@anuma/sdk 7.0.3-next.20261014212810 → 7.0.4

package/dist/expo/index.mjs

@@ -2158,8 +2158,9 @@ async function updateConversationPinnedOp(ctx, id, pinned) {
   }
   return false;
 }
-async function getMessagesOp(ctx, convId) {
-  const results = await ctx.messagesCollection.query(Q.where("conversation_id", convId), Q.sortBy("message_id", Q.asc)).unsafeFetchRaw();
+async function getMessagesOp(ctx, convId, opts) {
+  const fetched = await ctx.messagesCollection.query(Q.where("conversation_id", convId), Q.sortBy("message_id", Q.asc)).unsafeFetchRaw();
+  const results = opts?.rawFilter ? fetched.filter(opts.rawFilter) : fetched;
   return Promise.all(
     results.map(
       (raw) => messageRawToStored(raw, ctx.walletAddress, ctx.signMessage, ctx.embeddedWalletSigner)
@@ -3915,22 +3916,6 @@ var postApiV1Embeddings = (options) => (options.client ?? client).post({
 });
 var getApiV1Models = (options) => (options?.client ?? client).get({ url: "/api/v1/models", ...options });
 
-// src/lib/memoryEngine/embeddingCacheState.ts
-var EMBEDDING_CACHE_STATE_KEY = /* @__PURE__ */ Symbol.for("@anuma/sdk/embedding-cache/v1");
-var cacheGlobal = globalThis;
-var fallbackState = {
-  scopedViews: /* @__PURE__ */ new WeakMap(),
-  inFlightByCache: /* @__PURE__ */ new WeakMap(),
-  cachedEmbeddingModels: /* @__PURE__ */ new WeakMap()
-};
-function getEmbeddingCacheState() {
-  try {
-    return cacheGlobal[EMBEDDING_CACHE_STATE_KEY] ?? (cacheGlobal[EMBEDDING_CACHE_STATE_KEY] = fallbackState);
-  } catch {
-    return fallbackState;
-  }
-}
-
 // src/lib/memoryEngine/generate.ts
 var EMBED_MAX_ATTEMPTS = 4;
 var DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS = 15e3;
@@ -4022,18 +4007,15 @@ async function generateEmbedding(text14, options) {
   if (total === void 0) return embedOne(text14, options, void 0);
   return withDeadline((signal) => embedOne(text14, options, signal), total, "embedding");
 }
-var { inFlightByCache, cachedEmbeddingModels } = getEmbeddingCacheState();
+var inFlightByCache = /* @__PURE__ */ new WeakMap();
 async function embedOne(text14, options, outer) {
   const { cache } = options;
   if (!cache) return Array.from(await requestEmbedding(text14, options, outer));
-  const model = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const cached = cache.get(text14);
-  const cachedModel = cached && cachedEmbeddingModels.get(cached);
-  if (cached && (cachedModel === void 0 || cachedModel === model)) return Array.from(cached);
+  if (cached) return Array.from(cached);
   const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
   inFlightByCache.set(cache, inFlight);
-  const requestKey = JSON.stringify([model, text14]);
-  const pending = inFlight.get(requestKey);
+  const pending = inFlight.get(text14);
   if (pending) {
     try {
       return Array.from(await pending);
@@ -4042,14 +4024,13 @@ async function embedOne(text14, options, outer) {
     }
   }
   const request = requestEmbedding(text14, options, outer);
-  inFlight.set(requestKey, request);
+  inFlight.set(text14, request);
   try {
     const embedding2 = await request;
-    cachedEmbeddingModels.set(embedding2, model);
     cache.set(text14, embedding2);
     return Array.from(embedding2);
   } finally {
-    if (inFlight.get(requestKey) === request) inFlight.delete(requestKey);
+    if (inFlight.get(text14) === request) inFlight.delete(text14);
   }
 }
 async function requestEmbedding(text14, options, outer) {
@@ -4132,15 +4113,13 @@ async function embedMany(texts, options, outer) {
   const { baseUrl = BASE_URL, model, batchSize, cache } = options;
   const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const chunkSize = batchSize ?? DEFAULT_EMBEDDING_BATCH_SIZE;
-  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   const results = new Array(texts.length).fill(null);
   const uncachedIndices = [];
   const uncachedTexts = [];
   for (let i = 0; i < texts.length; i++) {
     if (cache) {
       const cached = cache.get(texts[i]);
-      const cachedModel = cached && cachedEmbeddingModels.get(cached);
-      if (cached && (cachedModel === void 0 || cachedModel === embeddingModel)) {
+      if (cached) {
         results[i] = Array.from(cached);
         continue;
       }
@@ -4156,6 +4135,7 @@ async function embedMany(texts, options, outer) {
     ...taskTypeHeader(options.taskType)
   };
   if (outer?.aborted) throw outer.reason;
+  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   let newEmbeddings;
   if (uncachedTexts.length <= chunkSize) {
     newEmbeddings = await generateEmbeddingsBatch(
@@ -4203,7 +4183,6 @@ async function embedMany(texts, options, outer) {
     const f32Embedding = Float32Array.from(newEmbeddings[i]);
     results[uncachedIndices[i]] = Array.from(f32Embedding);
     if (cache) {
-      cachedEmbeddingModels.set(f32Embedding, embeddingModel);
       cache.set(uncachedTexts[i], f32Embedding);
     }
   }
@@ -4328,6 +4307,37 @@ async function chunkAndEmbedMessage(ctx, messageId, options) {
   }));
   return updateMessageChunksOp(ctx, messageId, messageChunks, embeddingModel);
 }
+var MIN_SEALED_EMBEDDING_HEX = 256;
+function rawHasEmbedding(value) {
+  if (Array.isArray(value)) return value.length > 0;
+  if (typeof value !== "string" || value.length === 0) return false;
+  if (isEncrypted(value)) {
+    return value.length - value.indexOf(":", 4) - 1 >= MIN_SEALED_EMBEDDING_HEX;
+  }
+  try {
+    const parsed = JSON.parse(value);
+    return Array.isArray(parsed) && parsed.length > 0;
+  } catch {
+    return false;
+  }
+}
+function sealedVersion(value) {
+  if (typeof value !== "string" || !isEncrypted(value)) return null;
+  return value.startsWith("enc:v2:") ? "v2" : "v3";
+}
+function mayNeedChunkEmbedding(raw, embeddingModel, filter) {
+  const role = raw.role;
+  if (role === "system") return false;
+  if (filter?.roles && !filter.roles.includes(role ?? "")) return false;
+  if (isNonEmbeddableOrigin(raw.origin, filter?.reembedDiscarded)) {
+    return false;
+  }
+  const model = raw.embedding_model;
+  if (typeof model === "string" && model !== embeddingModel) return true;
+  if (rawHasEmbedding(raw.chunks)) return false;
+  if (rawHasEmbedding(raw.vector) && !filter?.rechunkExisting) return false;
+  return true;
+}
 async function chunkAndEmbedAllMessages(ctx, options, filter) {
   const embeddingModel = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const { chunkSize = DEFAULT_CHUNK_SIZE } = options;
@@ -4340,11 +4350,31 @@ async function chunkAndEmbedAllMessages(ctx, options, filter) {
   let considered = 0;
   let sealedRowsSeen = 0;
   let rowsSeen = 0;
+  const keptSealed = /* @__PURE__ */ new Map();
+  const candidatesStillSealed = /* @__PURE__ */ new Map();
+  const droppedSealed = /* @__PURE__ */ new Map();
+  const rawFilter = (raw) => {
+    rowsSeen++;
+    const keep = mayNeedChunkEmbedding(raw, embeddingModel, filter);
+    const version = sealedVersion(raw.content);
+    if (!version) return keep;
+    if (keep) {
+      keptSealed.set(version, (keptSealed.get(version) ?? 0) + 1);
+    } else {
+      const dropped = droppedSealed.get(version);
+      if (dropped) dropped.count++;
+      else droppedSealed.set(version, { count: 1, probeId: raw.id });
+    }
+    return keep;
+  };
   for (const conv of targetConversations) {
-    const messages = await getMessagesOp(ctx, conv.conversationId);
+    const messages = await getMessagesOp(ctx, conv.conversationId, { rawFilter });
     for (const message of messages) {
-      rowsSeen++;
-      if (isEncrypted(message.content)) sealedRowsSeen++;
+      const stillSealed = sealedVersion(message.content);
+      if (stillSealed) {
+        sealedRowsSeen++;
+        candidatesStillSealed.set(stillSealed, (candidatesStillSealed.get(stillSealed) ?? 0) + 1);
+      }
       const isStale = message.embeddingModel !== void 0 && message.embeddingModel !== null && message.embeddingModel !== embeddingModel;
       if (message.chunks && message.chunks.length > 0 && !isStale) continue;
       const hasVector = message.vector && message.vector.length > 0;
@@ -4357,6 +4387,7 @@ async function chunkAndEmbedAllMessages(ctx, options, filter) {
         stillEncrypted++;
         continue;
       }
+      if (isToolResultsRow(message)) continue;
       if (message.content.length < minLength) continue;
       if (shouldChunkMessage(message.content, chunkSize)) {
         longMessages.push({
@@ -4368,6 +4399,12 @@ async function chunkAndEmbedAllMessages(ctx, options, filter) {
       }
     }
   }
+  for (const [version, dropped] of droppedSealed) {
+    const keyOpenedACandidate = (keptSealed.get(version) ?? 0) > (candidatesStillSealed.get(version) ?? 0);
+    if (keyOpenedACandidate) continue;
+    const probe = await getMessageOp(ctx, dropped.probeId);
+    if (!probe || isEncrypted(probe.content)) sealedRowsSeen += dropped.count;
+  }
   if (stillEncrypted > 0 || sealedRowsSeen > 0) {
     getLogger().error(
       "memoryEngine: messages still encrypted (key unavailable?) \u2014 excluded from embedding",
@@ -9573,7 +9610,7 @@ var MaskScopedEmbeddingCache = class extends Map {
     return this;
   }
 };
-var { scopedViews } = getEmbeddingCacheState();
+var scopedViews = /* @__PURE__ */ new WeakMap();
 function maskScopedEmbeddingCache(cache, masked) {
   const views = scopedViews.get(cache) ?? {};
   scopedViews.set(cache, views);
@@ -17737,48 +17774,44 @@ async function recall(query, ctx, options = {}) {
       embeddingOptions,
       ctx.vaultCache,
       searchOptions
-    )))(
-      query,
-      {
-        limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
-        minSimilarity: vaultMinScore,
-        useFusion: true,
-        rerank: flags.rerank,
-        ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
-        ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
-        ...options.rerankLoadTimeoutMs !== void 0 && {
-          rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
-        },
-        ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
-        ...options.recency && { recency: options.recency },
-        ...options.mmr !== void 0 && { mmr: options.mmr },
-        ...options.supersessionBoost !== void 0 && {
-          supersessionBoost: options.supersessionBoost
-        },
-        ...options.supersessionWindow !== void 0 && {
-          supersessionWindow: options.supersessionWindow
-        },
-        ...options.proofCountAlpha !== void 0 && {
-          proofCountAlpha: options.proofCountAlpha
-        },
-        ...options.bm25AdmissionDivisor !== void 0 && {
-          bm25AdmissionDivisor: options.bm25AdmissionDivisor
-        },
-        ...options.rrfK !== void 0 && { rrfK: options.rrfK },
-        ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
-        ...subQueries && { subQueries },
-        ...options.scopes && { scopes: options.scopes },
-        ...options.folderId !== void 0 && { folderId: options.folderId },
-        ...options.factTypes?.length && { factTypes: options.factTypes },
-        ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
-        ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
-        ...entityRanking.length > 0 && { entityRanking },
-        ...temporalRanking.length > 0 && { temporalRanking },
-        ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
-        queryEmbedTotalTimeoutMs
+    )))(query, {
+      limit: types.includes("chunk") ? Math.max(limit * 2, 16) : limit,
+      minSimilarity: vaultMinScore,
+      useFusion: true,
+      rerank: flags.rerank,
+      ...options.rerankTopN !== void 0 && { rerankTopN: options.rerankTopN },
+      ...options.ceWeight !== void 0 && { ceWeight: options.ceWeight },
+      ...options.rerankLoadTimeoutMs !== void 0 && {
+        rerankLoadTimeoutMs: options.rerankLoadTimeoutMs
       },
-      options.embeddingCache
-    );
+      ...options.recencyAlpha !== void 0 && { recencyAlpha: options.recencyAlpha },
+      ...options.recency && { recency: options.recency },
+      ...options.mmr !== void 0 && { mmr: options.mmr },
+      ...options.supersessionBoost !== void 0 && {
+        supersessionBoost: options.supersessionBoost
+      },
+      ...options.supersessionWindow !== void 0 && {
+        supersessionWindow: options.supersessionWindow
+      },
+      ...options.proofCountAlpha !== void 0 && {
+        proofCountAlpha: options.proofCountAlpha
+      },
+      ...options.bm25AdmissionDivisor !== void 0 && {
+        bm25AdmissionDivisor: options.bm25AdmissionDivisor
+      },
+      ...options.rrfK !== void 0 && { rrfK: options.rrfK },
+      ...options.decryptLast !== void 0 && { decryptLast: options.decryptLast },
+      ...subQueries && { subQueries },
+      ...options.scopes && { scopes: options.scopes },
+      ...options.folderId !== void 0 && { folderId: options.folderId },
+      ...options.factTypes?.length && { factTypes: options.factTypes },
+      ...options.memoryIds !== void 0 && { memoryIds: options.memoryIds },
+      ...options.factTypeWeights && { factTypeWeights: options.factTypeWeights },
+      ...entityRanking.length > 0 && { entityRanking },
+      ...temporalRanking.length > 0 && { temporalRanking },
+      ...needsChunkEmbedding && { queryEmbedding: queryEmbedding ?? [] },
+      queryEmbedTotalTimeoutMs
+    });
     factResults.push(
       ...dedupeBy(
         results,
@@ -18917,19 +18950,10 @@ function createLocalMemoryStore(options) {
       retainOptions
     ),
     factSource: {
-      search: (query, searchOptions, embeddingCache) => searchVaultMemoriesWithSize(
+      search: (query, searchOptions) => searchVaultMemoriesWithSize(
         query,
         vaultCtx,
-        {
-          ...options.embeddingOptions,
-          taskType: "memory_recall_embed",
-          ...embeddingCache && !options.embeddingOptions.cache && {
-            cache: maskScopedEmbeddingCache(
-              embeddingCache,
-              Boolean(options.embeddingOptions.maskInput)
-            )
-          }
-        },
+        { ...options.embeddingOptions, taskType: "memory_recall_embed" },
         vaultCache,
         searchOptions
       ),
@@ -19559,7 +19583,7 @@ function stored(row) {
 function createRemoteMemoryPipeline(options) {
   const { persistence, embeddingOptions } = options;
   const model = embeddingOptions.model ?? DEFAULT_API_EMBEDDING_MODEL;
-  const operation = (mode, queryOptions = embeddingOptions) => {
+  const operation = (mode) => {
     const rejectUndecryptable = (failed) => {
       if (!failed.length) return;
       if (mode === "retain")
@@ -19590,7 +19614,7 @@ function createRemoteMemoryPipeline(options) {
       const start = Date.now();
       const budgetMs = search.queryEmbedTotalTimeoutMs ?? 8e3;
       const embed = (texts) => generateEmbeddings(texts, {
-        ...queryOptions,
+        ...embeddingOptions,
         taskType: mode === "retain" ? "memory_retain_embed" : "memory_recall_embed",
         totalTimeoutMs: budgetMs
       });
@@ -19610,11 +19634,7 @@ function createRemoteMemoryPipeline(options) {
         }
       };
       const [primary, facetEmbeddings] = await Promise.all([
-        search.queryEmbedding ?? (mode === "recall" && queryOptions.cache ? generateEmbedding(query, {
-          ...queryOptions,
-          taskType: "memory_recall_embed",
-          totalTimeoutMs: budgetMs
-        }) : embed([query]).then((vectors2) => vectors2[0])),
+        search.queryEmbedding ?? embed([query]).then((vectors2) => vectors2[0]),
         embedFacets()
       ]);
       const queryEmbedding = storedVector(primary);
@@ -19838,17 +19858,15 @@ function createRemoteMemoryPipeline(options) {
   const factSource = {
     graphRanking: options.graphRanking,
     temporalRanking: options.temporalRanking,
-    search: async (text14, searchOptions, embeddingCache) => {
-      const queryOptions = {
-        ...embeddingOptions,
-        taskType: "memory_recall_embed",
-        ...embeddingCache && !embeddingOptions.cache && {
-          cache: maskScopedEmbeddingCache(embeddingCache, Boolean(embeddingOptions.maskInput))
-        }
-      };
-      const { prepare } = operation("recall", queryOptions);
+    search: async (text14, searchOptions) => {
+      const { prepare } = operation("recall");
       const prepared = await prepare(text14, searchOptions);
-      const ranked = await rankPreparedVaultCandidates(text14, prepared, queryOptions, searchOptions);
+      const ranked = await rankPreparedVaultCandidates(
+        text14,
+        prepared,
+        { ...embeddingOptions, taskType: "memory_recall_embed" },
+        searchOptions
+      );
       return {
         ...ranked,
         rankedOnCosine: prepared.memories.length > 0 && !prepared.embeddingsUnavailable,

package/dist/index.cjs

@@ -2089,22 +2089,6 @@ function taskTypeHeader(taskType) {
   return taskType ? { [TASK_TYPE_HEADER]: taskType } : {};
 }
 
-// src/lib/memoryEngine/embeddingCacheState.ts
-var EMBEDDING_CACHE_STATE_KEY = /* @__PURE__ */ Symbol.for("@anuma/sdk/embedding-cache/v1");
-var cacheGlobal = globalThis;
-var fallbackState = {
-  scopedViews: /* @__PURE__ */ new WeakMap(),
-  inFlightByCache: /* @__PURE__ */ new WeakMap(),
-  cachedEmbeddingModels: /* @__PURE__ */ new WeakMap()
-};
-function getEmbeddingCacheState() {
-  try {
-    return cacheGlobal[EMBEDDING_CACHE_STATE_KEY] ?? (cacheGlobal[EMBEDDING_CACHE_STATE_KEY] = fallbackState);
-  } catch {
-    return fallbackState;
-  }
-}
-
 // src/lib/memoryEngine/generate.ts
 var EMBED_MAX_ATTEMPTS = 4;
 var DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS = 15e3;
@@ -2193,18 +2177,15 @@ 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, cachedEmbeddingModels } = getEmbeddingCacheState();
+var inFlightByCache = /* @__PURE__ */ new WeakMap();
 async function embedOne(text2, options, outer) {
   const { cache } = options;
   if (!cache) return Array.from(await requestEmbedding(text2, options, outer));
-  const model = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const cached = cache.get(text2);
-  const cachedModel = cached && cachedEmbeddingModels.get(cached);
-  if (cached && (cachedModel === void 0 || cachedModel === model)) return Array.from(cached);
+  if (cached) return Array.from(cached);
   const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
   inFlightByCache.set(cache, inFlight);
-  const requestKey = JSON.stringify([model, text2]);
-  const pending = inFlight.get(requestKey);
+  const pending = inFlight.get(text2);
   if (pending) {
     try {
       return Array.from(await pending);
@@ -2213,14 +2194,13 @@ async function embedOne(text2, options, outer) {
     }
   }
   const request = requestEmbedding(text2, options, outer);
-  inFlight.set(requestKey, request);
+  inFlight.set(text2, request);
   try {
     const embedding = await request;
-    cachedEmbeddingModels.set(embedding, model);
     cache.set(text2, embedding);
     return Array.from(embedding);
   } finally {
-    if (inFlight.get(requestKey) === request) inFlight.delete(requestKey);
+    if (inFlight.get(text2) === request) inFlight.delete(text2);
   }
 }
 async function requestEmbedding(text2, options, outer) {
@@ -2303,15 +2283,13 @@ async function embedMany(texts, options, outer) {
   const { baseUrl = BASE_URL, model, batchSize, cache } = options;
   const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const chunkSize = batchSize ?? DEFAULT_EMBEDDING_BATCH_SIZE;
-  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   const results = new Array(texts.length).fill(null);
   const uncachedIndices = [];
   const uncachedTexts = [];
   for (let i = 0; i < texts.length; i++) {
     if (cache) {
       const cached = cache.get(texts[i]);
-      const cachedModel = cached && cachedEmbeddingModels.get(cached);
-      if (cached && (cachedModel === void 0 || cachedModel === embeddingModel)) {
+      if (cached) {
         results[i] = Array.from(cached);
         continue;
       }
@@ -2327,6 +2305,7 @@ async function embedMany(texts, options, outer) {
     ...taskTypeHeader(options.taskType)
   };
   if (outer?.aborted) throw outer.reason;
+  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   let newEmbeddings;
   if (uncachedTexts.length <= chunkSize) {
     newEmbeddings = await generateEmbeddingsBatch(
@@ -2374,7 +2353,6 @@ async function embedMany(texts, options, outer) {
     const f32Embedding = Float32Array.from(newEmbeddings[i]);
     results[uncachedIndices[i]] = Array.from(f32Embedding);
     if (cache) {
-      cachedEmbeddingModels.set(f32Embedding, embeddingModel);
       cache.set(uncachedTexts[i], f32Embedding);
     }
   }

package/dist/index.mjs

@@ -1885,22 +1885,6 @@ function taskTypeHeader(taskType) {
   return taskType ? { [TASK_TYPE_HEADER]: taskType } : {};
 }
 
-// src/lib/memoryEngine/embeddingCacheState.ts
-var EMBEDDING_CACHE_STATE_KEY = /* @__PURE__ */ Symbol.for("@anuma/sdk/embedding-cache/v1");
-var cacheGlobal = globalThis;
-var fallbackState = {
-  scopedViews: /* @__PURE__ */ new WeakMap(),
-  inFlightByCache: /* @__PURE__ */ new WeakMap(),
-  cachedEmbeddingModels: /* @__PURE__ */ new WeakMap()
-};
-function getEmbeddingCacheState() {
-  try {
-    return cacheGlobal[EMBEDDING_CACHE_STATE_KEY] ?? (cacheGlobal[EMBEDDING_CACHE_STATE_KEY] = fallbackState);
-  } catch {
-    return fallbackState;
-  }
-}
-
 // src/lib/memoryEngine/generate.ts
 var EMBED_MAX_ATTEMPTS = 4;
 var DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS = 15e3;
@@ -1989,18 +1973,15 @@ 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, cachedEmbeddingModels } = getEmbeddingCacheState();
+var inFlightByCache = /* @__PURE__ */ new WeakMap();
 async function embedOne(text2, options, outer) {
   const { cache } = options;
   if (!cache) return Array.from(await requestEmbedding(text2, options, outer));
-  const model = options.model ?? DEFAULT_API_EMBEDDING_MODEL;
   const cached = cache.get(text2);
-  const cachedModel = cached && cachedEmbeddingModels.get(cached);
-  if (cached && (cachedModel === void 0 || cachedModel === model)) return Array.from(cached);
+  if (cached) return Array.from(cached);
   const inFlight = inFlightByCache.get(cache) ?? /* @__PURE__ */ new Map();
   inFlightByCache.set(cache, inFlight);
-  const requestKey = JSON.stringify([model, text2]);
-  const pending = inFlight.get(requestKey);
+  const pending = inFlight.get(text2);
   if (pending) {
     try {
       return Array.from(await pending);
@@ -2009,14 +1990,13 @@ async function embedOne(text2, options, outer) {
     }
   }
   const request = requestEmbedding(text2, options, outer);
-  inFlight.set(requestKey, request);
+  inFlight.set(text2, request);
   try {
     const embedding = await request;
-    cachedEmbeddingModels.set(embedding, model);
     cache.set(text2, embedding);
     return Array.from(embedding);
   } finally {
-    if (inFlight.get(requestKey) === request) inFlight.delete(requestKey);
+    if (inFlight.get(text2) === request) inFlight.delete(text2);
   }
 }
 async function requestEmbedding(text2, options, outer) {
@@ -2099,15 +2079,13 @@ async function embedMany(texts, options, outer) {
   const { baseUrl = BASE_URL, model, batchSize, cache } = options;
   const timeoutMs = options.timeoutMs ?? DEFAULT_EMBEDDING_REQUEST_TIMEOUT_MS;
   const chunkSize = batchSize ?? DEFAULT_EMBEDDING_BATCH_SIZE;
-  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   const results = new Array(texts.length).fill(null);
   const uncachedIndices = [];
   const uncachedTexts = [];
   for (let i = 0; i < texts.length; i++) {
     if (cache) {
       const cached = cache.get(texts[i]);
-      const cachedModel = cached && cachedEmbeddingModels.get(cached);
-      if (cached && (cachedModel === void 0 || cachedModel === embeddingModel)) {
+      if (cached) {
         results[i] = Array.from(cached);
         continue;
       }
@@ -2123,6 +2101,7 @@ async function embedMany(texts, options, outer) {
     ...taskTypeHeader(options.taskType)
   };
   if (outer?.aborted) throw outer.reason;
+  const embeddingModel = model ?? DEFAULT_API_EMBEDDING_MODEL;
   let newEmbeddings;
   if (uncachedTexts.length <= chunkSize) {
     newEmbeddings = await generateEmbeddingsBatch(
@@ -2170,7 +2149,6 @@ async function embedMany(texts, options, outer) {
     const f32Embedding = Float32Array.from(newEmbeddings[i]);
     results[uncachedIndices[i]] = Array.from(f32Embedding);
     if (cache) {
-      cachedEmbeddingModels.set(f32Embedding, embeddingModel);
       cache.set(uncachedTexts[i], f32Embedding);
     }
   }