@anuma/sdk 6.10.0 → 6.11.0

package/dist/expo/index.d.mts

@@ -5732,6 +5732,17 @@ interface RecallDiagnostics {
      * never the cost.
      */
     vaultRowsDecrypted?: number;
+    /**
+     * Rows the fact lane had to re-embed through the portal because their stored
+     * vector was unusable — stale `embedding_model`, wrong dimension, or
+     * unparseable. Absent when the fact lane didn't run.
+     *
+     * Expected to be 0 on a healthy vault: rows are embedded at write time and the
+     * re-embed writes the current model back, so a persistently non-zero value
+     * means the writeback is not sticking and every turn is paying for it. On the
+     * LEGACY read path this batch is uncapped, so it can be the whole vault.
+     */
+    vaultRowsEmbedded?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */
@@ -5744,6 +5755,37 @@ interface RecallDiagnostics {
         prep: number;
         /** Vault fact-lane search (`searchVaultMemoriesWithSize`). */
         factLane: number;
+        /**
+         * The cross-encoder's share of {@link factLane} — wall-clock spent inside
+         * `rerankPairs`. Billed even when the CE threw partway through, because a
+         * rerank that burned three seconds and then failed still cost them.
+         *
+         * A SUBSET of `factLane`, not a sibling: read `factLane - rerank` for
+         * everything else the lane did (query embed, vault read, fused ranking).
+         * 0 when the CE did not run; `reranked` is what distinguishes that from a
+         * rerank that cost nothing.
+         *
+         * Exists because #845 spent three rounds arguing about which stage inside
+         * the fact lane dominated — whole-vault read, then admission window, then
+         * the CE — with one aggregate number for all of them. Every hypothesis was
+         * an inference; this makes the question a query.
+         */
+        rerank: number;
+        /**
+         * The query embed's share of {@link factLane} — one portal round trip.
+         *
+         * Also a SUBSET of `factLane`. Together with `rerank` this makes the lane
+         * decomposable: `factLane - queryEmbed - rerank` is the local work (vault
+         * read, decrypt, BM25/cosine fusion), and each of the three implies a
+         * completely different fix.
+         *
+         * 0 when the lane returned before embedding, which includes every empty
+         * vault — `prepareVaultCandidates` short-circuits first. That is why the
+         * fast `vault_size = 0` population never established a baseline for this
+         * cost, and why the ~850ms floor on the smallest NON-empty vaults had no
+         * attributable owner.
+         */
+        queryEmbed: number;
         /** Chunk-lane search (`searchChunksOp`). */
         chunkLane: number;
         /** Cross-lane RRF fusion + provenance dedup after both lanes. */

package/dist/expo/index.d.ts

@@ -5732,6 +5732,17 @@ interface RecallDiagnostics {
      * never the cost.
      */
     vaultRowsDecrypted?: number;
+    /**
+     * Rows the fact lane had to re-embed through the portal because their stored
+     * vector was unusable — stale `embedding_model`, wrong dimension, or
+     * unparseable. Absent when the fact lane didn't run.
+     *
+     * Expected to be 0 on a healthy vault: rows are embedded at write time and the
+     * re-embed writes the current model back, so a persistently non-zero value
+     * means the writeback is not sticking and every turn is paying for it. On the
+     * LEGACY read path this batch is uncapped, so it can be the whole vault.
+     */
+    vaultRowsEmbedded?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */
@@ -5744,6 +5755,37 @@ interface RecallDiagnostics {
         prep: number;
         /** Vault fact-lane search (`searchVaultMemoriesWithSize`). */
         factLane: number;
+        /**
+         * The cross-encoder's share of {@link factLane} — wall-clock spent inside
+         * `rerankPairs`. Billed even when the CE threw partway through, because a
+         * rerank that burned three seconds and then failed still cost them.
+         *
+         * A SUBSET of `factLane`, not a sibling: read `factLane - rerank` for
+         * everything else the lane did (query embed, vault read, fused ranking).
+         * 0 when the CE did not run; `reranked` is what distinguishes that from a
+         * rerank that cost nothing.
+         *
+         * Exists because #845 spent three rounds arguing about which stage inside
+         * the fact lane dominated — whole-vault read, then admission window, then
+         * the CE — with one aggregate number for all of them. Every hypothesis was
+         * an inference; this makes the question a query.
+         */
+        rerank: number;
+        /**
+         * The query embed's share of {@link factLane} — one portal round trip.
+         *
+         * Also a SUBSET of `factLane`. Together with `rerank` this makes the lane
+         * decomposable: `factLane - queryEmbed - rerank` is the local work (vault
+         * read, decrypt, BM25/cosine fusion), and each of the three implies a
+         * completely different fix.
+         *
+         * 0 when the lane returned before embedding, which includes every empty
+         * vault — `prepareVaultCandidates` short-circuits first. That is why the
+         * fast `vault_size = 0` population never established a baseline for this
+         * cost, and why the ~850ms floor on the smallest NON-empty vaults had no
+         * attributable owner.
+         */
+        queryEmbed: number;
         /** Chunk-lane search (`searchChunksOp`). */
         chunkLane: number;
         /** Cross-lane RRF fusion + provenance dedup after both lanes. */

package/dist/expo/index.mjs

@@ -132,7 +132,7 @@ import {
   upsertMessageOp,
   useEncryption,
   vaultMemoryToStored
-} from "./chunk-4M56MKBV.mjs";
+} from "./chunk-ZO7S3TBT.mjs";
 import {
   PII_PATTERNS,
   PiiRedactor,

package/dist/expo/{recall-WIB7RQQZ.mjs → recall-OALPGL3H.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-4M56MKBV.mjs";
+} from "./chunk-ZO7S3TBT.mjs";
 import "./chunk-CJ32HO6Q.mjs";
 import "./chunk-GYKJOWEO.mjs";
 export {

package/dist/react/{chunk-EJ65X4GI.mjs → chunk-HNQ2ZEFF.mjs}

@@ -5553,9 +5553,9 @@ function recencyMultiplier(updatedAt, opts = {}) {
   const updatedMs = updatedAt.getTime();
   if (!Number.isFinite(updatedMs)) return noDate;
   const now = opts.now ?? /* @__PURE__ */ new Date();
-  const nowMs2 = now.getTime();
-  if (!Number.isFinite(nowMs2)) return noDate;
-  const ageDays = Math.max(0, (nowMs2 - updatedMs) / DAY_MS);
+  const nowMs3 = now.getTime();
+  if (!Number.isFinite(nowMs3)) return noDate;
+  const ageDays = Math.max(0, (nowMs3 - updatedMs) / DAY_MS);
   const perYearDecay = opts.perYearDecay ?? DEFAULT_PER_YEAR_DECAY;
   const floor = opts.floor ?? DEFAULT_FLOOR;
   const decay = perYearDecay * (ageDays / 365);
@@ -5901,6 +5901,7 @@ function createVaultEmbeddingCache(maxSize = DEFAULT_VAULT_CACHE_SIZE) {
 }
 
 // src/lib/memoryVault/searchTool.ts
+var nowMs = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
 var warnedProgrammaticDecomposeIgnored = false;
 function normalizeEventTimeKind(kind) {
   return kind === "point" || kind === "range" || kind === "ongoing" ? kind : null;
@@ -6133,6 +6134,7 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
     const rerankTopN = options.rerankTopN ?? 30;
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = v2Ranked.slice(0, rerankTopN);
+    const ceStart = nowMs();
     try {
       const reranked = await rerankPairs(
         query,
@@ -6160,6 +6162,8 @@ async function rankFusedVaultMemoriesAsync(query, queryEmbedding, items, options
         getLogger().warn("[memory/search] cross-encoder rerank failed; using V2 ranking", err);
       }
       combined = v2Ranked;
+    } finally {
+      if (options.rerankStats) options.rerankStats.ms += nowMs() - ceStart;
     }
   } else {
     combined = v2Ranked;
@@ -6326,6 +6330,7 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
     const ceWeight = options.ceWeight ?? 0.1;
     const headSlice = combined.slice(0, rerankTopN);
     const tailSlice = combined.slice(rerankTopN);
+    const ceStart = nowMs();
     try {
       const reranked = await rerankPairs(
         originalQuery,
@@ -6353,6 +6358,8 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
           err
         );
       }
+    } finally {
+      if (options.rerankStats) options.rerankStats.ms += nowMs() - ceStart;
     }
   }
   if (options?.mmr) {
@@ -6479,11 +6486,13 @@ async function buildProjectedCorpus(query, vaultCtx, embeddingOptions, cache, qu
       rowsDecrypted: 0
     };
   }
+  const queryEmbedStart = nowMs();
   const queryEmbedding = await embedQueryOrDegrade(
     query,
     embeddingOptions,
     opts.onEmbeddingDegraded
   );
+  if (opts.embedStats) opts.embedStats.queryMs += nowMs() - queryEmbedStart;
   const embeddingsDegraded = queryEmbedding.length === 0;
   const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const vectored = [];
@@ -6536,6 +6545,7 @@ async function buildProjectedCorpus(query, vaultCtx, embeddingOptions, cache, qu
     const laneRows = laneFetched.filter((m) => !isEncrypted(m.content));
     if (laneRows.length > 0) {
       let laneVecs;
+      if (opts.embedStats) opts.embedStats.rowsEmbedded += laneRows.length;
       try {
         laneVecs = await generateEmbeddings(
           laneRows.map((m) => m.content),
@@ -6614,6 +6624,10 @@ async function buildProjectedCorpus(query, vaultCtx, embeddingOptions, cache, qu
 async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache, searchOptions) {
   if (!query || typeof query !== "string") {
     return {
+      // Nothing was read and nothing was embedded — the short-circuit fires
+      // before either.
+      queryEmbedMs: 0,
+      rowsEmbedded: 0,
       memories: [],
       embeddedItems: [],
       queryEmbedding: [],
@@ -6646,6 +6660,7 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     embeddingsUnavailable = true;
     embeddingFailure = true;
   };
+  const embedStats = { queryMs: 0, rowsEmbedded: 0 };
   if (searchOptions?.decryptLast) {
     const forceIncludeIds = [
       ...searchOptions.entityRanking ?? [],
@@ -6657,7 +6672,8 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
       admitFloor: ADMIT_FLOOR,
       unembeddedCap: UNEMBEDDED_CAP,
       ...forceIncludeIds.length > 0 && { forceIncludeIds },
-      onEmbeddingDegraded
+      onEmbeddingDegraded,
+      embedStats
     });
     if (corpus.vaultSize === 0) {
       return {
@@ -6668,7 +6684,9 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
         embeddingsUnavailable: false,
         embeddingFailure: false,
         decryptLast: true,
-        rowsDecrypted: corpus.rowsDecrypted
+        rowsDecrypted: corpus.rowsDecrypted,
+        queryEmbedMs: embedStats.queryMs,
+        rowsEmbedded: embedStats.rowsEmbedded
       };
     }
     ({ memories, embeddedItems, queryEmbedding, vaultSize } = corpus);
@@ -6685,7 +6703,9 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
         decryptLast: true,
         // Rows WERE decrypt-attempted even though none came back readable — that
         // is the whole point of reporting attempts rather than successes.
-        rowsDecrypted: corpus.rowsDecrypted
+        rowsDecrypted: corpus.rowsDecrypted,
+        queryEmbedMs: embedStats.queryMs,
+        rowsEmbedded: embedStats.rowsEmbedded
       };
     }
   } else {
@@ -6710,10 +6730,14 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
         embeddingsUnavailable: false,
         embeddingFailure: false,
         decryptLast: false,
-        rowsDecrypted: loaded.length
+        rowsDecrypted: loaded.length,
+        queryEmbedMs: embedStats.queryMs,
+        rowsEmbedded: embedStats.rowsEmbedded
       };
     }
+    const legacyEmbedStart = nowMs();
     queryEmbedding = await embedQueryOrDegrade(query, embeddingOptions, onEmbeddingDegraded);
+    embedStats.queryMs += nowMs() - legacyEmbedStart;
     const currentModel = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
     const uncachedTexts = [];
     const uncachedIndices = [];
@@ -6747,6 +6771,7 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     }
     if (uncachedTexts.length > 0) {
       let newEmbeddings;
+      embedStats.rowsEmbedded += uncachedTexts.length;
       try {
         newEmbeddings = await generateEmbeddings(uncachedTexts, embeddingOptions);
       } catch (err) {
@@ -6790,6 +6815,8 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     embeddingsUnavailable = true;
   }
   return {
+    queryEmbedMs: embedStats.queryMs,
+    rowsEmbedded: embedStats.rowsEmbedded,
     memories,
     embeddedItems,
     queryEmbedding,
@@ -6840,11 +6867,15 @@ async function rankPreparedVaultCandidates(query, prepared, embeddingOptions, se
       results: results2,
       vaultSize: out.vaultSize,
       reranked: out.reranked ?? false,
+      // Read straight off the shared stats object rather than being passed in at
+      // each call site: every return in this function funnels through here, so a
+      // path added later reports the CE bill without having to remember to.
+      rerankMs: rerankStats.ms,
       hadV2Head: out.hadV2Head ?? false,
       embeddingsUnavailable
     };
   };
-  const rerankStats = { applied: false };
+  const rerankStats = { applied: false, ms: 0 };
   const useFusion = searchOptions?.useFusion ?? true;
   const tuning = {
     ...searchOptions?.recencyAlpha !== void 0 && { recencyAlpha: searchOptions.recencyAlpha },
@@ -6963,13 +6994,16 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
       results: [],
       vaultSize: 0,
       reranked: false,
+      rerankMs: 0,
       hadV2Head: false,
       embeddingsUnavailable: false,
       rankedOnCosine: false,
       // No storage read happened, so nothing was decrypted; report the path the
       // caller asked for so this turn is still attributable.
       decryptLast: !!searchOptions?.decryptLast,
-      rowsDecrypted: 0
+      rowsDecrypted: 0,
+      queryEmbedMs: 0,
+      rowsEmbedded: 0
     };
   }
   if (searchOptions?.decompose === "llm" && normalizeSubQueries(searchOptions.subQueries).length < 2 && !warnedProgrammaticDecomposeIgnored) {
@@ -6990,11 +7024,14 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
       results: [],
       vaultSize: 0,
       reranked: false,
+      rerankMs: 0,
       hadV2Head: false,
       embeddingsUnavailable: prepared.embeddingsUnavailable,
       rankedOnCosine: false,
       decryptLast: prepared.decryptLast,
-      rowsDecrypted: prepared.rowsDecrypted
+      rowsDecrypted: prepared.rowsDecrypted,
+      queryEmbedMs: prepared.queryEmbedMs,
+      rowsEmbedded: prepared.rowsEmbedded
     };
   }
   if (prepared.memories.length === 0) {
@@ -7002,11 +7039,14 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
       results: [],
       vaultSize: prepared.vaultSize,
       reranked: false,
+      rerankMs: 0,
       hadV2Head: false,
       embeddingsUnavailable: prepared.embeddingsUnavailable,
       rankedOnCosine: false,
       decryptLast: prepared.decryptLast,
-      rowsDecrypted: prepared.rowsDecrypted
+      rowsDecrypted: prepared.rowsDecrypted,
+      queryEmbedMs: prepared.queryEmbedMs,
+      rowsEmbedded: prepared.rowsEmbedded
     };
   }
   const ranked = await rankPreparedVaultCandidates(
@@ -7019,7 +7059,9 @@ async function searchVaultMemoriesWithSize(query, vaultCtx, embeddingOptions, ca
     ...ranked,
     rankedOnCosine: !ranked.embeddingsUnavailable,
     decryptLast: prepared.decryptLast,
-    rowsDecrypted: prepared.rowsDecrypted
+    rowsDecrypted: prepared.rowsDecrypted,
+    queryEmbedMs: prepared.queryEmbedMs,
+    rowsEmbedded: prepared.rowsEmbedded
   };
 }
 async function searchVaultMemories(query, vaultCtx, embeddingOptions, cache, searchOptions) {
@@ -7124,7 +7166,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             subQueries = decomp.subQueries;
           }
         }
-        const { recall: recall2 } = await import("./recall-FWILZSFH.mjs");
+        const { recall: recall2 } = await import("./recall-F5DBVCU2.mjs");
         let recallDegraded = [];
         const result = await recall2(
           query,
@@ -7664,7 +7706,7 @@ function scoreEventTimeOverlap(memoryStart, memoryEnd, memoryKind, window2) {
 var DEFAULT_LIMIT = 8;
 var DEFAULT_BUDGET = "low";
 var DEFAULT_FACT_MIN_SCORE = 0.1;
-var nowMs = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
+var nowMs2 = () => typeof performance !== "undefined" && typeof performance.now === "function" ? performance.now() : Date.now();
 var DEFAULT_CHUNK_MIN_SCORE = 0.5;
 function flagsForBudget(budget) {
   switch (budget) {
@@ -7697,7 +7739,7 @@ async function recall(query, ctx, options = {}) {
   const flags = flagsForBudget(usedBudget);
   const normalizedFacets = normalizeSubQueries(options.subQueries);
   const subQueries = normalizedFacets.length >= 2 ? normalizedFacets : void 0;
-  const t0 = nowMs();
+  const t0 = nowMs2();
   let prepMs = 0;
   let factLaneMs = 0;
   let chunkLaneMs = 0;
@@ -7706,6 +7748,9 @@ async function recall(query, ctx, options = {}) {
   const chunkResults = [];
   let vaultSize;
   let didRerank = false;
+  let rerankMs = 0;
+  let queryEmbedMs = 0;
+  let vaultRowsEmbedded;
   let hadV2Head = false;
   let embeddingsUnavailable = false;
   let chunkEmbedFailed = false;
@@ -7730,12 +7775,15 @@ async function recall(query, ctx, options = {}) {
       ...vaultSize !== void 0 && { vaultSize },
       ...decryptLastRan !== void 0 && { decryptLast: decryptLastRan },
       ...vaultRowsDecrypted !== void 0 && { vaultRowsDecrypted },
+      ...vaultRowsEmbedded !== void 0 && { vaultRowsEmbedded },
       factCount: factResults.length,
       chunkCount: chunkResults.length,
       timings: {
-        total: nowMs() - t0,
+        total: nowMs2() - t0,
         prep: prepMs,
         factLane: factLaneMs,
+        rerank: rerankMs,
+        queryEmbed: queryEmbedMs,
         chunkLane: chunkLaneMs,
         fuse: fuseMs
       },
@@ -7753,7 +7801,7 @@ async function recall(query, ctx, options = {}) {
   const needsChunkEmbedding = types.includes("chunk") && ctx.storageCtx;
   const wantsTemporal = types.includes("fact") && ctx.vaultCtx;
   const graphRefiner = flags.traverse && options.graphRefine && options.decomposeOptions ? createLlmNeighborRefiner(options.decomposeOptions) : void 0;
-  const prepStart = nowMs();
+  const prepStart = nowMs2();
   const [queryEmbedding, entityRanking, temporalRanking] = await Promise.all([
     // The chunk lane is cosine-only — `searchChunksOp` needs a real vector, and
     // there is no lexical fallback for it — so an embeddings outage must SKIP the
@@ -7791,14 +7839,17 @@ async function recall(query, ctx, options = {}) {
     ),
     wantsTemporal ? safeLane("temporal", () => buildTemporalLaneRanking(query, ctx.vaultCtx, options.now)) : Promise.resolve([])
   ]);
-  prepMs = nowMs() - prepStart;
+  prepMs = nowMs2() - prepStart;
   if (types.includes("fact") && ctx.vaultCtx && ctx.vaultCache) {
-    const factStart = nowMs();
+    const factStart = nowMs2();
     const vaultMinScore = options.minScore ?? DEFAULT_FACT_MIN_SCORE;
     const {
       results,
       vaultSize: size,
       reranked,
+      rerankMs: factRerankMs,
+      queryEmbedMs: factQueryEmbedMs,
+      rowsEmbedded: factRowsEmbedded,
       hadV2Head: v2Head,
       embeddingsUnavailable: factEmbeddingsUnavailable,
       rankedOnCosine: factRankedOnCosine,
@@ -7857,16 +7908,19 @@ async function recall(query, ctx, options = {}) {
     );
     vaultSize = size;
     didRerank = reranked;
+    rerankMs = factRerankMs;
+    queryEmbedMs = factQueryEmbedMs;
+    vaultRowsEmbedded = factRowsEmbedded;
     hadV2Head = v2Head;
     if (factEmbeddingsUnavailable) embeddingsUnavailable = true;
     factLaneRankedOnCosine = factRankedOnCosine;
     decryptLastRan = factDecryptLast;
     vaultRowsDecrypted = factRowsDecrypted;
-    factLaneMs = nowMs() - factStart;
+    factLaneMs = nowMs2() - factStart;
   }
   if (chunkEmbedFailed && !factLaneRankedOnCosine) embeddingsUnavailable = true;
   if (types.includes("chunk") && ctx.storageCtx && queryEmbedding) {
-    const chunkStart = nowMs();
+    const chunkStart = nowMs2();
     const chunkMinScore = options.minScore ?? DEFAULT_CHUNK_MIN_SCORE;
     const results = await searchChunksOp(ctx.storageCtx, queryEmbedding, {
       limit: types.includes("fact") ? Math.max(limit * 2, 16) : limit,
@@ -7883,9 +7937,9 @@ async function recall(query, ctx, options = {}) {
         (r) => r.chunkText.trim()
       )
     );
-    chunkLaneMs = nowMs() - chunkStart;
+    chunkLaneMs = nowMs2() - chunkStart;
   }
-  const fuseStart = nowMs();
+  const fuseStart = nowMs2();
   if (types.length === 1 || factResults.length === 0 || chunkResults.length === 0) {
     const memories2 = [
       ...factResults.map((r) => toFactMemory(r, options.now)),
@@ -7893,7 +7947,7 @@ async function recall(query, ctx, options = {}) {
     ];
     memories2.sort((a, b) => b.score - a.score);
     const candidateCount = factResults.length + chunkResults.length;
-    fuseMs = nowMs() - fuseStart;
+    fuseMs = nowMs2() - fuseStart;
     emitDiagnostics(candidateCount);
     return {
       memories: memories2.slice(0, limit),
@@ -7948,7 +8002,7 @@ async function recall(query, ctx, options = {}) {
     suppressed = next;
     memories = selectWith(suppressed);
   }
-  fuseMs = nowMs() - fuseStart;
+  fuseMs = nowMs2() - fuseStart;
   emitDiagnostics(byId.size);
   return {
     memories,