@anuma/sdk 3.8.3 → 3.8.4

package/dist/server/index.mjs

@@ -125,7 +125,7 @@ import {
   updateVaultMemoryOp,
   upsertMessageOp,
   vaultMemoryToStored
-} from "./chunk-GKN5NJKZ.mjs";
+} from "./chunk-FS56IDOW.mjs";
 import {
   createStreamingDeAnonymizer,
   resolvePiiRedactor
@@ -4831,6 +4831,162 @@ function createMemoryVaultTool(vaultCtx, options, embeddingOptions, cache) {
 // src/lib/memory/autoExtract.ts
 var import_watermelondb20 = __toESM(require_watermelondb());
 
+// src/lib/memory/entitySalience.ts
+var GENERIC_ENTITY_NAMES = /* @__PURE__ */ new Set([
+  // Relative and absolute time
+  "today",
+  "tomorrow",
+  "yesterday",
+  "tonight",
+  "morning",
+  "afternoon",
+  "evening",
+  "night",
+  "day",
+  "date",
+  "time",
+  "week",
+  "weekend",
+  "weekday",
+  "month",
+  "year",
+  "monday",
+  "tuesday",
+  "wednesday",
+  "thursday",
+  "friday",
+  "saturday",
+  "sunday",
+  "january",
+  "february",
+  "march",
+  "april",
+  "may",
+  "june",
+  "july",
+  "august",
+  "september",
+  "october",
+  "november",
+  "december",
+  // Generic containers and venues — the named ones ("Blue Bottle on Valencia",
+  // "Stanford") are unaffected because they are not bare tokens.
+  "home",
+  "house",
+  "apartment",
+  "room",
+  "office",
+  "work",
+  "workplace",
+  "school",
+  "gym",
+  "store",
+  "shop",
+  "hospital",
+  "airport",
+  "hotel",
+  "city",
+  "town",
+  "country",
+  "place",
+  // Generic calendar entries — "Chicago Marathon" survives, "meeting" doesn't.
+  "meeting",
+  "call",
+  "appointment",
+  "reminder",
+  "event",
+  "birthday",
+  "breakfast",
+  "lunch",
+  "dinner",
+  "brunch",
+  "vacation",
+  "holiday",
+  "trip",
+  // Plurals are listed explicitly rather than stemmed: naive `s`-stripping
+  // would fold real titles into generics ("The Times" → "time"). Only plurals
+  // that are junk as topics in their own right belong here — note the absence
+  // of "times", "places", and anything else that reads as a name.
+  "days",
+  "weeks",
+  "weekends",
+  "months",
+  "years",
+  "rooms",
+  "offices",
+  "houses",
+  "apartments",
+  "stores",
+  "shops",
+  "hotels",
+  "meetings",
+  "calls",
+  "appointments",
+  "reminders",
+  "events",
+  "birthdays",
+  "trips",
+  "vacations",
+  "holidays",
+  // Meta / filler the models occasionally echo back as an entity
+  "user",
+  "assistant",
+  "memory",
+  "memories",
+  "note",
+  "notes",
+  "thing",
+  "things",
+  "stuff",
+  "something",
+  "someone",
+  "anything",
+  "everything",
+  "none",
+  "other",
+  "misc",
+  "general",
+  "unknown"
+]);
+function isGenericEntityName(name) {
+  const tokens = name.toLowerCase().split(/[^\p{L}\p{N}]+/u).filter((token) => token.length > 0);
+  if (tokens.length === 0) return true;
+  const distinctive = tokens.filter(
+    (token) => !GENERIC_ENTITY_NAMES.has(token) && !ARTICLES.has(token) && !WEAK_MODIFIERS.has(token)
+  );
+  if (distinctive.length > 0) return false;
+  const generic = tokens.filter((token) => GENERIC_ENTITY_NAMES.has(token));
+  if (generic.length === 0) return true;
+  const titleShaped = generic.length === 1 && tokens.some((token) => ARTICLES.has(token)) && !tokens.some((token) => WEAK_MODIFIERS.has(token));
+  return !titleShaped;
+}
+var ARTICLES = /* @__PURE__ */ new Set(["the", "a", "an"]);
+var WEAK_MODIFIERS = /* @__PURE__ */ new Set([
+  "my",
+  "our",
+  "their",
+  "his",
+  "her",
+  "its",
+  "this",
+  "that",
+  "next",
+  "last",
+  "upcoming",
+  "every",
+  "at",
+  "in",
+  "on",
+  "of",
+  "for",
+  "with",
+  "from",
+  "to",
+  "by",
+  "and",
+  "or"
+]);
+
 // src/lib/memory/injectionClassifier.ts
 var DEFAULT_MODEL = "inclusionai/ling-2.6-flash";
 var DEFAULT_ATTEMPTS = 2;
@@ -5759,13 +5915,13 @@ function parseEntities(raw) {
   for (const e of raw) {
     if (typeof e === "string") {
       const name2 = e.trim();
-      if (name2.length > 0) out.push({ name: name2 });
+      if (name2.length > 0 && !isGenericEntityName(name2)) out.push({ name: name2 });
       continue;
     }
     if (typeof e !== "object" || e === null) continue;
     const obj = e;
     const name = typeof obj.name === "string" ? obj.name.trim() : "";
-    if (name.length === 0) continue;
+    if (name.length === 0 || isGenericEntityName(name)) continue;
     const kind = typeof obj.kind === "string" && ENTITY_KINDS.includes(obj.kind) ? obj.kind : void 0;
     out.push(kind !== void 0 ? { name, kind } : { name });
   }
@@ -7068,7 +7224,11 @@ var SYSTEM_PROMPT4 = `You assign topics to saved memories for a personal memory
 
 Each memory is a short statement about the user. For each memory, list the NAMED entities it mentions \u2014 these become the memory's topics, used to connect related memories in a knowledge graph.
 
-Include only NAMED entities, skip generic/common nouns. ${ENTITY_KIND_GUIDELINES}
+Include only NAMED entities, skip generic/common nouns. A topic must identify a specific thing the user could recognize by name; if it would be equally true of thousands of unrelated memories, leave it out:
+- SKIP bare generic nouns even when the memory capitalizes them or a calendar entry is titled that way \u2014 "Home", "Work", "Meeting", "Lunch", "Birthday", "Appointment", "Trip", "Today", "Friday", "March" are NOT topics.
+- KEEP the specific thing instead, when the memory names one: "Blue Bottle on Valencia" over "cafe", "Chicago Marathon" over "event", "Hollowpoint Labs" over "work".
+- Prefer 0 entities over a weak one. A memory whose only candidate is a generic noun has NO entities.
+${ENTITY_KIND_GUIDELINES}
 
 Each memory below is written as "<id>: <text>". Output strict JSON: {"memories": [{"id": string, "entities": [{"name": string, "kind": string}]}]}
 - exactly one element per input memory; copy its id EXACTLY as written (the token before the first ": "), with no brackets, quotes, or extra characters
@@ -7080,7 +7240,7 @@ async function extractEntitiesForMemories(memories, options) {
   const redactor = resolvePiiRedactor(options.piiRedaction);
   const vocabulary = (options.existingEntityNames ?? []).filter((n) => n.trim().length > 0).slice(0, MAX_VOCABULARY_NAMES).map((n) => redactor ? redactor.redactText(n).text : n);
   const vocabularyNote = vocabulary.length > 0 ? `The user's existing topics: ${vocabulary.join(", ")}.
-When a memory refers to one of these, reuse the EXACT existing name (same spelling and casing) instead of a variant.
+When a memory refers to one of these, reuse the EXACT existing name (same spelling, casing, spacing and punctuation) instead of a variant \u2014 if "Foo.ai" is listed, never emit "Foo ai" or "Foo AI", and if "Jane Roe" is listed, emit that rather than a bare "Jane".
 
 ` : "";
   for (let i = 0; i < memories.length; i += TOPIC_EXTRACTION_BATCH_SIZE) {

package/dist/server/{recall-XZ6WLFPJ.mjs → recall-VDMDNQNB.mjs}

@@ -1,6 +1,6 @@
 import {
   recall
-} from "./chunk-GKN5NJKZ.mjs";
+} from "./chunk-FS56IDOW.mjs";
 import "./chunk-MIXEASUP.mjs";
 import "./chunk-3RPGXXPZ.mjs";
 export {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@anuma/sdk",
-  "version": "3.8.3",
+  "version": "3.8.4",
   "description": "",
   "main": "./dist/index.cjs",
   "module": "./dist/index.mjs",