Spaces:

bep40
/

VSEARCH

Running

VSEARCH / src /assets.js

fix(build): restore assets.js exports so the Space can rebuild + ship rebuilt dist

ef50ce7 15 days ago

9.81 kB

/**
 * Image database: 1,222 Vietnamese interior-design photographs, encoded offline with
 * CLIP-ViT-B-32 and stored as raw little-endian float32 in 512-d rows.
 */
export const DATASET = 'https://huggingface.co/datasets/bep40/interior-image-field-assets/resolve/main/';

/**
 * Text tower. `sentence-transformers/clip-ViT-B-32-multilingual-v1` maps 50+
 * languages (including Vietnamese) into the same 512-d space as CLIP-ViT-B-32,
 * so a Vietnamese query retrieves the matching interior photograph. The stock
 * repository ships a SentenceTransformer wrapper whose ONNX export stops at
 * 768-d `last_hidden_state`, 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. 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. Naming an explicit commit changes the cache key, so visitors
 * who cached the broken graph fetch the known-good file instead.
 */
export const TEXT_MODEL_REVISION = '9e1d53553228b17abfe6986fbe92e22bd6bd6bad';
export const TOKENIZER = 'sentence-transformers/clip-ViT-B-32-multilingual-v1';
export const DIMENSIONS = 512;

/** The URL of an interior photograph in the dataset repo, at a CDN width. */
export function imageURL(image, width = 360) {
  return `${DATASET}images/${image.url}?w=${width}&q=80`;
}

/**
 * The thumbnail URL for a tile, whichever kind of record it holds.
 *
 * Interior photographs are files in the dataset repo, so they go through
 * imageURL and its CDN resize. A Malloca product points at a third-party shop
 * CDN that this Space knows nothing about -- there is no filename to build a
 * path from and no query string it would honour -- so its own URL is used
 * verbatim.
 *
 * Dispatch is on `src`, which merge.js sets only for products, rather than on
 * the presence of `url`: an interior record always has `url`, and keying on the
 * absence of a field would mis-classify a record that is merely malformed.
 */
export function tileURL(image, width) {
  return image?.src ? image.src : imageURL(image, width);
}

/** Cache is an optimization: unavailable storage never prevents a search. */
export async function cachedFile(url, onProgress) {
  let cache;
  try {
    cache = await caches.open('interior-image-field-v1');
    const cached = await cache.match(url);
    if (cached) return cached.arrayBuffer();
  } catch { /* Private browsing / storage disabled. */ }

  const response = await fetch(url);
  if (!response.ok) throw new Error(`Could not download image database (${response.status}).`);
  const total = Number(response.headers.get('content-length'));
  let buffer;
  if (response.body && onProgress) {
    const reader = response.body.getReader();
    const chunks = [];
    let loaded = 0;
    while (true) {
      const { done, value } = await reader.read();
      if (done) break;
      chunks.push(value);
      loaded += value.length;
      onProgress(loaded, total);
    }
    const joined = new Uint8Array(loaded);
    let offset = 0;
    for (const chunk of chunks) { joined.set(chunk, offset); offset += chunk.length; }
    buffer = joined.buffer;
  } else buffer = await response.arrayBuffer();
  try { await cache?.put(url, new Response(buffer)); } catch { /* Cache quota reached. */ }
  return buffer;
}

/**
 * The V.AI STUDIO share page for whatever is currently open in the viewer.
 *
 * Both routes live on the vai-avatar Space, which serves them server-side with
 * OpenGraph tags so a pasted link becomes a card in Facebook or Zalo rather
 * than a bare page:
 *
 *   /share/i/<image_id>  interior photos  (the gallery that image came from)
 *   /share/p/<sku>       products          (name, price and gallery)
 *
 * The product route did not exist until it was added there; before that the
 * only product-ish URL was `?product=<sku>`, which is an OG-tag handler and
 * renders nothing but the chat page. `/share/p/` resolves a sku, a model code or
 * a slug against the same catalogue the Space already loads, so all 3,966
 * identity forms in it resolve.
 *
 * Returns null when there is no id to share, so the caller can omit the button
 * rather than render a link that would 404.
 */
const VAI_STUDIO = 'https://bep40-vai-avatar.hf.space';

/**
 * The prefix avatar-interiors.js puts on a vai-avatar photograph's `sku`.
 *
 * Namespacing it there is deliberate -- it is what stops a 16-hex image_id
 * colliding with a baked interior's id in merge.js's de-duplication set, and it
 * is what lets a photograph travel the same code path as a catalogue product
 * (same viewer, same panel, same neon, same share anchor).
 *
 * It is NOT a catalogue identity, and /share/p/ resolves against the product
 * catalogue -- so a photograph sent through that route answers "Not Found".
 * Nothing about that is fixable on vai-avatar's side: no SKU of that shape
 * exists there to find. It has to be routed as the interior it is.
 */
const AVATAR_SKU_PREFIX = 'avatarimg:';

/**
 * Whether this record is an interior photograph published on vai-avatar.
 *
 * Either field is conclusive on its own. `avatarImageId` is set on every such
 * photograph; the namespaced `sku` covers a record that has lost it.
 */
function isAvatarInterior(photo) {
  return Boolean(photo.avatarImageId) || String(photo.sku ?? '').startsWith(AVATAR_SKU_PREFIX);
}

/**
 * The interior id behind a record, or '' when it has none.
 *
 * Two sources, in order, and the second is a fallback rather than an
 * alternative: `avatarImageId` is the field avatar-interiors.js sets, present on
 * every vai-avatar photograph. But a record can reach shareURL() having lost it
 * -- merge.js rewrites `id` into a namespaced tile key, and any path that
 * rebuilds a record from its `sku` alone would carry the prefix but not the
 * original field. That is exactly the shape that produced
 * /share/p/avatarimg%3Ab68dc779370dc70e, which is a "Not Found" page for every
 * single photograph in the vai-avatar interior set.
 *
 * So the namespaced sku is unwrapped as well, and a prefix that is present but
 * empty (`avatarimg:` with nothing after it) yields ''.
 */
function interiorIdOf(photo) {
  const direct = String(photo.avatarImageId ?? '').trim();
  if (direct) return direct;
  const sku = String(photo.sku ?? '').trim();
  return sku.startsWith(AVATAR_SKU_PREFIX)
    ? sku.slice(AVATAR_SKU_PREFIX.length).trim()
    : '';
}

export function shareURL(photo) {
  if (!photo || typeof photo !== 'object') return null;
  /**
   * A photograph published on vai-avatar carries BOTH `avatarImageId` and a
   * namespaced `sku` of `avatarimg:<image_id>`.
   *
   * `sku` exists so the record travels the same path as a catalogue product --
   * same viewer, same detail panel, same neon, same share anchor -- and it has
   * to stay for that. But it is not an identity the product catalogue knows,
   * and it must not decide the URL: `/share/p/` resolves against the product
   * catalogue, so every such photograph sent through it answered "Not Found".
   * The whole vai-avatar interior set was affected, on the tile's share button
   * and on the viewer's "Xem trang" alike.
   *
   * The interior identity is therefore resolved first and unconditionally, and
   * /share/i/ gets the raw id -- no prefix to strip, and nothing to keep in
   * step with whatever namespace merge.js uses for its tile ids.
   */
  const interiorId = interiorIdOf(photo);
  if (interiorId) return `${VAI_STUDIO}/share/i/${encodeURIComponent(interiorId)}`;
  /**
   * A record that still says it is an interior but carries no usable id gets NO
   * share URL, rather than falling through to the product branch below.
   *
   * That fall-through is a trap worth naming, because it is exactly how the bug
   * comes back. `avatarimg:` with nothing after it has no id to share, and every
   * remaining branch here would still produce a URL for it: the product branch
   * would encode the prefix itself into `/share/p/avatarimg%3A`, which is the
   * dead route all over again, and the id branch would send it to `/share/i/`
   * with an empty segment, which 404s in its own right. Returning null lets the
   * caller omit the button instead, which is the only outcome that is true.
   */
  if (isAvatarInterior(photo)) return null;
  if (photo.sku || photo.slug) {
    // slug first, then sku, then mod. A shared URL is a permanent, public,
    // indexable address and the one thing about a product a reader can see:
    // `MH800TC` says nothing about what is linked, while
    // `mh800tc-malloca-may-hut-mui-gan-tu-am-tu-80cm` reads as the product's name
    // to a person and to a crawler. The slug is already the canonical key this
    // Space's catalogue uses, and vai-avatar's /share/p/ matches on it first, so
    // this changes how the link reads, not whether it resolves.
    const id = String(photo.slug || photo.sku || photo.mod || '').trim();
    return id ? `${VAI_STUDIO}/share/p/${encodeURIComponent(id)}` : null;
  }
  const id = String(photo.id ?? '').trim();
  return id ? `${VAI_STUDIO}/share/i/${encodeURIComponent(id)}` : null;
}

	/**
	* Image database: 1,222 Vietnamese interior-design photographs, encoded offline with
	* CLIP-ViT-B-32 and stored as raw little-endian float32 in 512-d rows.
	*/
	export const DATASET = 'https://huggingface.co/datasets/bep40/interior-image-field-assets/resolve/main/';

	/**
	* Text tower. `sentence-transformers/clip-ViT-B-32-multilingual-v1` maps 50+
	* languages (including Vietnamese) into the same 512-d space as CLIP-ViT-B-32,
	* so a Vietnamese query retrieves the matching interior photograph. The stock
	* repository ships a SentenceTransformer wrapper whose ONNX export stops at
	* 768-d `last_hidden_state`, 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. 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. Naming an explicit commit changes the cache key, so visitors
	* who cached the broken graph fetch the known-good file instead.
	*/
	export const TEXT_MODEL_REVISION = '9e1d53553228b17abfe6986fbe92e22bd6bd6bad';
	export const TOKENIZER = 'sentence-transformers/clip-ViT-B-32-multilingual-v1';
	export const DIMENSIONS = 512;

	/** The URL of an interior photograph in the dataset repo, at a CDN width. */
	export function imageURL(image, width = 360) {
	return `${DATASET}images/${image.url}?w=${width}&q=80`;
	}

	/**
	* The thumbnail URL for a tile, whichever kind of record it holds.
	*
	* Interior photographs are files in the dataset repo, so they go through
	* imageURL and its CDN resize. A Malloca product points at a third-party shop
	* CDN that this Space knows nothing about -- there is no filename to build a
	* path from and no query string it would honour -- so its own URL is used
	* verbatim.
	*
	* Dispatch is on `src`, which merge.js sets only for products, rather than on
	* the presence of `url`: an interior record always has `url`, and keying on the
	* absence of a field would mis-classify a record that is merely malformed.
	*/
	export function tileURL(image, width) {
	return image?.src ? image.src : imageURL(image, width);
	}

	/** Cache is an optimization: unavailable storage never prevents a search. */
	export async function cachedFile(url, onProgress) {
	let cache;
	try {
	cache = await caches.open('interior-image-field-v1');
	const cached = await cache.match(url);
	if (cached) return cached.arrayBuffer();
	} catch { /* Private browsing / storage disabled. */ }

	const response = await fetch(url);
	if (!response.ok) throw new Error(`Could not download image database (${response.status}).`);
	const total = Number(response.headers.get('content-length'));
	let buffer;
	if (response.body && onProgress) {
	const reader = response.body.getReader();
	const chunks = [];
	let loaded = 0;
	while (true) {
	const { done, value } = await reader.read();
	if (done) break;
	chunks.push(value);
	loaded += value.length;
	onProgress(loaded, total);
	}
	const joined = new Uint8Array(loaded);
	let offset = 0;
	for (const chunk of chunks) { joined.set(chunk, offset); offset += chunk.length; }
	buffer = joined.buffer;
	} else buffer = await response.arrayBuffer();
	try { await cache?.put(url, new Response(buffer)); } catch { /* Cache quota reached. */ }
	return buffer;
	}

	/**
	* The V.AI STUDIO share page for whatever is currently open in the viewer.
	*
	* Both routes live on the vai-avatar Space, which serves them server-side with
	* OpenGraph tags so a pasted link becomes a card in Facebook or Zalo rather
	* than a bare page:
	*
	* /share/i/<image_id> interior photos (the gallery that image came from)
	* /share/p/<sku> products (name, price and gallery)
	*
	* The product route did not exist until it was added there; before that the
	* only product-ish URL was `?product=<sku>`, which is an OG-tag handler and
	* renders nothing but the chat page. `/share/p/` resolves a sku, a model code or
	* a slug against the same catalogue the Space already loads, so all 3,966
	* identity forms in it resolve.
	*
	* Returns null when there is no id to share, so the caller can omit the button
	* rather than render a link that would 404.
	*/
	const VAI_STUDIO = 'https://bep40-vai-avatar.hf.space';

	/**
	* The prefix avatar-interiors.js puts on a vai-avatar photograph's `sku`.
	*
	* Namespacing it there is deliberate -- it is what stops a 16-hex image_id
	* colliding with a baked interior's id in merge.js's de-duplication set, and it
	* is what lets a photograph travel the same code path as a catalogue product
	* (same viewer, same panel, same neon, same share anchor).
	*
	* It is NOT a catalogue identity, and /share/p/ resolves against the product
	* catalogue -- so a photograph sent through that route answers "Not Found".
	* Nothing about that is fixable on vai-avatar's side: no SKU of that shape
	* exists there to find. It has to be routed as the interior it is.
	*/
	const AVATAR_SKU_PREFIX = 'avatarimg:';

	/**
	* Whether this record is an interior photograph published on vai-avatar.
	*
	* Either field is conclusive on its own. `avatarImageId` is set on every such
	* photograph; the namespaced `sku` covers a record that has lost it.
	*/
	function isAvatarInterior(photo) {
	return Boolean(photo.avatarImageId) \|\| String(photo.sku ?? '').startsWith(AVATAR_SKU_PREFIX);
	}

	/**
	* The interior id behind a record, or '' when it has none.
	*
	* Two sources, in order, and the second is a fallback rather than an
	* alternative: `avatarImageId` is the field avatar-interiors.js sets, present on
	* every vai-avatar photograph. But a record can reach shareURL() having lost it
	* -- merge.js rewrites `id` into a namespaced tile key, and any path that
	* rebuilds a record from its `sku` alone would carry the prefix but not the
	* original field. That is exactly the shape that produced
	* /share/p/avatarimg%3Ab68dc779370dc70e, which is a "Not Found" page for every
	* single photograph in the vai-avatar interior set.
	*
	* So the namespaced sku is unwrapped as well, and a prefix that is present but
	* empty (`avatarimg:` with nothing after it) yields ''.
	*/
	function interiorIdOf(photo) {
	const direct = String(photo.avatarImageId ?? '').trim();
	if (direct) return direct;
	const sku = String(photo.sku ?? '').trim();
	return sku.startsWith(AVATAR_SKU_PREFIX)
	? sku.slice(AVATAR_SKU_PREFIX.length).trim()
	: '';
	}

	export function shareURL(photo) {
	if (!photo \|\| typeof photo !== 'object') return null;
	/**
	* A photograph published on vai-avatar carries BOTH `avatarImageId` and a
	* namespaced `sku` of `avatarimg:<image_id>`.
	*
	* `sku` exists so the record travels the same path as a catalogue product --
	* same viewer, same detail panel, same neon, same share anchor -- and it has
	* to stay for that. But it is not an identity the product catalogue knows,
	* and it must not decide the URL: `/share/p/` resolves against the product
	* catalogue, so every such photograph sent through it answered "Not Found".
	* The whole vai-avatar interior set was affected, on the tile's share button
	* and on the viewer's "Xem trang" alike.
	*
	* The interior identity is therefore resolved first and unconditionally, and
	* /share/i/ gets the raw id -- no prefix to strip, and nothing to keep in
	* step with whatever namespace merge.js uses for its tile ids.
	*/
	const interiorId = interiorIdOf(photo);
	if (interiorId) return `${VAI_STUDIO}/share/i/${encodeURIComponent(interiorId)}`;
	/**
	* A record that still says it is an interior but carries no usable id gets NO
	* share URL, rather than falling through to the product branch below.
	*
	* That fall-through is a trap worth naming, because it is exactly how the bug
	* comes back. `avatarimg:` with nothing after it has no id to share, and every
	* remaining branch here would still produce a URL for it: the product branch
	* would encode the prefix itself into `/share/p/avatarimg%3A`, which is the
	* dead route all over again, and the id branch would send it to `/share/i/`
	* with an empty segment, which 404s in its own right. Returning null lets the
	* caller omit the button instead, which is the only outcome that is true.
	*/
	if (isAvatarInterior(photo)) return null;
	if (photo.sku \|\| photo.slug) {
	// slug first, then sku, then mod. A shared URL is a permanent, public,
	// indexable address and the one thing about a product a reader can see:
	// `MH800TC` says nothing about what is linked, while
	// `mh800tc-malloca-may-hut-mui-gan-tu-am-tu-80cm` reads as the product's name
	// to a person and to a crawler. The slug is already the canonical key this
	// Space's catalogue uses, and vai-avatar's /share/p/ matches on it first, so
	// this changes how the link reads, not whether it resolves.
	const id = String(photo.slug \|\| photo.sku \|\| photo.mod \|\| '').trim();
	return id ? `${VAI_STUDIO}/share/p/${encodeURIComponent(id)}` : null;
	}
	const id = String(photo.id ?? '').trim();
	return id ? `${VAI_STUDIO}/share/i/${encodeURIComponent(id)}` : null;
	}