@anuma/sdk 6.10.0 → 6.12.0

package/dist/react/{chunk-EJ65X4GI.mjs → chunk-6IBFFWCE.mjs}

@@ -5439,6 +5439,9 @@ async function decomposeQuery(query, options) {
       ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
       model: options.model ?? DEFAULT_MODEL,
       systemPrompt: SYSTEM_PROMPT,
+      // The query itself rides the user turn below, so SYSTEM_PROMPT is the whole
+      // fixed instruction — which is what lets the portal own it outright.
+      taskType: "memory_decompose",
       // Wrap the bare query in a task framing so the model treats it as
       // input to classify, not as a question to answer conversationally.
       // Bare-query inputs caused Anthropic models to respond with prose
@@ -5553,9 +5556,9 @@ 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 +5904,7 @@ 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 +6137,7 @@ 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 +6165,8 @@ 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 +6333,7 @@ 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 +6361,8 @@ async function rankComposite(originalQuery, originalQueryEmbedding, subQueries,
           err
         );
       }
+    } finally {
+      if (options.rerankStats) options.rerankStats.ms += nowMs() - ceStart;
     }
   }
   if (options?.mmr) {
@@ -6479,11 +6489,13 @@ 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 +6548,7 @@ 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 +6627,10 @@ 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 +6663,7 @@ 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 +6675,8 @@ 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 +6687,9 @@ 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 +6706,9 @@ 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 +6733,14 @@ 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 +6774,7 @@ 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 +6818,8 @@ async function prepareVaultCandidates(query, vaultCtx, embeddingOptions, cache,
     embeddingsUnavailable = true;
   }
   return {
+    queryEmbedMs: embedStats.queryMs,
+    rowsEmbedded: embedStats.rowsEmbedded,
     memories,
     embeddedItems,
     queryEmbedding,
@@ -6840,11 +6870,15 @@ 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 +6997,16 @@ 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 +7027,14 @@ 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 +7042,14 @@ 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 +7062,9 @@ 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 +7169,7 @@ function createMemoryVaultSearchTool(vaultCtx, embeddingOptions, cache, searchOp
             subQueries = decomp.subQueries;
           }
         }
-        const { recall: recall2 } = await import("./recall-FWILZSFH.mjs");
+        const { recall: recall2 } = await import("./recall-QUWN7DJ7.mjs");
         let recallDegraded = [];
         const result = await recall2(
           query,
@@ -7377,29 +7422,29 @@ async function traverseGraphLane(query, entityCtx, options = {}) {
 var DEFAULT_REFINER_MODEL = "inclusionai/ling-2.6-flash";
 var DEFAULT_REFINER_ATTEMPTS = 1;
 var DEFAULT_REFINER_TOTAL_TIMEOUT_MS = 8e3;
+var REFINER_SYSTEM_PROMPT = `You help a memory-retrieval system decide which related topics to explore.
+
+Given a user's question and a numbered list of candidate topics/entities linked to memories found so far, pick the ones most likely to lead to memories that help ANSWER the question. Never return more than the maximum the user turn asks for. Prefer topics semantically related to the question; ignore incidental ones.
+
+Output strict JSON, no prose: { "expand": [<entity names to expand, verbatim from the list>] }`;
 function createLlmNeighborRefiner(options) {
   return {
     async refine(query, candidates, limit) {
       if (candidates.length === 0) return [];
       const numbered = candidates.map((name, i) => `[${i + 1}] ${name}`).join("\n");
-      const systemPrompt = `You help a memory-retrieval system decide which related topics to explore.
-
-Given a user's question and a numbered list of candidate topics/entities linked to memories found so far, pick the ones most likely to lead to memories that help ANSWER the question. Choose at most ${limit}. Prefer topics semantically related to the question; ignore incidental ones.
-
-Output strict JSON, no prose: { "expand": [<entity names to expand, verbatim from the list>] }`;
       const parsed = await callPortalJsonCompletion({
         ...options.apiKey !== void 0 && { apiKey: options.apiKey },
         ...options.getToken !== void 0 && { getToken: options.getToken },
         ...options.baseUrl !== void 0 && { baseUrl: options.baseUrl },
         taskType: "memory_graph",
         model: options.model ?? DEFAULT_REFINER_MODEL,
-        systemPrompt,
+        systemPrompt: REFINER_SYSTEM_PROMPT,
         userMessage: `Question: ${query}
 
 Candidate topics:
 ${numbered}
 
-Which should be expanded?`,
+Which should be expanded? Choose at most ${limit}.`,
         tag: "memory/graph-refine",
         maxAttempts: options.maxAttempts ?? DEFAULT_REFINER_ATTEMPTS,
         totalTimeoutMs: options.totalTimeoutMs ?? DEFAULT_REFINER_TOTAL_TIMEOUT_MS,
@@ -7664,7 +7709,7 @@ 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 +7742,7 @@ 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 +7751,9 @@ 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 +7778,15 @@ 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 +7804,7 @@ 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 +7842,17 @@ 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 +7911,19 @@ 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 +7940,9 @@ 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 +7950,7 @@ 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 +8005,7 @@ async function recall(query, ctx, options = {}) {
     suppressed = next;
     memories = selectWith(suppressed);
   }
-  fuseMs = nowMs() - fuseStart;
+  fuseMs = nowMs2() - fuseStart;
   emitDiagnostics(byId.size);
   return {
     memories,