@bendyline/gezk 1.0.3 → 1.0.5

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/node.js CHANGED
@@ -28,6 +28,43 @@ var RepoRelativePathSchema = z2.string().min(1).refine(
28
28
  );
29
29
  var DEFAULT_EMBEDDING_ONNX_FILE = "onnx/model.onnx";
30
30
  var DEFAULT_EMBEDDING_TOKENIZER_FILE = "tokenizer.json";
31
+ var EmbeddingModelFileSchema = z2.object({
32
+ path: RepoRelativePathSchema,
33
+ digest: ArtifactDigestSchema.optional()
34
+ });
35
+ var EmbeddingMediaEncoderSchema = z2.object({
36
+ onnxFile: RepoRelativePathSchema,
37
+ onnxDigest: ArtifactDigestSchema.optional(),
38
+ files: z2.array(EmbeddingModelFileSchema).optional()
39
+ });
40
+ var EmbeddingMediaSchema = z2.object({
41
+ image: z2.object({
42
+ encoder: EmbeddingMediaEncoderSchema,
43
+ /** Vision tokens per image; changing it moves every image vector. */
44
+ tokenBudget: z2.number().int().positive(),
45
+ resample: z2.enum(["bilinear", "bicubic"]),
46
+ /** Transparent pixels are composited onto white before encoding. */
47
+ alpha: z2.literal("composite-white")
48
+ }).optional(),
49
+ /** Video frames go through the image encoder, so `video` requires `image`. */
50
+ video: z2.object({
51
+ framesPerSecond: z2.number().positive(),
52
+ maxFrames: z2.number().int().positive(),
53
+ tokenBudgetPerFrame: z2.number().int().positive()
54
+ }).optional(),
55
+ audio: z2.object({
56
+ encoder: EmbeddingMediaEncoderSchema,
57
+ sampleRate: z2.number().int().positive(),
58
+ channels: z2.literal(1),
59
+ /** Longest window one audio row may cover. */
60
+ maxWindowMs: z2.number().int().positive()
61
+ }).optional()
62
+ }).refine((media) => Boolean(media.image || media.video || media.audio), {
63
+ message: "media must describe at least one modality"
64
+ }).refine((media) => !media.video || Boolean(media.image), {
65
+ message: "media.video requires media.image (frames go through the image encoder)",
66
+ path: ["video"]
67
+ });
31
68
  var KnowledgeEmbeddingProfileObject = z2.object({
32
69
  /** e.g. `multilingual-e5-small@1`. A new model, prefix or encoding is a new id. */
33
70
  id: z2.string().min(1),
@@ -44,7 +81,9 @@ var KnowledgeEmbeddingProfileObject = z2.object({
44
81
  */
45
82
  onnxFile: RepoRelativePathSchema.optional(),
46
83
  /** sha256 of that file's bytes at `revision`. */
47
- onnxDigest: ArtifactDigestSchema.optional()
84
+ onnxDigest: ArtifactDigestSchema.optional(),
85
+ /** Every further file the runtime loads: weight sidecars, configs (0.8). */
86
+ files: z2.array(EmbeddingModelFileSchema).optional()
48
87
  }),
49
88
  tokenizer: z2.object({
50
89
  kind: z2.string().min(1),
@@ -55,7 +94,17 @@ var KnowledgeEmbeddingProfileObject = z2.object({
55
94
  }),
56
95
  pooling: z2.enum(["mean", "cls", "last"]),
57
96
  normalized: z2.boolean(),
97
+ /** The stored vector width — after truncation, when the profile truncates. */
58
98
  dimensions: z2.number().int().positive(),
99
+ /**
100
+ * Matryoshka truncation (0.8): the model emits `sourceDimensions` values,
101
+ * the first `dimensions` are kept, and the result is L2-normalized again.
102
+ * One rule for passages, queries and media; `profileUnitVector` owns it.
103
+ */
104
+ truncation: z2.object({
105
+ method: z2.literal("prefix"),
106
+ sourceDimensions: z2.number().int().positive()
107
+ }).optional(),
59
108
  maxTokens: z2.number().int().positive(),
60
109
  queryInstruction: z2.string(),
61
110
  passageInstruction: z2.string(),
@@ -91,7 +140,8 @@ var KnowledgeEmbeddingProfileObject = z2.object({
91
140
  /** Required by `centered-sign`, refused by `sign`; one entry per dimension. */
92
141
  center: z2.array(z2.number()).optional()
93
142
  })
94
- })
143
+ }),
144
+ media: EmbeddingMediaSchema.optional()
95
145
  });
96
146
  function embeddingProfileCenterProblem(profile) {
97
147
  const { method, center } = profile.quantization.binary;
@@ -110,6 +160,19 @@ function embeddingProfileCenter(profile) {
110
160
  const { center } = profile.quantization.binary;
111
161
  return center ? Float32Array.from(center) : null;
112
162
  }
163
+ function embeddingProfileTruncationProblem(profile) {
164
+ if (!profile.truncation) return null;
165
+ if (profile.truncation.sourceDimensions <= profile.dimensions) {
166
+ return `truncation.sourceDimensions (${profile.truncation.sourceDimensions}) must exceed dimensions (${profile.dimensions})`;
167
+ }
168
+ return profile.normalized ? null : "truncation requires a normalized profile";
169
+ }
170
+ function embeddingProfileSourceDimensions(profile) {
171
+ return profile.truncation?.sourceDimensions ?? profile.dimensions;
172
+ }
173
+ function embeddingProfileMinimumFormat(profile) {
174
+ return profile.truncation || profile.model.files || profile.media ? "0.8" : "0.5";
175
+ }
113
176
  var KnowledgeEmbeddingProfileSchema = KnowledgeEmbeddingProfileObject.superRefine(
114
177
  (profile, ctx) => {
115
178
  const problem = embeddingProfileCenterProblem(profile);
@@ -120,6 +183,10 @@ var KnowledgeEmbeddingProfileSchema = KnowledgeEmbeddingProfileObject.superRefin
120
183
  path: ["quantization", "binary", "center"]
121
184
  });
122
185
  }
186
+ const truncation = embeddingProfileTruncationProblem(profile);
187
+ if (truncation) {
188
+ ctx.addIssue({ code: "custom", message: truncation, path: ["truncation"] });
189
+ }
123
190
  }
124
191
  );
125
192
  function embeddingProfileArtifacts(profile) {
@@ -127,14 +194,19 @@ function embeddingProfileArtifacts(profile) {
127
194
  onnxFile: profile.model.onnxFile ?? DEFAULT_EMBEDDING_ONNX_FILE,
128
195
  onnxDigest: profile.model.onnxDigest ?? null,
129
196
  tokenizerFile: profile.tokenizer.file ?? DEFAULT_EMBEDDING_TOKENIZER_FILE,
130
- tokenizerDigest: profile.tokenizer.digest ?? null
197
+ tokenizerDigest: profile.tokenizer.digest ?? null,
198
+ files: (profile.model.files ?? []).map((f) => ({ path: f.path, digest: f.digest ?? null }))
131
199
  };
132
200
  }
133
201
  function sameVectorSpace(a, b) {
134
202
  const aa = embeddingProfileArtifacts(a);
135
203
  const ab = embeddingProfileArtifacts(b);
136
204
  const digestsAgree = (x, y) => !x || !y || x === y;
137
- return a.model.repo === b.model.repo && a.model.revision === b.model.revision && aa.onnxFile === ab.onnxFile && digestsAgree(aa.onnxDigest, ab.onnxDigest) && a.tokenizer.kind === b.tokenizer.kind && aa.tokenizerFile === ab.tokenizerFile && digestsAgree(aa.tokenizerDigest, ab.tokenizerDigest) && a.pooling === b.pooling && a.normalized === b.normalized && a.dimensions === b.dimensions && a.queryInstruction === b.queryInstruction && a.passageInstruction === b.passageInstruction && a.vectorEncoding === b.vectorEncoding && a.distance.stage1 === b.distance.stage1 && a.distance.stage2 === b.distance.stage2 && a.quantization.int8.method === b.quantization.int8.method && a.quantization.int8.scale === b.quantization.int8.scale;
205
+ const bFiles = new Map(ab.files.map((f) => [f.path, f.digest]));
206
+ const filesAgree = aa.files.every(
207
+ (f) => !bFiles.has(f.path) || digestsAgree(f.digest, bFiles.get(f.path) ?? null)
208
+ );
209
+ return filesAgree && a.truncation?.method === b.truncation?.method && embeddingProfileSourceDimensions(a) === embeddingProfileSourceDimensions(b) && a.model.repo === b.model.repo && a.model.revision === b.model.revision && aa.onnxFile === ab.onnxFile && digestsAgree(aa.onnxDigest, ab.onnxDigest) && a.tokenizer.kind === b.tokenizer.kind && aa.tokenizerFile === ab.tokenizerFile && digestsAgree(aa.tokenizerDigest, ab.tokenizerDigest) && a.pooling === b.pooling && a.normalized === b.normalized && a.dimensions === b.dimensions && a.queryInstruction === b.queryInstruction && a.passageInstruction === b.passageInstruction && a.vectorEncoding === b.vectorEncoding && a.distance.stage1 === b.distance.stage1 && a.distance.stage2 === b.distance.stage2 && a.quantization.int8.method === b.quantization.int8.method && a.quantization.int8.scale === b.quantization.int8.scale;
138
210
  }
139
211
  var KnowledgeChunkingProfileSchema = z2.object({
140
212
  /** e.g. `markdown-chunks@2`. */
@@ -150,14 +222,17 @@ var KnowledgeChunkingProfileSchema = z2.object({
150
222
  import { z as z4 } from "zod";
151
223
 
152
224
  // src/format/constants.ts
153
- var GEZK_FORMAT_VERSION = "0.7";
154
- var GEZK_INDEX_SCHEMA_VERSION = 4;
155
- var GEZK_FORMAT_GENERATIONS = { "0.5": 2, "0.6": 3, "0.7": 4 };
225
+ var GEZK_FORMAT_VERSION = "0.8";
226
+ var GEZK_INDEX_SCHEMA_VERSION = 5;
227
+ var GEZK_FORMAT_GENERATIONS = { "0.5": 2, "0.6": 3, "0.7": 4, "0.8": 5 };
156
228
  var GEZK_SUPPORTED_FORMAT_VERSIONS = Object.keys(GEZK_FORMAT_GENERATIONS);
157
229
  var GEZK_SUPPORTED_INDEX_SCHEMA_VERSIONS = Object.values(GEZK_FORMAT_GENERATIONS);
158
230
  function isSupportedFormatVersion(value) {
159
231
  return typeof value === "string" && Object.hasOwn(GEZK_FORMAT_GENERATIONS, value);
160
232
  }
233
+ function formatAtLeast(version, minimum) {
234
+ return GEZK_FORMAT_GENERATIONS[version] >= GEZK_FORMAT_GENERATIONS[minimum];
235
+ }
161
236
  function isSupportedIndexSchemaVersion(value) {
162
237
  return typeof value === "number" && GEZK_SUPPORTED_INDEX_SCHEMA_VERSIONS.includes(value);
163
238
  }
@@ -321,7 +396,7 @@ var KnowledgeSignatureSchema = z4.object({
321
396
  var KnowledgeCatalogManifestSchema = z4.object({
322
397
  kind: z4.literal(GEZK_MANIFEST_KIND),
323
398
  formatVersion: z4.enum(GEZK_SUPPORTED_FORMAT_VERSIONS),
324
- indexSchemaVersion: z4.union([z4.literal(2), z4.literal(3), z4.literal(4)]),
399
+ indexSchemaVersion: z4.literal(GEZK_SUPPORTED_INDEX_SCHEMA_VERSIONS),
325
400
  id: KnowledgeIdSchema,
326
401
  version: KnowledgeVersionSchema,
327
402
  name: z4.string().min(1),
@@ -378,7 +453,16 @@ var KnowledgeCatalogManifestSchema = z4.object({
378
453
  chunks: z4.number().int().nonnegative(),
379
454
  shards: z4.number().int().positive(),
380
455
  /** Files under `assets/` (0.6+). */
381
- assets: z4.number().int().nonnegative().optional()
456
+ assets: z4.number().int().nonnegative().optional(),
457
+ /**
458
+ * Media rows by modality (0.8, required there). They are counted in
459
+ * `chunks` too: every row of the shard `chunks` table is a chunk.
460
+ */
461
+ media: z4.object({
462
+ image: z4.number().int().nonnegative(),
463
+ video: z4.number().int().nonnegative(),
464
+ audio: z4.number().int().nonnegative()
465
+ }).optional()
382
466
  }),
383
467
  files: z4.array(KnowledgeManifestFileSchema).min(1),
384
468
  spatial: KnowledgeSpatialManifestSchema.optional(),
@@ -405,18 +489,61 @@ var KnowledgeCatalogManifestSchema = z4.object({
405
489
  }).optional(),
406
490
  signature: KnowledgeSignatureSchema.optional()
407
491
  }).superRefine((manifest, ctx) => {
408
- if (manifest.formatVersion === "0.7" && !manifest.spatial) {
492
+ const spatialEra = formatAtLeast(manifest.formatVersion, "0.7");
493
+ if (spatialEra && !manifest.spatial) {
409
494
  ctx.addIssue({
410
495
  code: "custom",
411
496
  path: ["spatial"],
412
- message: "0.7 requires a spatial summary, including catalogs with zero locations"
497
+ message: "0.7+ requires a spatial summary, including catalogs with zero locations"
413
498
  });
414
499
  }
415
- if (manifest.spatial && (manifest.formatVersion !== "0.7" || manifest.spatial.locatedDocuments > manifest.counts.documents)) {
500
+ if (manifest.spatial && (!spatialEra || manifest.spatial.locatedDocuments > manifest.counts.documents)) {
416
501
  ctx.addIssue({
417
502
  code: "custom",
418
503
  path: ["spatial"],
419
- message: "spatial requires 0.7 and valid document counts"
504
+ message: "spatial requires 0.7+ and valid document counts"
505
+ });
506
+ }
507
+ const mediaEra = formatAtLeast(manifest.formatVersion, "0.8");
508
+ const media = manifest.counts.media;
509
+ if (mediaEra && !media) {
510
+ ctx.addIssue({
511
+ code: "custom",
512
+ path: ["counts", "media"],
513
+ message: "0.8 requires counts.media, including catalogs with no media rows"
514
+ });
515
+ }
516
+ if (media && !mediaEra) {
517
+ ctx.addIssue({
518
+ code: "custom",
519
+ path: ["counts", "media"],
520
+ message: "counts.media is a 0.8 field"
521
+ });
522
+ }
523
+ if (media && media.image + media.video + media.audio > manifest.counts.chunks) {
524
+ ctx.addIssue({
525
+ code: "custom",
526
+ path: ["counts", "media"],
527
+ message: "media rows are chunks, so counts.media cannot exceed counts.chunks"
528
+ });
529
+ }
530
+ if (media) {
531
+ for (const modality of ["image", "video", "audio"]) {
532
+ if (media[modality] > 0 && !manifest.embedding.media?.[modality]) {
533
+ ctx.addIssue({
534
+ code: "custom",
535
+ path: ["embedding", "media", modality],
536
+ message: `${media[modality]} ${modality} rows need embedding.media.${modality}`
537
+ });
538
+ }
539
+ }
540
+ }
541
+ const profileFormat = embeddingProfileMinimumFormat(manifest.embedding);
542
+ if (!formatAtLeast(manifest.formatVersion, profileFormat)) {
543
+ ctx.addIssue({
544
+ code: "custom",
545
+ path: ["embedding"],
546
+ message: `embedding profile ${manifest.embedding.id} uses ${profileFormat} vocabulary (truncation, model.files or media), which ${manifest.formatVersion} cannot carry`
420
547
  });
421
548
  }
422
549
  if (GEZK_FORMAT_GENERATIONS[manifest.formatVersion] !== manifest.indexSchemaVersion) {
@@ -669,7 +796,8 @@ CREATE VIRTUAL TABLE fts_documents USING fts5(
669
796
  tokenize = 'unicode61 remove_diacritics 2', prefix = '2 3'
670
797
  );
671
798
  `;
672
- var SHARD_DDL = `
799
+ function shardDdl(mediaColumns, mediaIndex) {
800
+ return `
673
801
  CREATE TABLE meta(key TEXT PRIMARY KEY, value TEXT NOT NULL) WITHOUT ROWID;
674
802
 
675
803
  CREATE TABLE chunks(
@@ -683,10 +811,10 @@ CREATE TABLE chunks(
683
811
  line_start INTEGER NOT NULL, line_end INTEGER NOT NULL,
684
812
  token_count INTEGER NOT NULL,
685
813
  content_hash TEXT NOT NULL,
686
- text TEXT NOT NULL
814
+ text TEXT NOT NULL${mediaColumns}
687
815
  );
688
816
  CREATE UNIQUE INDEX chunks_uid ON chunks(chunk_uid);
689
- CREATE INDEX chunks_document ON chunks(document_id, ordinal);
817
+ CREATE INDEX chunks_document ON chunks(document_id, ordinal);${mediaIndex}
690
818
 
691
819
  CREATE VIRTUAL TABLE fts_chunks USING fts5(
692
820
  title, heading_text, text,
@@ -704,6 +832,30 @@ CREATE TABLE chunk_vectors_int8(
704
832
  v BLOB NOT NULL
705
833
  );
706
834
  `;
835
+ }
836
+ var MEDIA_COLUMNS = `,
837
+ modality TEXT NOT NULL DEFAULT 'text'
838
+ CHECK (modality IN ('text','image','video','audio')),
839
+ asset_path TEXT,
840
+ mime_type TEXT,
841
+ width INTEGER, height INTEGER,
842
+ start_ms INTEGER, end_ms INTEGER,
843
+ thumbnail_path TEXT,
844
+ attribution_json TEXT,
845
+ CHECK (modality <> 'text' OR (asset_path IS NULL AND mime_type IS NULL
846
+ AND width IS NULL AND height IS NULL AND start_ms IS NULL AND end_ms IS NULL
847
+ AND thumbnail_path IS NULL AND attribution_json IS NULL)),
848
+ CHECK (modality = 'text' OR (asset_path IS NOT NULL AND mime_type IS NOT NULL)),
849
+ CHECK (modality <> 'image' OR (start_ms IS NULL AND end_ms IS NULL)),
850
+ CHECK (modality NOT IN ('video','audio') OR (start_ms >= 0 AND end_ms > start_ms))`;
851
+ var MEDIA_INDEX = `
852
+ CREATE INDEX chunks_media ON chunks(modality, id) WHERE modality <> 'text';`;
853
+ var SHARD_DDL_V4 = shardDdl("", "");
854
+ var SHARD_DDL_V5 = shardDdl(MEDIA_COLUMNS, MEDIA_INDEX);
855
+ var SHARD_DDL = SHARD_DDL_V5;
856
+ function shardDdlFor(indexSchemaVersion) {
857
+ return indexSchemaVersion === 5 ? SHARD_DDL_V5 : SHARD_DDL_V4;
858
+ }
707
859
 
708
860
  // src/format/assets.ts
709
861
  import { z as z7 } from "zod";
@@ -716,11 +868,23 @@ var KNOWLEDGE_ASSET_TYPES = {
716
868
  webp: "image/webp",
717
869
  svg: "image/svg+xml"
718
870
  };
871
+ var KNOWLEDGE_MEDIA_ASSET_TYPES = {
872
+ mp4: "video/mp4",
873
+ webm: "video/webm",
874
+ mp3: "audio/mpeg",
875
+ m4a: "audio/mp4",
876
+ ogg: "audio/ogg",
877
+ opus: "audio/ogg",
878
+ wav: "audio/wav",
879
+ flac: "audio/flac"
880
+ };
719
881
  var MAX_KNOWLEDGE_ASSET_BYTES = 8 * 1024 * 1024;
720
882
  var MAX_KNOWLEDGE_ASSETS_TOTAL_BYTES = 256 * 1024 * 1024;
883
+ var MAX_KNOWLEDGE_MEDIA_ASSET_BYTES = 512 * 1024 * 1024;
884
+ var MAX_KNOWLEDGE_MEDIA_ASSETS_TOTAL_BYTES = 8 * 1024 * 1024 * 1024;
721
885
  var MAX_KNOWLEDGE_ASSET_COUNT = 8192;
722
886
  var MAX_KNOWLEDGE_ASSET_PATH_LENGTH = 512;
723
- var KNOWLEDGE_ASSET_PATH_PATTERN = /^assets\/(?:[A-Za-z0-9_-][A-Za-z0-9._-]{0,127}\/){0,15}[A-Za-z0-9_-][A-Za-z0-9._-]{0,127}\.(?:png|jpe?g|gif|webp|svg)$/i;
887
+ var KNOWLEDGE_ASSET_PATH_PATTERN = /^assets\/(?:[A-Za-z0-9_-][A-Za-z0-9._-]{0,127}\/){0,15}[A-Za-z0-9_-][A-Za-z0-9._-]{0,127}\.(?:png|jpe?g|gif|webp|svg|mp4|webm|mp3|m4a|ogg|opus|wav|flac)$/i;
724
888
  var KnowledgeAssetPathSchema = z7.string().max(MAX_KNOWLEDGE_ASSET_PATH_LENGTH).regex(KNOWLEDGE_ASSET_PATH_PATTERN, "not a valid asset path");
725
889
  function isKnowledgeAssetPath(path) {
726
890
  return path.length <= MAX_KNOWLEDGE_ASSET_PATH_LENGTH && KNOWLEDGE_ASSET_PATH_PATTERN.test(path);
@@ -729,14 +893,61 @@ function assetExtension(path) {
729
893
  const dot = path.lastIndexOf(".");
730
894
  if (dot < 0) return null;
731
895
  const ext = path.slice(dot + 1).toLowerCase();
732
- return Object.hasOwn(KNOWLEDGE_ASSET_TYPES, ext) ? ext : null;
896
+ if (Object.hasOwn(KNOWLEDGE_ASSET_TYPES, ext)) return ext;
897
+ if (Object.hasOwn(KNOWLEDGE_MEDIA_ASSET_TYPES, ext)) return ext;
898
+ return null;
733
899
  }
734
900
  function assetContentType(path) {
735
901
  const ext = assetExtension(path);
736
- return ext ? KNOWLEDGE_ASSET_TYPES[ext] : null;
902
+ if (!ext) return null;
903
+ return ext in KNOWLEDGE_ASSET_TYPES ? KNOWLEDGE_ASSET_TYPES[ext] : KNOWLEDGE_MEDIA_ASSET_TYPES[ext];
904
+ }
905
+ function isKnowledgeMediaAssetPath(path) {
906
+ const ext = assetExtension(path);
907
+ return ext !== null && Object.hasOwn(KNOWLEDGE_MEDIA_ASSET_TYPES, ext);
908
+ }
909
+ function assetModality(path) {
910
+ const ext = assetExtension(path);
911
+ if (!ext) return null;
912
+ if (ext === "mp4" || ext === "webm") return "video";
913
+ return Object.hasOwn(KNOWLEDGE_MEDIA_ASSET_TYPES, ext) ? "audio" : "image";
914
+ }
915
+ function maxKnowledgeAssetBytes(path) {
916
+ return isKnowledgeMediaAssetPath(path) ? MAX_KNOWLEDGE_MEDIA_ASSET_BYTES : MAX_KNOWLEDGE_ASSET_BYTES;
917
+ }
918
+ function knowledgeAssetLimitsProblem(files) {
919
+ if (files.length > MAX_KNOWLEDGE_ASSET_COUNT) {
920
+ return `${files.length} assets exceed the ${MAX_KNOWLEDGE_ASSET_COUNT}-file limit`;
921
+ }
922
+ let imageBytes = 0;
923
+ let mediaBytes = 0;
924
+ for (const file of files) {
925
+ if (file.sizeBytes > maxKnowledgeAssetBytes(file.path)) {
926
+ return `${file.path} is ${file.sizeBytes} bytes, over its ${maxKnowledgeAssetBytes(file.path)}-byte limit`;
927
+ }
928
+ if (isKnowledgeMediaAssetPath(file.path)) mediaBytes += file.sizeBytes;
929
+ else imageBytes += file.sizeBytes;
930
+ }
931
+ if (imageBytes > MAX_KNOWLEDGE_ASSETS_TOTAL_BYTES) {
932
+ return `image assets total ${imageBytes} bytes, over the ${MAX_KNOWLEDGE_ASSETS_TOTAL_BYTES}-byte limit`;
933
+ }
934
+ if (mediaBytes > MAX_KNOWLEDGE_MEDIA_ASSETS_TOTAL_BYTES) {
935
+ return `audio and video assets total ${mediaBytes} bytes, over the ${MAX_KNOWLEDGE_MEDIA_ASSETS_TOTAL_BYTES}-byte limit`;
936
+ }
937
+ return null;
737
938
  }
738
939
  function assetKindForExtension(ext) {
739
- return ext === "jpg" ? "jpeg" : ext;
940
+ switch (ext) {
941
+ case "jpg":
942
+ return "jpeg";
943
+ case "mp4":
944
+ case "m4a":
945
+ return "isobmff";
946
+ case "opus":
947
+ return "ogg";
948
+ default:
949
+ return ext;
950
+ }
740
951
  }
741
952
  var PNG_MAGIC = [137, 80, 78, 71, 13, 10, 26, 10];
742
953
  var JPEG_MAGIC = [255, 216, 255];
@@ -744,6 +955,12 @@ var GIF87 = [71, 73, 70, 56, 55, 97];
744
955
  var GIF89 = [71, 73, 70, 56, 57, 97];
745
956
  var RIFF = [82, 73, 70, 70];
746
957
  var WEBP = [87, 69, 66, 80];
958
+ var WAVE = [87, 65, 86, 69];
959
+ var FTYP = [102, 116, 121, 112];
960
+ var EBML = [26, 69, 223, 163];
961
+ var OGGS = [79, 103, 103, 83];
962
+ var FLAC = [102, 76, 97, 67];
963
+ var ID3 = [73, 68, 51];
747
964
  function startsWith(bytes, magic, offset = 0) {
748
965
  if (bytes.length < offset + magic.length) return false;
749
966
  for (let i = 0; i < magic.length; i++) {
@@ -756,8 +973,20 @@ function sniffAssetType(bytes) {
756
973
  if (startsWith(bytes, JPEG_MAGIC)) return "jpeg";
757
974
  if (startsWith(bytes, GIF87) || startsWith(bytes, GIF89)) return "gif";
758
975
  if (startsWith(bytes, RIFF) && startsWith(bytes, WEBP, 8)) return "webp";
976
+ if (startsWith(bytes, RIFF) && startsWith(bytes, WAVE, 8)) return "wav";
977
+ if (startsWith(bytes, FTYP, 4)) return "isobmff";
978
+ if (startsWith(bytes, EBML)) return "webm";
979
+ if (startsWith(bytes, OGGS)) return "ogg";
980
+ if (startsWith(bytes, FLAC)) return "flac";
981
+ if (startsWith(bytes, ID3) || isMpegAudioFrame(bytes)) return "mp3";
759
982
  return looksLikeSvg(decodeText(bytes)) ? "svg" : null;
760
983
  }
984
+ function isMpegAudioFrame(bytes) {
985
+ if (bytes.length < 2) return false;
986
+ const b0 = bytes[0];
987
+ const b1 = bytes[1];
988
+ return b0 === 255 && (b1 & 224) === 224 && (b1 & 6) !== 0;
989
+ }
761
990
  function decodeText(bytes) {
762
991
  try {
763
992
  return new TextDecoder("utf-8", { fatal: true, ignoreBOM: false }).decode(bytes);
@@ -852,6 +1081,16 @@ function topicSortKeyForOrder(order) {
852
1081
  }
853
1082
 
854
1083
  // src/format/quantize.ts
1084
+ function profileUnitVector(profile, raw) {
1085
+ const source = embeddingProfileSourceDimensions(profile);
1086
+ if (raw.length !== source) {
1087
+ throw new Error(
1088
+ `embedding has ${raw.length} dimensions, profile ${profile.id} expects ${source}`
1089
+ );
1090
+ }
1091
+ if (source === profile.dimensions) return l2Normalize(raw);
1092
+ return l2Normalize(Array.prototype.slice.call(raw, 0, profile.dimensions));
1093
+ }
855
1094
  function quantizeInt8(unitVector) {
856
1095
  const out = new Int8Array(unitVector.length);
857
1096
  for (let i = 0; i < unitVector.length; i++) {
@@ -883,9 +1122,13 @@ function quantizeBinary(unitVector) {
883
1122
  return out;
884
1123
  }
885
1124
  function rerankScore(query, passageInt8) {
1125
+ if (query.length !== passageInt8.length) {
1126
+ throw new Error(`query has ${query.length} dimensions, passage has ${passageInt8.length}`);
1127
+ }
886
1128
  let dot = 0;
887
- const n = Math.min(query.length, passageInt8.length);
888
- for (let i = 0; i < n; i++) dot += query[i] * (passageInt8[i] / 127);
1129
+ for (let i = 0; i < query.length; i++) {
1130
+ dot += query[i] * (passageInt8[i] / 127);
1131
+ }
889
1132
  return dot;
890
1133
  }
891
1134
  function l2Normalize(vector) {
@@ -952,6 +1195,178 @@ function documentSlug(title) {
952
1195
  return slug || "untitled";
953
1196
  }
954
1197
 
1198
+ // src/query-stopwords.ts
1199
+ var QUERY_STOP_WORDS = /* @__PURE__ */ new Set([
1200
+ // articles + determiners
1201
+ "a",
1202
+ "an",
1203
+ "the",
1204
+ "this",
1205
+ "that",
1206
+ "these",
1207
+ "those",
1208
+ "some",
1209
+ "any",
1210
+ "all",
1211
+ "each",
1212
+ "both",
1213
+ "every",
1214
+ "other",
1215
+ "another",
1216
+ "same",
1217
+ "such",
1218
+ // pronouns
1219
+ "i",
1220
+ "me",
1221
+ "my",
1222
+ "we",
1223
+ "us",
1224
+ "our",
1225
+ "you",
1226
+ "your",
1227
+ "it",
1228
+ "its",
1229
+ "he",
1230
+ "him",
1231
+ "his",
1232
+ "she",
1233
+ "her",
1234
+ "they",
1235
+ "them",
1236
+ "their",
1237
+ // be / have / do
1238
+ "am",
1239
+ "is",
1240
+ "are",
1241
+ "was",
1242
+ "were",
1243
+ "be",
1244
+ "been",
1245
+ "being",
1246
+ "have",
1247
+ "has",
1248
+ "had",
1249
+ "do",
1250
+ "does",
1251
+ "did",
1252
+ "done",
1253
+ // modals
1254
+ "can",
1255
+ "could",
1256
+ "will",
1257
+ "would",
1258
+ "shall",
1259
+ "should",
1260
+ "may",
1261
+ "might",
1262
+ "must",
1263
+ // question words
1264
+ "how",
1265
+ "what",
1266
+ "when",
1267
+ "where",
1268
+ "which",
1269
+ "who",
1270
+ "whom",
1271
+ "why",
1272
+ // prepositions + conjunctions
1273
+ "about",
1274
+ "and",
1275
+ "as",
1276
+ "at",
1277
+ "but",
1278
+ "by",
1279
+ "for",
1280
+ "from",
1281
+ "if",
1282
+ "in",
1283
+ "into",
1284
+ "of",
1285
+ "on",
1286
+ "or",
1287
+ "so",
1288
+ "than",
1289
+ "then",
1290
+ "there",
1291
+ "to",
1292
+ "with",
1293
+ // request scaffolding + degree words
1294
+ "please",
1295
+ "help",
1296
+ "let",
1297
+ "tell",
1298
+ "want",
1299
+ "need",
1300
+ "create",
1301
+ "make",
1302
+ "give",
1303
+ "get",
1304
+ "got",
1305
+ "use",
1306
+ "using",
1307
+ "just",
1308
+ "like",
1309
+ "also",
1310
+ "very",
1311
+ "only",
1312
+ "new",
1313
+ "more",
1314
+ "most",
1315
+ "much",
1316
+ "many",
1317
+ "few",
1318
+ "own",
1319
+ // conversational scaffolding
1320
+ "hey",
1321
+ "hi",
1322
+ "hiya",
1323
+ "hello",
1324
+ "thanks",
1325
+ "thank",
1326
+ "okay",
1327
+ "ok",
1328
+ "yep",
1329
+ "yeah",
1330
+ "yes",
1331
+ "nope",
1332
+ "good",
1333
+ "great",
1334
+ "fine",
1335
+ "nice",
1336
+ "cool",
1337
+ "ready",
1338
+ "going",
1339
+ "doing",
1340
+ "sounds",
1341
+ "morning",
1342
+ "afternoon",
1343
+ "evening",
1344
+ "night",
1345
+ // Apostrophes are token boundaries, so contractions otherwise leave these
1346
+ // fragments looking like rare, highly distinctive query terms.
1347
+ "s",
1348
+ "t",
1349
+ "re",
1350
+ "ve",
1351
+ "ll",
1352
+ "d",
1353
+ "m",
1354
+ "isn",
1355
+ "aren",
1356
+ "wasn",
1357
+ "weren",
1358
+ "don",
1359
+ "doesn",
1360
+ "didn",
1361
+ "hasn",
1362
+ "haven",
1363
+ "hadn",
1364
+ "won",
1365
+ "wouldn",
1366
+ "shouldn",
1367
+ "couldn"
1368
+ ]);
1369
+
955
1370
  // src/node/ids.ts
956
1371
  import { createHash } from "crypto";
957
1372
  function chunkUid(documentId, ordinal, chunkText) {
@@ -1078,6 +1493,8 @@ export {
1078
1493
  CatalogDocumentSchema,
1079
1494
  DEFAULT_EMBEDDING_ONNX_FILE,
1080
1495
  DEFAULT_EMBEDDING_TOKENIZER_FILE,
1496
+ EmbeddingMediaSchema,
1497
+ EmbeddingModelFileSchema,
1081
1498
  GEZK_APPLICATION_ID,
1082
1499
  GEZK_EARTH_RADIUS_METERS,
1083
1500
  GEZK_FORMAT_GENERATIONS,
@@ -1092,6 +1509,7 @@ export {
1092
1509
  KNOWLEDGE_ASSET_TYPES,
1093
1510
  KNOWLEDGE_ID_PATTERN,
1094
1511
  KNOWLEDGE_MANIFEST_INDEX_SCHEMA_VERSIONS,
1512
+ KNOWLEDGE_MEDIA_ASSET_TYPES,
1095
1513
  KNOWLEDGE_VERSION_PATTERN,
1096
1514
  KnowledgeAssetPathSchema,
1097
1515
  KnowledgeBoundsSchema,
@@ -1123,13 +1541,18 @@ export {
1123
1541
  MAX_KNOWLEDGE_ASSET_PATH_LENGTH,
1124
1542
  MAX_KNOWLEDGE_DOCUMENT_BYTES,
1125
1543
  MAX_KNOWLEDGE_DOCUMENT_META_BYTES,
1544
+ MAX_KNOWLEDGE_MEDIA_ASSETS_TOTAL_BYTES,
1545
+ MAX_KNOWLEDGE_MEDIA_ASSET_BYTES,
1126
1546
  MAX_KNOWLEDGE_TOPIC_DEPTH,
1127
1547
  MIMETYPE_PATH,
1548
+ QUERY_STOP_WORDS,
1128
1549
  README_PATH,
1129
1550
  ROUTER_DB_PATH,
1130
1551
  ROUTER_DDL,
1131
1552
  RepoRelativePathSchema,
1132
1553
  SHARD_DDL,
1554
+ SHARD_DDL_V4,
1555
+ SHARD_DDL_V5,
1133
1556
  SOURCE_NOTICES_PATH,
1134
1557
  Sha256HexSchema,
1135
1558
  SourceNoticesSchema,
@@ -1137,6 +1560,7 @@ export {
1137
1560
  assetContentType,
1138
1561
  assetExtension,
1139
1562
  assetKindForExtension,
1563
+ assetModality,
1140
1564
  canonicalizeJson,
1141
1565
  centerVector,
1142
1566
  chunkContentHash,
@@ -1145,25 +1569,34 @@ export {
1145
1569
  embeddingProfileArtifacts,
1146
1570
  embeddingProfileCenter,
1147
1571
  embeddingProfileCenterProblem,
1572
+ embeddingProfileMinimumFormat,
1573
+ embeddingProfileSourceDimensions,
1574
+ embeddingProfileTruncationProblem,
1575
+ formatAtLeast,
1148
1576
  formatKnowledgeUri,
1149
1577
  generateKnowledgeSigningKeyPair,
1150
1578
  hashFileStreaming,
1151
1579
  isKnowledgeAssetPath,
1580
+ isKnowledgeMediaAssetPath,
1152
1581
  isSupportedFormatVersion,
1153
1582
  isSupportedIndexSchemaVersion,
1583
+ knowledgeAssetLimitsProblem,
1154
1584
  knowledgeKeyId,
1155
1585
  l2Normalize,
1156
1586
  locationDistanceMeters,
1157
1587
  locationInBounds,
1158
1588
  manifestSigningPayload,
1589
+ maxKnowledgeAssetBytes,
1159
1590
  normalizeLongitude,
1160
1591
  parseKnowledgeUri,
1592
+ profileUnitVector,
1161
1593
  quantizeBinary,
1162
1594
  quantizeBinaryForProfile,
1163
1595
  quantizeInt8,
1164
1596
  radiusBounds,
1165
1597
  rerankScore,
1166
1598
  sameVectorSpace,
1599
+ shardDdlFor,
1167
1600
  signManifest,
1168
1601
  signRegistryIndex,
1169
1602
  sniffAssetType,