newIdx !== -1 && storedIdx >= newIdx) return;
+      }
+      cursorStore.set(conversationId, advancedToId);
+    } catch {
+    }
+  };
   const consolidateBaseUrl = options.consolidate?.baseUrl ?? options.extract.baseUrl;
   const consolidateOptions = options.consolidate ? {
     ...options.extract.apiKey !== void 0 && { apiKey: options.extract.apiKey },
@@ -5283,7 +5330,7 @@ function createAutoExtractor(options) {
     return dispatch(messages, conversationId);
   }
   function dispatch(messages, conversationId) {
-    const window2 = computeWindow(messages, conversationId);
+    const { window: window2, contiguous } = computeWindow(messages, conversationId);
     if (window2.length === 0) {
       options.onSkipped?.({ reason: "no-new-content", conversationId });
       drainPending();
@@ -5308,7 +5355,9 @@ function createAutoExtractor(options) {
             }
           }
         );
-        stateFor(conversationId).watermark = window2[window2.length - 1].id;
+        const advancedTo = window2[window2.length - 1].id;
+        stateFor(conversationId).watermark = advancedTo;
+        persistCursor(conversationId, messages, advancedTo, contiguous);
         for (let i = 0; i < results.length; i++) {
           if (results[i].action === "suppressed") continue;
           options.onMemoryExtracted?.({
@@ -5355,6 +5404,12 @@ function createAutoExtractor(options) {
   };
 }
 
+// src/lib/memory/chunkVectorCache.ts
+var DEFAULT_CHUNK_CACHE_SIZE = 5e3;
+function createChunkVectorCache(maxSize = DEFAULT_CHUNK_CACHE_SIZE) {
+  return new LRUCache(maxSize);
+}
+
 // src/lib/memory/recallTool.ts
 var RECALL_TOOL_NAME = "recall_memory";
 var RECALL_MAX_LIMIT = 50;
@@ -11273,6 +11328,7 @@ export {
   DEFAULT_ANALYSIS_PROMPT,
   DEFAULT_ANALYZE_AFTER_MESSAGES,
   DEFAULT_CACHE_EXPIRATION_MS,
+  DEFAULT_CHUNK_CACHE_SIZE,
   DEFAULT_CHUNK_OVERLAP,
   DEFAULT_CHUNK_SIZE,
   DEFAULT_DOCUMENT_PATTERN,
@@ -11303,6 +11359,7 @@ export {
   QueueManager,
   RECALL_MAX_LIMIT,
   RECALL_TOOL_NAME,
+  RerankerUnavailableError,
   SDK_SCHEMA_VERSION,
   SERVER_TOOL_DEPENDENCY_SETS,
   SLIDER_CONFIG,
@@ -11344,6 +11401,7 @@ export {
   clearServerToolsCache,
   composeHooks,
   createAutoExtractor,
+  createChunkVectorCache,
   createConversationOp,
   createCryptoPricePreProcessor,
   createMediaBatchOp,
@@ -11352,6 +11410,7 @@ export {
   createMemoryVaultSearchTool,
   createMemoryVaultTool,
   createMessageOp,
+  createPlatformCursorStore,
   createProjectOp,
   createRecallTool,
   createServerToolsFilter,
@@ -11448,6 +11507,7 @@ export {
   hardDeleteMediaOp,
   hasEncryptionKey,
   hasKeyPair,
+  isRerankerAvailable,
   isSupportedFile,
   isSupportedMediaType,
   makeSyntheticStoredConversation,

package/dist/server/{recall-EWMI64SA.mjs → recall-MJHGXMHZ.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-3AWYZRTS.mjs";
+} from "./chunk-3MLG7SUH.mjs";
 import "./chunk-AHJT66EO.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "2.3.1",
+  "version": "2.4.1",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",