fix(model): pin the CLIP text tower to a loadable ONNX revision

Fixes 'Không tải được bộ tìm kiếm' / [ShapeInferenceError] Inferred=1 Declared=768.
The model repo's first ONNX export declared 'masked' as [...,768]; commit 9e1d535
re-exported it cleanly. transformers.js cached the broken bytes under the
unversioned /resolve/main/ URL, so reload and 'Thử lại' replayed them. Pinning
'revision' changes the cache key and lets those browsers recover.

src/assets.js

@@ -10,6 +10,29 @@ export const DATASET = 'https://huggingface.co/datasets/bep40/interior-image-fie
 // so this fork adds a `text_model.onnx` that includes mean pooling and the 768->512
 // projection, letting Transformers.js run it under the usual CLIP file naming.
 export const TEXT_MODEL = 'bep40/clip-vit-b32-multilingual-vi-text';
+
+/**
+ * Pinned to the commit whose `onnx/text_model.onnx` actually loads, instead of
+ * following the `main` branch.
+ *
+ * The repo uploaded a broken export first (commit 68dc949, 136,951,792 bytes).
+ * Its graph declares `masked` as [batch_size, sequence_length, 768] while an
+ * attention mask only ever has 1 feature, so ONNX shape inference aborts with
+ *
+ *   [ShapeInferenceError] Can't merge shape info ... Inferred=1 Declared=768
+ *
+ * and the app dies at "Không tải được bộ tìm kiếm". A later commit (9e1d535,
+ * 136,915,522 bytes) re-exported the graph with no `value_info` at all, which
+ * loads fine -- that is the revision pinned here.
+ *
+ * Pinning matters beyond reproducibility. transformers.js requests the
+ * *unversioned* `/resolve/main/...` URL unless `revision` is given, and caches
+ * under that same URL. Anyone who loaded the app while `main` still pointed at
+ * the broken graph keeps the bad bytes indefinitely: reload and "Thử lại" both
+ * replay from that cache. Naming an explicit commit changes the cache key, so
+ * those visitors fetch the known-good file instead of staying stuck.
+ */
+export const TEXT_MODEL_REVISION = '9e1d53553228b17abfe6986fbe92e22bd6bd6bad';
 export const TOKENIZER = 'sentence-transformers/clip-ViT-B-32-multilingual-v1';
 export const DIMENSIONS = 512;
 

src/worker.js

@@ -1,5 +1,5 @@
 import { AutoTokenizer, CLIPTextModelWithProjection, env } from '@huggingface/transformers';
-import { cachedFile, DATASET, TOKENIZER, TEXT_MODEL, DIMENSIONS } from './assets.js';
+import { cachedFile, DATASET, TOKENIZER, TEXT_MODEL, TEXT_MODEL_REVISION, DIMENSIONS } from './assets.js';
 import { normalizeDatabase, rankImages } from './search.js';
 import { loadCatalogue, searchCatalogue, tryLoadCatalogue } from './catalogue.js';
 
@@ -35,7 +35,16 @@ function initialize() {
       // onnx/ contains exactly one graph, text_model.onnx (130.6 MB), and no
       // *_quantized variant -- verified on the Hub. Any dtype other than 'fp32'
       // would ask for a file that does not exist, so pin 'fp32' and device 'wasm'.
-      CLIPTextModelWithProjection.from_pretrained(TEXT_MODEL, { dtype: 'fp32', device: 'wasm', progress_callback }),
+      //
+      // `revision` is load-bearing, not a reproducibility nicety: without it the
+      // graph is fetched and cached under the unversioned /resolve/main/ URL, so
+      // browsers that cached the broken 68dc949 export keep failing with
+      // [ShapeInferenceError] Inferred=1 Declared=768 on every reload. See
+      // TEXT_MODEL_REVISION in assets.js.
+      CLIPTextModelWithProjection.from_pretrained(
+        TEXT_MODEL,
+        { dtype: 'fp32', device: 'wasm', revision: TEXT_MODEL_REVISION, progress_callback },
+      ),
       cachedFile(DATASET + 'image-embeddings.json').then(buffer => JSON.parse(new TextDecoder().decode(buffer))),
       cachedFile(DATASET + 'image-embeddings_25k-512-32bit.bin'),
     ]);