@anuma/sdk 7.0.1 → 7.0.2

package/dist/index.mjs

@@ -1973,13 +1973,35 @@ async function generateEmbedding(text2, options) {
   if (total === void 0) return embedOne(text2, options, void 0);
   return withDeadline((signal) => embedOne(text2, options, signal), total, "embedding");
 }
+var inFlightByCache = /* @__PURE__ */ new WeakMap();
 async function embedOne(text2, options, outer) {
-  const { baseUrl = BASE_URL, model, cache } = options;
-  const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
-  if (cache) {
-    const cached = cache.get(text2);
-    if (cached) return Array.from(cached);
+  const { cache } = options;
+  if (!cache) return Array.from(await requestEmbedding(text2, options, outer));
+  const cached = cache.get(text2);
+  if (cached) return Array.from(cached);
+  const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
+  inFlightByCache.set(cache, inFlight);
+  const pending = inFlight.get(text2);
+  if (pending) {
+    try {
+      return Array.from(await pending);
+    } catch (err) {
+      if (outer?.aborted) throw err;
+    }
+  }
+  const request = requestEmbedding(text2, options, outer);
+  inFlight.set(text2, request);
+  try {
+    const embedding = await request;
+    cache.set(text2, embedding);
+    return Array.from(embedding);
+  } finally {
+    if (inFlight.get(text2) === request) inFlight.delete(text2);
   }
+}
+async function requestEmbedding(text2, options, outer) {
+  const { baseUrl = BASE_URL, model } = options;
+  const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const headers = {
     ...await resolveAuthHeaders(options, outer),
     ...taskTypeHeader(options.taskType)
@@ -2014,11 +2036,7 @@ async function embedOne(text2, options, outer) {
       totalTokens: response.data.usage.total_tokens ?? 0
     });
   }
-  const f32Embedding = Float32Array.from(embedding);
-  if (cache) {
-    cache.set(text2, f32Embedding);
-  }
-  return Array.from(f32Embedding);
+  return Float32Array.from(embedding);
 }
 var DEFAULT_EMBEDDING_BATCH_SIZE = 100;
 var DEFAULT_EMBEDDING_BATCH_CONCURRENCY = 3;

package/dist/memory/context.d.mts

@@ -299,6 +299,8 @@ interface RankedMemory {
     updatedAt: Date;
 }
 interface RecallOptions {
+    /** Embedding cache shared with the rest of the turn, so the query is embedded once across lanes and the send. */
+    embeddingCache?: Map<string, Float32Array>;
     /** Which kinds to search. Default: ['fact']. */
     types?: MemoryKind[];
     /** Max items returned. Default: 8. */
@@ -554,6 +556,12 @@ interface RecallDiagnostics {
      * LEGACY read path this batch is uncapped, so it can be the whole vault.
      */
     vaultRowsEmbedded?: number;
+    /** Embedded messages the chunk lane considered; absent when the lane did not run. */
+    chunkRowsScanned?: number;
+    /** Messages whose vectors the chunk lane read from storage on a cache miss. */
+    chunkRowsFetched?: number;
+    /** Messages the chunk lane scored from the vector cache. */
+    chunkCacheHits?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */

package/dist/memory/context.d.ts

@@ -299,6 +299,8 @@ interface RankedMemory {
     updatedAt: Date;
 }
 interface RecallOptions {
+    /** Embedding cache shared with the rest of the turn, so the query is embedded once across lanes and the send. */
+    embeddingCache?: Map<string, Float32Array>;
     /** Which kinds to search. Default: ['fact']. */
     types?: MemoryKind[];
     /** Max items returned. Default: 8. */
@@ -554,6 +556,12 @@ interface RecallDiagnostics {
      * LEGACY read path this batch is uncapped, so it can be the whole vault.
      */
     vaultRowsEmbedded?: number;
+    /** Embedded messages the chunk lane considered; absent when the lane did not run. */
+    chunkRowsScanned?: number;
+    /** Messages whose vectors the chunk lane read from storage on a cache miss. */
+    chunkRowsFetched?: number;
+    /** Messages the chunk lane scored from the vector cache. */
+    chunkCacheHits?: number;
     /** Facts the fact lane returned (post-dedupe, pre-fusion). */
     factCount: number;
     /** Chunks the chunk lane returned (post-dedupe, pre-fusion). */