@equationalapplications/core-llm-wiki 4.2.0 → 4.4.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.js CHANGED
@@ -209,6 +209,40 @@ function parseEmbedding(blob, text) {
209
209
  return null;
210
210
  }
211
211
 
212
+ // src/readOptions.ts
213
+ function normalizeEntityIds(entityId) {
214
+ const input = Array.isArray(entityId) ? entityId : [entityId];
215
+ const seen = /* @__PURE__ */ new Set();
216
+ const normalized = [];
217
+ for (const id of input) {
218
+ if (seen.has(id)) continue;
219
+ seen.add(id);
220
+ normalized.push(id);
221
+ }
222
+ return normalized;
223
+ }
224
+ function sanitizeTierWeights(entityIds, tierWeights) {
225
+ if (tierWeights === void 0) return void 0;
226
+ const sanitized = /* @__PURE__ */ Object.create(null);
227
+ for (const entityId of entityIds) {
228
+ const raw = tierWeights[entityId];
229
+ if (raw === void 0 || !Number.isFinite(raw)) {
230
+ sanitized[entityId] = 1;
231
+ } else {
232
+ sanitized[entityId] = Math.max(0, raw);
233
+ }
234
+ }
235
+ return sanitized;
236
+ }
237
+ function applyTierWeight(score, entityId, sanitizedTierWeights) {
238
+ const weight = sanitizedTierWeights?.[entityId] ?? 1;
239
+ if (weight === 0) return -Infinity;
240
+ return score * weight;
241
+ }
242
+ function shouldExposeReadMetadata(entityId) {
243
+ return Array.isArray(entityId);
244
+ }
245
+
212
246
  // src/WikiMemory.ts
213
247
  var HOOK_TIMEOUT_MARKER = /* @__PURE__ */ Symbol("WikiMemoryHookTimeout");
214
248
  function parseJsonResponse(text) {
@@ -974,6 +1008,25 @@ After running the migration SQL, restart your application.`
974
1008
  }
975
1009
  async read(entityId, query, options) {
976
1010
  const config = this.options.config;
1011
+ const entityIds = normalizeEntityIds(entityId);
1012
+ const sanitizedTierWeights = sanitizeTierWeights(entityIds, options?.tierWeights);
1013
+ const exposeMetadata = shouldExposeReadMetadata(entityId);
1014
+ if (entityIds.length === 0) {
1015
+ const empty = { facts: [], tasks: [], events: [] };
1016
+ if (exposeMetadata) {
1017
+ empty.metadata = { query, entityIds: [] };
1018
+ if (sanitizedTierWeights && Object.keys(sanitizedTierWeights).length > 0) empty.metadata.tierWeights = sanitizedTierWeights;
1019
+ }
1020
+ return empty;
1021
+ }
1022
+ const MAX_ENTITY_IDS = 100;
1023
+ if (entityIds.length > MAX_ENTITY_IDS) {
1024
+ throw new RangeError(`read() accepts at most ${MAX_ENTITY_IDS} entity IDs; received ${entityIds.length}`);
1025
+ }
1026
+ const nullByteId = entityIds.find((id) => id.includes("\0"));
1027
+ if (nullByteId !== void 0) {
1028
+ throw new TypeError(`entity_id values must not contain the null byte (\\x00); got "${nullByteId}"`);
1029
+ }
977
1030
  const rawMaxResults = options?.maxResults ?? config?.maxResults ?? config?.maxFtsResults ?? 10;
978
1031
  const maxResults = Number.isFinite(rawMaxResults) ? Math.max(0, Math.trunc(rawMaxResults)) : 10;
979
1032
  const rawPreFilterLimit = options?.preFilterLimit === null ? void 0 : options?.preFilterLimit ?? config?.preFilterLimit;
@@ -984,13 +1037,15 @@ After running the migration SQL, restart your application.`
984
1037
  const embedFn = this.options.llmProvider.embed;
985
1038
  const trimmedQuery = query.trim();
986
1039
  let facts = [];
1040
+ let scoreByFactId;
987
1041
  if (maxResults === 0) ; else if (trimmedQuery) {
988
1042
  let usedEmbed = false;
989
- if (!skipEmbed && embedFn) {
1043
+ const scoredEntityIds = this._filterScoredEntities(entityIds, sanitizedTierWeights, options?.includeZeroWeightEntities);
1044
+ if (scoredEntityIds.length === 0) {
1045
+ usedEmbed = true;
1046
+ } else if (!skipEmbed && embedFn) {
990
1047
  let rankerShouldRethrow = false;
991
1048
  let pendingRankerFallbackError;
992
- let usedKeywordFallback = false;
993
- let scoredAlreadySortedAndLimited = false;
994
1049
  try {
995
1050
  const queryVec = await embedFn(trimmedQuery);
996
1051
  if (queryVec.length === 0 || !queryVec.every((v) => typeof v === "number" && isFinite(v))) {
@@ -1009,13 +1064,14 @@ After running the migration SQL, restart your application.`
1009
1064
  );
1010
1065
  }
1011
1066
  }
1067
+ const mismatchScope = this._entityInClause(entityIds);
1012
1068
  const mismatchedCount = await this.db.getFirstAsync(
1013
1069
  `SELECT COUNT(*) AS cnt FROM ${this.prefix}entries
1014
- WHERE entity_id = ? AND deleted_at IS NULL
1070
+ WHERE ${mismatchScope.clause} AND deleted_at IS NULL
1015
1071
  AND embedding_blob IS NOT NULL
1016
1072
  AND (CAST(length(embedding_blob) AS INTEGER) % 4 = 0)
1017
1073
  AND (CAST(length(embedding_blob) AS INTEGER) / 4) != ?`,
1018
- [entityId, queryVec.length]
1074
+ [...mismatchScope.params, queryVec.length]
1019
1075
  );
1020
1076
  if (mismatchedCount && mismatchedCount.cnt > 0) {
1021
1077
  throw new Error(
@@ -1024,12 +1080,13 @@ After running the migration SQL, restart your application.`
1024
1080
  }
1025
1081
  const useRanker = Boolean(this.options.vectorRanker);
1026
1082
  let candidateRows;
1027
- let populateCache = true;
1083
+ let populateCache = entityIds.length === 1;
1028
1084
  let miniSearchScores;
1029
1085
  if (effectivePreFilterLimit !== void 0) {
1030
1086
  populateCache = false;
1087
+ const entityIdSet = new Set(scoredEntityIds);
1031
1088
  const preResults = this.miniSearch.search(trimmedQuery, {
1032
- filter: (r) => r.entity_id === entityId,
1089
+ filter: (r) => entityIdSet.has(r.entity_id),
1033
1090
  combineWith: "OR"
1034
1091
  });
1035
1092
  if (preResults.length === 0) {
@@ -1047,7 +1104,7 @@ After running the migration SQL, restart your application.`
1047
1104
  const idChunk = topKIds.slice(i, i + inClauseChunkSize);
1048
1105
  const placeholders = idChunk.map(() => "?").join(",");
1049
1106
  const chunkRows = await this.db.getAllAsync(
1050
- `SELECT id, updated_at, access_count FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND deleted_at IS NULL`,
1107
+ `SELECT id, entity_id, updated_at, access_count FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND deleted_at IS NULL`,
1051
1108
  idChunk
1052
1109
  );
1053
1110
  rows.push(...chunkRows);
@@ -1059,7 +1116,7 @@ After running the migration SQL, restart your application.`
1059
1116
  const idChunk = topKIds.slice(i, i + inClauseChunkSize);
1060
1117
  const placeholders = idChunk.map(() => "?").join(",");
1061
1118
  const chunkRows = await this.db.getAllAsync(
1062
- `SELECT id, embedding_blob, embedding, updated_at, access_count FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND deleted_at IS NULL`,
1119
+ `SELECT id, entity_id, embedding_blob, embedding, updated_at, access_count FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND deleted_at IS NULL`,
1063
1120
  idChunk
1064
1121
  );
1065
1122
  rows.push(...chunkRows);
@@ -1074,19 +1131,22 @@ After running the migration SQL, restart your application.`
1074
1131
  }
1075
1132
  } else {
1076
1133
  if (useRanker) {
1134
+ const entityScope = this._entityInClause(scoredEntityIds);
1077
1135
  candidateRows = await this.db.getAllAsync(
1078
- `SELECT id, updated_at, access_count FROM ${this.prefix}entries WHERE entity_id = ? AND deleted_at IS NULL`,
1079
- [entityId]
1136
+ `SELECT id, entity_id, updated_at, access_count FROM ${this.prefix}entries WHERE ${entityScope.clause} AND deleted_at IS NULL`,
1137
+ entityScope.params
1080
1138
  );
1081
1139
  } else {
1140
+ const entityScope = this._entityInClause(scoredEntityIds);
1082
1141
  candidateRows = await this.db.getAllAsync(
1083
- `SELECT id, embedding_blob, embedding, updated_at, access_count FROM ${this.prefix}entries WHERE entity_id = ? AND deleted_at IS NULL`,
1084
- [entityId]
1142
+ `SELECT id, entity_id, embedding_blob, embedding, updated_at, access_count FROM ${this.prefix}entries WHERE ${entityScope.clause} AND deleted_at IS NULL`,
1143
+ entityScope.params
1085
1144
  );
1086
1145
  }
1087
1146
  if (weight !== void 0 && weight < 1) {
1147
+ const entityIdSet = new Set(scoredEntityIds);
1088
1148
  const msResults = this.miniSearch.search(trimmedQuery, {
1089
- filter: (r) => r.entity_id === entityId,
1149
+ filter: (r) => entityIdSet.has(r.entity_id),
1090
1150
  combineWith: "OR"
1091
1151
  });
1092
1152
  const maxMsScore = Math.max(1, msResults[0]?.score ?? 1);
@@ -1096,33 +1156,45 @@ After running the migration SQL, restart your application.`
1096
1156
  if (candidateRows === null) {
1097
1157
  usedEmbed = true;
1098
1158
  } else {
1159
+ const entityCacheKey = entityIds.length === 1 ? entityIds[0] : entityIds.join("\0");
1099
1160
  let scored;
1100
1161
  if (useRanker) {
1101
- const candidateIds = effectivePreFilterLimit !== void 0 ? candidateRows.map((r) => r.id) : void 0;
1162
+ const candidateRowsByEntity = /* @__PURE__ */ new Map();
1163
+ for (const row of candidateRows) {
1164
+ const rows = candidateRowsByEntity.get(row.entity_id) ?? [];
1165
+ rows.push(row);
1166
+ candidateRowsByEntity.set(row.entity_id, rows);
1167
+ }
1102
1168
  try {
1103
- const oversampledLimit = Math.max(maxResults * 2, maxResults + 50);
1104
- scored = await this._rankWithVectorRanker({
1105
- entityId,
1106
- queryVec,
1107
- candidateIds,
1108
- weight,
1109
- miniSearchScores,
1110
- limit: oversampledLimit
1111
- });
1112
- if (scored.length > 0) {
1113
- const scoredIds2 = new Set(scored.map((s) => s.id));
1114
- const metaMap = /* @__PURE__ */ new Map();
1115
- for (const r of candidateRows) {
1116
- if (scoredIds2.has(r.id)) {
1117
- metaMap.set(r.id, { updated_at: r.updated_at, access_count: r.access_count });
1118
- }
1119
- }
1120
- scored = scored.map((s) => {
1121
- const meta = metaMap.get(s.id);
1122
- return { ...s, updated_at: meta?.updated_at ?? null, access_count: meta?.access_count ?? null };
1123
- });
1124
- }
1169
+ const rankerResultsByEntity = await Promise.all(
1170
+ scoredEntityIds.filter((id) => (candidateRowsByEntity.get(id)?.length ?? 0) > 0).map(async (scopedEntityId) => {
1171
+ const rowsForEntity = candidateRowsByEntity.get(scopedEntityId) ?? [];
1172
+ const candidateIds = effectivePreFilterLimit !== void 0 ? rowsForEntity.map((row) => row.id) : void 0;
1173
+ const ranked = await this._rankWithVectorRanker({
1174
+ entityId: scopedEntityId,
1175
+ queryVec,
1176
+ candidateIds,
1177
+ candidateRows: rowsForEntity,
1178
+ weight,
1179
+ miniSearchScores,
1180
+ limit: Math.max(maxResults * 2, maxResults + 50)
1181
+ });
1182
+ return ranked.map((row) => ({ ...row, entity_id: scopedEntityId }));
1183
+ })
1184
+ );
1185
+ scored = rankerResultsByEntity.flat();
1125
1186
  const scoredIds = new Set(scored.map((s) => s.id));
1187
+ const metadataById = new Map(
1188
+ candidateRows.filter((row) => scoredIds.has(row.id)).map((row) => [row.id, row])
1189
+ );
1190
+ scored = scored.map((row) => {
1191
+ const metadata = metadataById.get(row.id);
1192
+ return {
1193
+ ...row,
1194
+ updated_at: metadata?.updated_at ?? null,
1195
+ access_count: metadata?.access_count ?? null
1196
+ };
1197
+ });
1126
1198
  const isHybrid = weight !== void 0 && weight < 1;
1127
1199
  const maxBackfill = isHybrid ? maxResults : Math.max(0, maxResults - scored.length);
1128
1200
  if (maxBackfill > 0) {
@@ -1200,6 +1272,7 @@ After running the migration SQL, restart your application.`
1200
1272
  for (const { row, kwScore } of topK) {
1201
1273
  scored.push({
1202
1274
  id: row.id,
1275
+ entity_id: row.entity_id,
1203
1276
  score: (1 - weight) * kwScore,
1204
1277
  updated_at: row.updated_at,
1205
1278
  access_count: row.access_count
@@ -1209,7 +1282,7 @@ After running the migration SQL, restart your application.`
1209
1282
  const omitted = [];
1210
1283
  for (const row of candidateRows) {
1211
1284
  if (scoredIds.has(row.id)) continue;
1212
- omitted.push({ id: row.id, score: -2, updated_at: row.updated_at, access_count: row.access_count });
1285
+ omitted.push({ id: row.id, entity_id: row.entity_id, score: -2, updated_at: row.updated_at, access_count: row.access_count });
1213
1286
  }
1214
1287
  if (omitted.length > 0) {
1215
1288
  this._tieBreakSort(omitted);
@@ -1237,8 +1310,8 @@ After running the migration SQL, restart your application.`
1237
1310
  const idChunk = rowIds.slice(i, i + chunkSize);
1238
1311
  const placeholders = idChunk.map(() => "?").join(",");
1239
1312
  const embeddingRows = await this.db.getAllAsync(
1240
- `SELECT id, embedding_blob, embedding FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND entity_id = ? AND deleted_at IS NULL`,
1241
- [...idChunk, entityId]
1313
+ `SELECT id, embedding_blob, embedding FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND deleted_at IS NULL`,
1314
+ idChunk
1242
1315
  );
1243
1316
  for (const row of embeddingRows) {
1244
1317
  embeddingsMap.set(row.id, { embedding_blob: row.embedding_blob, embedding: row.embedding });
@@ -1251,38 +1324,37 @@ After running the migration SQL, restart your application.`
1251
1324
  }));
1252
1325
  }
1253
1326
  scored = await this._rankWithJsCosine({
1254
- entityId,
1327
+ entityId: entityCacheKey,
1255
1328
  queryVec,
1256
1329
  candidateRows: fallbackRows,
1257
1330
  weight,
1258
1331
  miniSearchScores,
1259
1332
  populateCache,
1260
- limit: maxResults
1333
+ limit: fallbackRows.length,
1334
+ skipSort: true
1335
+ // read() re-sorts after applying tier weights
1261
1336
  });
1262
- scoredAlreadySortedAndLimited = true;
1263
1337
  } else if (policy === "keyword") {
1338
+ const scoredEntityIdSet = new Set(scoredEntityIds);
1264
1339
  const msResults = this.miniSearch.search(trimmedQuery, {
1265
- filter: (r) => r.entity_id === entityId,
1340
+ filter: (r) => scoredEntityIdSet.has(r.entity_id),
1266
1341
  combineWith: "OR"
1267
1342
  });
1268
- const topResults = msResults.slice(0, maxResults);
1269
- const resultIds = new Set(topResults.map((r) => r.id));
1270
- const candidateMap = /* @__PURE__ */ new Map();
1271
- for (const r of candidateRows) {
1272
- if (resultIds.has(r.id)) {
1273
- candidateMap.set(r.id, { updated_at: r.updated_at, access_count: r.access_count });
1274
- }
1275
- }
1276
- scored = topResults.map((r) => {
1277
- const meta = candidateMap.get(r.id);
1343
+ const keywordOversampledLimit = Math.max(maxResults * 2, maxResults + 50);
1344
+ const topResults = msResults.slice(0, keywordOversampledLimit);
1345
+ const topResultIds = new Set(topResults.map((r) => r.id));
1346
+ const candidateMap = new Map(candidateRows.filter((r) => topResultIds.has(r.id)).map((row) => [row.id, row]));
1347
+ scored = topResults.map((result) => {
1348
+ const metadata = candidateMap.get(result.id);
1349
+ const entityForScore = metadata?.entity_id ?? result.entity_id ?? "";
1278
1350
  return {
1279
- id: r.id,
1280
- score: r.score ?? 0,
1281
- access_count: meta?.access_count ?? null,
1282
- updated_at: meta?.updated_at ?? null
1351
+ id: result.id,
1352
+ entity_id: entityForScore,
1353
+ score: result.score ?? 0,
1354
+ access_count: metadata?.access_count ?? null,
1355
+ updated_at: metadata?.updated_at ?? null
1283
1356
  };
1284
1357
  });
1285
- usedKeywordFallback = true;
1286
1358
  } else {
1287
1359
  scored = [];
1288
1360
  }
@@ -1294,46 +1366,44 @@ After running the migration SQL, restart your application.`
1294
1366
  }
1295
1367
  }
1296
1368
  } else {
1369
+ const jsCosineNeedsTierSort = sanitizedTierWeights !== void 0 && Object.values(sanitizedTierWeights).some((w) => w !== 1);
1297
1370
  scored = await this._rankWithJsCosine({
1298
- entityId,
1371
+ entityId: entityCacheKey,
1299
1372
  queryVec,
1300
1373
  candidateRows,
1301
1374
  weight,
1302
1375
  miniSearchScores,
1303
1376
  populateCache,
1304
- limit: maxResults
1377
+ limit: jsCosineNeedsTierSort ? candidateRows.length : maxResults,
1378
+ skipSort: jsCosineNeedsTierSort
1379
+ // read() re-sorts after applying tier weights
1305
1380
  });
1306
- scoredAlreadySortedAndLimited = true;
1307
1381
  }
1308
1382
  if (scored.length > 0) {
1309
- if (!usedKeywordFallback && !scoredAlreadySortedAndLimited) {
1310
- this._tieBreakSort(scored);
1383
+ scored = scored.map((row) => ({
1384
+ ...row,
1385
+ score: applyTierWeight(row.score, row.entity_id, sanitizedTierWeights)
1386
+ }));
1387
+ this._tieBreakSort(scored);
1388
+ const selectedScored = scored.slice(0, maxResults);
1389
+ const topIds = selectedScored.map((s) => s.id);
1390
+ if (exposeMetadata && trimmedQuery) {
1391
+ scoreByFactId = new Map(selectedScored.map((s) => [s.id, Number.isFinite(s.score) ? s.score : 0]));
1311
1392
  }
1312
- const topIds = (scoredAlreadySortedAndLimited ? scored : scored.slice(0, maxResults)).map((s) => s.id);
1313
1393
  if (topIds.length > 0) {
1314
- const fullRows = [];
1315
- const phase2ChunkSize = 500;
1316
- for (let i = 0; i < topIds.length; i += phase2ChunkSize) {
1317
- const idChunk = topIds.slice(i, i + phase2ChunkSize);
1318
- const placeholders = idChunk.map(() => "?").join(",");
1319
- const chunkRows = await this.db.getAllAsync(
1320
- `SELECT * FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND entity_id = ? AND deleted_at IS NULL`,
1321
- [...idChunk, entityId]
1322
- );
1323
- fullRows.push(...chunkRows);
1324
- }
1325
- const byId = new Map(fullRows.map((r) => [r.id, r]));
1326
- facts = topIds.map((id) => byId.get(id)).filter((f) => f !== void 0);
1327
- if (facts.length < topIds.length) {
1328
- const missingIds = topIds.filter((id) => !byId.has(id));
1394
+ const facts2 = await this._hydrateFactsByIds(topIds, entityIds);
1395
+ if (facts2.length < topIds.length) {
1396
+ const hydrationById = new Set(facts2.map((f) => f.id));
1397
+ const missingIds = topIds.filter((id) => !hydrationById.has(id));
1329
1398
  const missingCount = missingIds.length;
1330
1399
  const sample = missingIds.slice(0, 5);
1331
1400
  const sampleSuffix = sample.length > 0 ? ` Missing ID sample: ${sample.join(", ")}${missingIds.length > sample.length ? ", ..." : ""}.` : "";
1332
1401
  const error = new Error(
1333
- `Phase 2 fact hydration returned ${missingCount} fewer row(s) than ranked IDs for entity ${entityId}. Rows may have been concurrently soft-deleted or filtered by deleted_at during hydration, or vector ranker output may include IDs that do not exist for this entity.` + sampleSuffix
1402
+ `Phase 2 fact hydration returned ${missingCount} fewer row(s) than ranked IDs. Rows may have been concurrently soft-deleted or filtered by deleted_at during hydration, or vector ranker output may include IDs that do not exist in requested entities.` + sampleSuffix
1334
1403
  );
1335
1404
  this.options.onRetrievalFallback?.(error);
1336
1405
  }
1406
+ facts = facts2;
1337
1407
  }
1338
1408
  if (pendingRankerFallbackError) {
1339
1409
  this.options.onRetrievalFallback?.(pendingRankerFallbackError);
@@ -1360,26 +1430,28 @@ After running the migration SQL, restart your application.`
1360
1430
  this.options.onRetrievalFallback?.(error);
1361
1431
  }
1362
1432
  }
1363
- if (!usedEmbed) {
1433
+ if (!usedEmbed && scoredEntityIds.length > 0) {
1434
+ const fallbackEntityIdSet = new Set(scoredEntityIds);
1435
+ const fallbackOversampledLimit = Math.max(maxResults * 2, maxResults + 50);
1364
1436
  const results = this.miniSearch.search(trimmedQuery, {
1365
- filter: (r) => r.entity_id === entityId,
1437
+ filter: (r) => fallbackEntityIdSet.has(r.entity_id),
1366
1438
  combineWith: "OR"
1367
1439
  });
1368
- const topIds = results.slice(0, maxResults).map((r) => r.id);
1440
+ const candidates = results.slice(0, fallbackOversampledLimit).map((r) => ({
1441
+ id: r.id,
1442
+ entity_id: r.entity_id,
1443
+ score: applyTierWeight(r.score ?? 0, r.entity_id, sanitizedTierWeights),
1444
+ updated_at: null,
1445
+ access_count: null
1446
+ }));
1447
+ this._tieBreakSort(candidates);
1448
+ const topCandidates = candidates.slice(0, maxResults);
1449
+ const topIds = topCandidates.map((c) => c.id);
1369
1450
  if (topIds.length > 0) {
1370
- const kwRows = [];
1371
- const kwChunkSize = 500;
1372
- for (let i = 0; i < topIds.length; i += kwChunkSize) {
1373
- const idChunk = topIds.slice(i, i + kwChunkSize);
1374
- const placeholders = idChunk.map(() => "?").join(",");
1375
- const chunkRows = await this.db.getAllAsync(
1376
- `SELECT * FROM ${this.prefix}entries WHERE id IN (${placeholders}) AND entity_id = ? AND deleted_at IS NULL`,
1377
- [...idChunk, entityId]
1378
- );
1379
- kwRows.push(...chunkRows);
1451
+ facts = await this._hydrateFactsByIds(topIds, entityIds);
1452
+ if (exposeMetadata) {
1453
+ scoreByFactId = new Map(topCandidates.map((c) => [c.id, Number.isFinite(c.score) ? c.score : 0]));
1380
1454
  }
1381
- const byId = new Map(kwRows.map((r) => [r.id, r]));
1382
- facts = topIds.map((id) => byId.get(id)).filter((f) => f !== void 0);
1383
1455
  }
1384
1456
  }
1385
1457
  if (facts.length > 0) {
@@ -1398,37 +1470,66 @@ After running the migration SQL, restart your application.`
1398
1470
  }
1399
1471
  }
1400
1472
  } else {
1401
- facts = await this.db.getAllAsync(
1473
+ const entityScope = this._entityInClause(entityIds);
1474
+ const rawFacts = await this.db.getAllAsync(
1402
1475
  `SELECT * FROM ${this.prefix}entries
1403
- WHERE entity_id = ? AND deleted_at IS NULL
1476
+ WHERE ${entityScope.clause} AND deleted_at IS NULL
1404
1477
  ORDER BY updated_at DESC
1405
1478
  LIMIT ?`,
1406
- [entityId, maxResults]
1479
+ [...entityScope.params, maxResults]
1407
1480
  );
1481
+ facts = rawFacts.map((f) => {
1482
+ const { embedding: _embedding, embedding_blob: _blob, ...rest } = f;
1483
+ return {
1484
+ ...rest,
1485
+ tags: typeof rest.tags === "string" ? JSON.parse(rest.tags) : rest.tags
1486
+ };
1487
+ });
1408
1488
  }
1409
1489
  const [tasks, events] = await Promise.all([
1410
- this.db.getAllAsync(
1411
- `SELECT * FROM ${this.prefix}tasks
1412
- WHERE entity_id = ? AND status IN ('pending', 'in_progress') AND deleted_at IS NULL
1413
- ORDER BY priority DESC, created_at ASC`,
1414
- [entityId]
1415
- ),
1416
- this.db.getAllAsync(
1417
- `SELECT * FROM ${this.prefix}events
1418
- WHERE entity_id = ?
1419
- ORDER BY created_at DESC
1420
- LIMIT 10`,
1421
- [entityId]
1422
- )
1490
+ (async () => {
1491
+ const entityScope = this._entityInClause(entityIds);
1492
+ const tasksLimit = entityIds.length === 1 ? void 0 : Math.min(20 * entityIds.length, 200);
1493
+ return this.db.getAllAsync(
1494
+ `SELECT * FROM ${this.prefix}tasks
1495
+ WHERE ${entityScope.clause} AND status IN ('pending', 'in_progress') AND deleted_at IS NULL
1496
+ ORDER BY priority DESC, created_at ASC${tasksLimit !== void 0 ? "\n LIMIT ?" : ""}`,
1497
+ tasksLimit !== void 0 ? [...entityScope.params, tasksLimit] : entityScope.params
1498
+ );
1499
+ })(),
1500
+ (async () => {
1501
+ const entityScope = this._entityInClause(entityIds);
1502
+ const eventsLimit = Math.min(10 * entityIds.length, 100);
1503
+ return this.db.getAllAsync(
1504
+ `SELECT * FROM ${this.prefix}events
1505
+ WHERE ${entityScope.clause}
1506
+ ORDER BY created_at DESC
1507
+ LIMIT ?`,
1508
+ [...entityScope.params, eventsLimit]
1509
+ );
1510
+ })()
1423
1511
  ]);
1424
- const parsedFacts = facts.map((f) => {
1425
- const { embedding: _embedding, embedding_blob: _blob, ...rest } = f;
1426
- return {
1427
- ...rest,
1428
- tags: typeof rest.tags === "string" ? JSON.parse(rest.tags) : rest.tags
1429
- };
1512
+ let factScores;
1513
+ if (exposeMetadata && trimmedQuery && scoreByFactId) {
1514
+ factScores = Object.fromEntries(facts.map((fact) => [fact.id, scoreByFactId.get(fact.id) ?? 0]));
1515
+ }
1516
+ const bundle = { facts, tasks, events: events.reverse() };
1517
+ if (exposeMetadata) {
1518
+ bundle.metadata = { query, entityIds };
1519
+ if (sanitizedTierWeights && Object.keys(sanitizedTierWeights).length > 0) bundle.metadata.tierWeights = sanitizedTierWeights;
1520
+ if (factScores && Object.keys(factScores).length > 0) bundle.factScores = factScores;
1521
+ }
1522
+ return bundle;
1523
+ }
1524
+ /**
1525
+ * Returns entity IDs that will participate in scored retrieval.
1526
+ * Excludes zero-weight entities unless includeZeroWeightEntities is true.
1527
+ */
1528
+ _filterScoredEntities(entityIds, sanitizedTierWeights, includeZeroWeightEntities) {
1529
+ return entityIds.filter((id) => {
1530
+ const w = sanitizedTierWeights?.[id] ?? 1;
1531
+ return includeZeroWeightEntities === true || w !== 0;
1430
1532
  });
1431
- return { facts: parsedFacts, tasks, events: events.reverse() };
1432
1533
  }
1433
1534
  /**
1434
1535
  * Stable tie-break sort: score desc → access_count desc → updated_at desc → id asc.
@@ -1442,13 +1543,48 @@ After running the migration SQL, restart your application.`
1442
1543
  */
1443
1544
  _compareScoredRows(a, b) {
1444
1545
  const scoreDiff = b.score - a.score;
1445
- if (scoreDiff !== 0) return scoreDiff;
1546
+ if (!Number.isNaN(scoreDiff) && scoreDiff !== 0) return scoreDiff;
1446
1547
  const accessCountDiff = (b.access_count ?? 0) - (a.access_count ?? 0);
1447
1548
  if (accessCountDiff !== 0) return accessCountDiff;
1448
1549
  const updatedAtDiff = (b.updated_at ?? 0) - (a.updated_at ?? 0);
1449
1550
  if (updatedAtDiff !== 0) return updatedAtDiff;
1450
1551
  return a.id.localeCompare(b.id);
1451
1552
  }
1553
+ /**
1554
+ * Build SQL IN clause with placeholders for multiple entity IDs.
1555
+ */
1556
+ _entityInClause(entityIds) {
1557
+ if (entityIds.length === 0) return { clause: "1=0", params: [] };
1558
+ const placeholders = entityIds.map(() => "?").join(",");
1559
+ return { clause: `entity_id IN (${placeholders})`, params: [...entityIds] };
1560
+ }
1561
+ /**
1562
+ * Hydrate full facts by ID. Pass scopedEntityIds to restrict to requested namespaces in SQL
1563
+ * (defense-in-depth against a rogue VectorRanker returning cross-entity IDs).
1564
+ */
1565
+ async _hydrateFactsByIds(ids, scopedEntityIds) {
1566
+ const fullRows = [];
1567
+ const chunkSize = 500;
1568
+ const entityClause = scopedEntityIds && scopedEntityIds.length > 0 ? ` AND entity_id IN (${scopedEntityIds.map(() => "?").join(",")})` : "";
1569
+ const entityParams = scopedEntityIds && scopedEntityIds.length > 0 ? [...scopedEntityIds] : [];
1570
+ for (let i = 0; i < ids.length; i += chunkSize) {
1571
+ const idChunk = ids.slice(i, i + chunkSize);
1572
+ const placeholders = idChunk.map(() => "?").join(",");
1573
+ const chunkRows = await this.db.getAllAsync(
1574
+ `SELECT * FROM ${this.prefix}entries WHERE id IN (${placeholders})${entityClause} AND deleted_at IS NULL`,
1575
+ [...idChunk, ...entityParams]
1576
+ );
1577
+ fullRows.push(...chunkRows);
1578
+ }
1579
+ const byId = new Map(fullRows.map((row) => [row.id, row]));
1580
+ return ids.map((id) => byId.get(id)).filter((fact) => fact !== void 0).map((fact) => {
1581
+ const { embedding: _embedding, embedding_blob: _blob, ...rest } = fact;
1582
+ return {
1583
+ ...rest,
1584
+ tags: typeof rest.tags === "string" ? JSON.parse(rest.tags) : rest.tags
1585
+ };
1586
+ });
1587
+ }
1452
1588
  /**
1453
1589
  * Strip potentially sensitive data from ranker errors before exposing to host callbacks.
1454
1590
  * Preserves error type for debugging but removes message/stack that may contain credentials.
@@ -1473,7 +1609,7 @@ After running the migration SQL, restart your application.`
1473
1609
  */
1474
1610
  async _rankWithJsCosine(args) {
1475
1611
  const queryVec = args.queryVec instanceof Float32Array ? args.queryVec.slice() : Array.from(args.queryVec);
1476
- const { entityId, candidateRows, weight, miniSearchScores, populateCache, limit } = args;
1612
+ const { entityId, candidateRows, weight, miniSearchScores, populateCache, limit, skipSort } = args;
1477
1613
  let entityCache = this.vectorCache.get(entityId);
1478
1614
  const tooLarge = populateCache && candidateRows.length > _WikiMemory.MAX_VECTOR_CACHE_FACTS_PER_ENTITY;
1479
1615
  if (tooLarge && entityCache) {
@@ -1504,7 +1640,13 @@ After running the migration SQL, restart your application.`
1504
1640
  } else {
1505
1641
  score = -2;
1506
1642
  }
1507
- return { id: row.id, score, updated_at: row.updated_at, access_count: row.access_count };
1643
+ return {
1644
+ id: row.id,
1645
+ entity_id: row.entity_id,
1646
+ score,
1647
+ updated_at: row.updated_at,
1648
+ access_count: row.access_count
1649
+ };
1508
1650
  });
1509
1651
  if (canCache && entityCache && entityCache.size > 0) {
1510
1652
  if (!this.vectorCache.has(entityId)) {
@@ -1515,7 +1657,7 @@ After running the migration SQL, restart your application.`
1515
1657
  this.vectorCache.set(entityId, entityCache);
1516
1658
  }
1517
1659
  }
1518
- this._tieBreakSort(scored);
1660
+ if (!skipSort) this._tieBreakSort(scored);
1519
1661
  return scored.slice(0, limit);
1520
1662
  }
1521
1663
  /**
@@ -1524,7 +1666,7 @@ After running the migration SQL, restart your application.`
1524
1666
  * Returns scored results ready for hybrid blending and tie-break sorting.
1525
1667
  */
1526
1668
  async _rankWithVectorRanker(args) {
1527
- const { entityId, candidateIds, weight, miniSearchScores, limit } = args;
1669
+ const { entityId, candidateIds, candidateRows, weight, miniSearchScores, limit } = args;
1528
1670
  const ranker = this.options.vectorRanker;
1529
1671
  if (!ranker) {
1530
1672
  throw new Error("vectorRanker not configured");
@@ -1536,7 +1678,7 @@ After running the migration SQL, restart your application.`
1536
1678
  candidateIds,
1537
1679
  limit
1538
1680
  });
1539
- const allowedIds = candidateIds ? new Set(candidateIds) : void 0;
1681
+ const allowedIds = new Set(candidateRows.map((row) => row.id));
1540
1682
  const seen = /* @__PURE__ */ new Set();
1541
1683
  const normalized = [];
1542
1684
  for (const r of rankerResults) {
@@ -1547,13 +1689,19 @@ After running the migration SQL, restart your application.`
1547
1689
  seen.add(r.id);
1548
1690
  normalized.push(r);
1549
1691
  }
1692
+ const entityIdByCandidateId = new Map(candidateRows.map((row) => [row.id, row.entity_id]));
1550
1693
  const scored = normalized.map((r) => {
1551
1694
  let score = r.semanticScore;
1552
1695
  if (weight !== void 0) {
1553
1696
  const kwScore = miniSearchScores?.get(r.id) ?? 0;
1554
1697
  score = weight * Math.max(0, r.semanticScore) + (1 - weight) * kwScore;
1555
1698
  }
1556
- return { id: r.id, score };
1699
+ return {
1700
+ id: r.id,
1701
+ entity_id: entityIdByCandidateId.get(r.id),
1702
+ // allowedIds filter above guarantees membership
1703
+ score
1704
+ };
1557
1705
  });
1558
1706
  return scored;
1559
1707
  }