Harden vision vector search and embedding jobs.

Add point search, circuit-breaking, and safer gateway limits so artwork similarity lookups fail closed instead of hanging the request path.
This commit is contained in:
test
2026-09-20 14:49:19 +02:00
parent 455b514d50
commit 62aaa2f704
12 changed files with 903 additions and 64 deletions
@@ -0,0 +1,109 @@
<?php
declare(strict_types=1);
use App\Models\Artwork;
use Illuminate\Foundation\Testing\RefreshDatabase;
use Illuminate\Http\Client\ConnectionException;
use Illuminate\Support\Facades\Cache;
use Illuminate\Support\Facades\Http;
use Illuminate\Support\Facades\Storage;
use function Pest\Laravel\getJson;
uses(RefreshDatabase::class);
beforeEach(function (): void {
config()->set('vision.vector_gateway.enabled', true);
config()->set('vision.vector_gateway.base_url', 'https://vision.klevze.net');
config()->set('vision.vector_gateway.api_key', 'test-key');
config()->set('vision.vector_gateway.search_point_endpoint', '/vectors/search/point');
config()->set('cdn.files_url', 'https://files.skinbase.org');
config()->set('app.url', 'https://skinbase.test');
Cache::flush();
Storage::fake('public');
});
function pointSearchArtwork(): Artwork
{
return Artwork::factory()->create([
'hash' => 'point-source-hash',
'thumb_ext' => 'webp',
'is_public' => true,
'is_approved' => true,
'published_at' => now()->subHour(),
]);
}
it('uses the stored point and does not fetch or embed the source image', function (): void {
$source = pointSearchArtwork();
$match = Artwork::factory()->create(['title' => 'Point match']);
Http::fake([
'https://vision.klevze.net/vectors/search/point' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
['id' => $match->id, 'score' => 0.91],
],
], 200),
'https://files.skinbase.org/*' => Http::response('should not download', 500),
'https://vision.klevze.net/vectors/search/file' => Http::response(['error' => 'should not embed'], 500),
'https://vision.klevze.net/vectors/search' => Http::response(['error' => 'should not search by URL'], 500),
]);
getJson('/api/art/' . $source->id . '/similar-ai')
->assertOk()
->assertJsonPath('data.0.id', $match->id);
Http::assertSent(fn ($request): bool => str_contains($request->url(), '/vectors/search/point')
&& ($request->data()['id'] ?? null) === $source->id);
Http::assertNotSent(fn ($request): bool => str_contains($request->url(), '/vectors/search/file'));
Http::assertNotSent(fn ($request): bool => str_contains($request->url(), '/vectors/search') && ! str_contains($request->url(), '/point'));
Http::assertNotSent(fn ($request): bool => str_contains($request->url(), 'files.skinbase.org'));
});
it('falls back to the existing image path only when the point is missing', function (): void {
$source = pointSearchArtwork();
$match = Artwork::factory()->create(['title' => 'Fallback match']);
Http::fake([
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'missing'], 404),
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/file' => Http::response([
'results' => [['id' => $match->id, 'score' => 0.8]],
], 200),
]);
getJson('/api/art/' . $source->id . '/similar-ai')
->assertOk()
->assertJsonPath('data.0.id', $match->id);
});
it('does not fall back to CLIP after a point-search server failure', function (): void {
$source = pointSearchArtwork();
Http::fake([
'https://vision.klevze.net/vectors/search/point' => Http::response(['error' => 'upstream'], 500),
'https://files.skinbase.org/*' => Http::response('should not download', 500),
'https://vision.klevze.net/vectors/search/file' => Http::response(['error' => 'should not embed'], 500),
'https://vision.klevze.net/vectors/search' => Http::response(['error' => 'should not search by URL'], 500),
]);
getJson('/api/art/' . $source->id . '/similar-ai')->assertStatus(502);
Http::assertNotSent(fn ($request): bool => str_contains($request->url(), 'files.skinbase.org'));
Http::assertNotSent(fn ($request): bool => str_contains($request->url(), '/vectors/search/file'));
});
it('does not fall back to CLIP after a point-search timeout', function (): void {
$source = pointSearchArtwork();
Http::fake(function ($request) {
if (str_contains($request->url(), '/vectors/search/point')) {
throw new ConnectionException('cURL error 28: Operation timed out');
}
throw new RuntimeException('fallback must not be called');
});
getJson('/api/art/' . $source->id . '/similar-ai')->assertStatus(502);
Http::assertNothingSent();
});
@@ -46,6 +46,7 @@ it('returns AI similar artworks for a public artwork', function (): void {
Http::fake([
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search/file' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
@@ -84,6 +85,7 @@ it('falls back to URL search when the file vector endpoint fails for similar-ai'
Http::fake([
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search/file' => Http::response(['error' => 'missing endpoint'], 404),
'https://vision.klevze.net/vectors/search' => Http::response([
'results' => [
@@ -141,4 +143,4 @@ it('searches by uploaded image through the vector gateway', function (): void {
return $request->url() === 'https://vision.klevze.net/vectors/search/file'
&& $request->hasHeader('X-API-Key', 'test-key');
});
});
});
@@ -18,7 +18,9 @@ beforeEach(function (): void {
config()->set('vision.vector_gateway.base_url', 'https://vision.klevze.net');
config()->set('vision.vector_gateway.api_key', 'test-key');
config()->set('vision.vector_gateway.upsert_endpoint', '/vectors/upsert');
config()->set('vision.vector_gateway.upsert_file_endpoint', '/vectors/upsert/file');
config()->set('vision.vector_gateway.search_endpoint', '/vectors/search');
config()->set('vision.vector_gateway.search_point_endpoint', '/vectors/search/point');
config()->set('vision.image_variant', 'md');
config()->set('cdn.files_url', 'https://files.skinbase.org');
});
@@ -49,35 +51,110 @@ it('indexes artworks into the vector gateway with artwork metadata', function ()
$artwork->tags()->attach($tag->id, ['source' => 'ai', 'confidence' => 0.88]);
Http::fake([
'https://vision.klevze.net/vectors/upsert' => Http::response(['ok' => true], 200),
'https://files.skinbase.org/*'
=> Http::response(
'fake-image-bytes',
200,
['Content-Type' => 'image/webp']
),
'https://vision.klevze.net/vectors/upsert/file'
=> Http::response([
'id' => $artwork->id,
'collection' => 'images',
'dim' => 512,
], 200),
]);
artisan('artworks:vectors-index', ['--limit' => 1])
->assertSuccessful();
$artwork->refresh();
expect($artwork->last_vector_indexed_at)->not->toBeNull();
expect($artwork->last_vector_indexed_at)
->not->toBeNull();
Http::assertSentCount(2);
Http::assertSent(function ($request): bool {
return $request->method() === 'GET'
&& $request->url()
=== 'https://files.skinbase.org/artworks/md/aa/bb/aabbcc112233.webp';
});
Http::assertSent(function ($request) use ($artwork): bool {
if ($request->url() !== 'https://vision.klevze.net/vectors/upsert') {
if (
$request->url()
!== 'https://vision.klevze.net/vectors/upsert/file'
) {
return false;
}
$payload = json_decode($request->body(), true);
$parts = collect(
$request->data()
)->keyBy('name');
return $request->hasHeader('X-API-Key', 'test-key')
&& is_array($payload)
&& ($payload['id'] ?? null) === (string) $artwork->id
&& ($payload['url'] ?? null) === 'https://files.skinbase.org/artworks/md/aa/bb/aabbcc112233.webp'
&& ($payload['metadata']['content_type'] ?? null) === 'Photography'
&& ($payload['metadata']['category'] ?? null) === 'Abstract'
&& ($payload['metadata']['tags'] ?? null) === ['skyline']
&& array_key_exists('is_public', $payload['metadata'])
&& array_key_exists('is_deleted', $payload['metadata'])
&& array_key_exists('is_nsfw', $payload['metadata'])
&& array_key_exists('category_id', $payload['metadata'])
&& array_key_exists('content_type_id', $payload['metadata'])
&& array_key_exists('status', $payload['metadata']);
$idPart = $parts->get('id', []);
$metadataPart = $parts->get(
'metadata_json',
[]
);
$filePart = $parts->get('file', []);
$metadata = json_decode(
(string) (
$metadataPart['contents'] ?? ''
),
true
);
return $request->hasHeader(
'X-API-Key',
'test-key'
)
&& (string) (
$idPart['contents'] ?? ''
) === (string) $artwork->id
&& (string) (
$filePart['contents'] ?? ''
) === 'fake-image-bytes'
&& (string) (
$filePart['filename'] ?? ''
) === sprintf(
'artwork-%d.webp',
$artwork->id
)
&& is_array($metadata)
&& ($metadata['content_type'] ?? null)
=== 'Photography'
&& ($metadata['category'] ?? null)
=== 'Abstract'
&& ($metadata['tags'] ?? null)
=== ['skyline']
&& array_key_exists(
'is_public',
$metadata
)
&& array_key_exists(
'is_deleted',
$metadata
)
&& array_key_exists(
'is_nsfw',
$metadata
)
&& array_key_exists(
'category_id',
$metadata
)
&& array_key_exists(
'content_type_id',
$metadata
)
&& array_key_exists(
'status',
$metadata
);
});
});
@@ -113,12 +190,19 @@ it('searches similar artworks through the vector gateway', function (): void {
$similar->categories()->attach($category->id);
Http::fake([
'https://vision.klevze.net/vectors/search' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
['id' => $similar->id, 'score' => 0.9876],
],
], 200),
'https://vision.klevze.net/vectors/search/point'
=> Http::response([
'results' => [
[
'id' => $source->id,
'score' => 1.0,
],
[
'id' => $similar->id,
'score' => 0.9876,
],
],
], 200),
]);
artisan('artworks:vectors-search', [
@@ -133,6 +217,25 @@ it('searches similar artworks through the vector gateway', function (): void {
'category' => 'Nature',
]])
->assertSuccessful();
Http::assertSentCount(1);
Http::assertSent(
function ($request) use ($source): bool {
return $request->url()
=== 'https://vision.klevze.net/vectors/search/point'
&& ($request->data()['id'] ?? null)
=== $source->id;
}
);
Http::assertNotSent(
fn ($request): bool =>
str_contains(
$request->url(),
'files.skinbase.org'
)
);
});
it('can re-upsert only artworks that already have local embeddings', function (): void {
@@ -184,28 +287,81 @@ it('can re-upsert only artworks that already have local embeddings', function ()
$nonEmbeddedArtwork->categories()->attach($category->id);
Http::fake([
'https://vision.klevze.net/vectors/upsert' => Http::response(['ok' => true], 200),
'https://files.skinbase.org/*'
=> Http::response(
'fake-image-bytes',
200,
['Content-Type' => 'image/webp']
),
'https://vision.klevze.net/vectors/upsert/file'
=> Http::response([
'id' => $embeddedArtwork->id,
'collection' => 'images',
'dim' => 512,
], 200),
]);
artisan('artworks:vectors-index', ['--embedded-only' => true, '--limit' => 10])
->assertSuccessful();
artisan('artworks:vectors-index', [
'--embedded-only' => true,
'--limit' => 10,
])->assertSuccessful();
$embeddedArtwork->refresh();
$nonEmbeddedArtwork->refresh();
expect($embeddedArtwork->last_vector_indexed_at)->not->toBeNull()
->and($nonEmbeddedArtwork->last_vector_indexed_at)->toBeNull();
expect($embeddedArtwork->last_vector_indexed_at)
->not->toBeNull()
->and(
$nonEmbeddedArtwork->last_vector_indexed_at
)
->toBeNull();
Http::assertSentCount(1);
Http::assertSent(function ($request) use ($embeddedArtwork): bool {
if ($request->url() !== 'https://vision.klevze.net/vectors/upsert') {
return false;
/*
* One indexed artwork:
* 1x CDN image download
* 1x /vectors/upsert/file
*/
Http::assertSentCount(2);
Http::assertSent(
function ($request): bool {
return $request->method() === 'GET'
&& $request->url()
=== 'https://files.skinbase.org/artworks/md/11/22/112233445566.webp';
}
);
$payload = json_decode($request->body(), true);
Http::assertSent(
function ($request) use (
$embeddedArtwork
): bool {
if (
$request->url()
!== 'https://vision.klevze.net/vectors/upsert/file'
) {
return false;
}
return is_array($payload)
&& ($payload['id'] ?? null) === (string) $embeddedArtwork->id
&& ($payload['url'] ?? null) === 'https://files.skinbase.org/artworks/md/11/22/112233445566.webp';
});
$parts = collect(
$request->data()
)->keyBy('name');
$idPart = $parts->get('id', []);
$filePart = $parts->get('file', []);
return (string) (
$idPart['contents'] ?? ''
) === (string) $embeddedArtwork->id
&& (string) (
$filePart['contents'] ?? ''
) === 'fake-image-bytes'
&& (string) (
$filePart['filename'] ?? ''
) === sprintf(
'artwork-%d.webp',
$embeddedArtwork->id
);
}
);
});
@@ -21,15 +21,14 @@ it('persists a normalized embedding and upserts the artwork to the vector gatewa
config()->set('vision.vector_gateway.enabled', true);
config()->set('vision.vector_gateway.base_url', 'https://vision.local');
config()->set('vision.vector_gateway.api_key', 'test-key');
config()->set('vision.vector_gateway.upsert_file_endpoint', '/vectors/upsert/file');
config()->set('vision.vector_gateway.upsert_vector_endpoint', '/vectors/upsert/vector');
config()->set('cdn.files_url', 'https://files.local');
Http::fake([
'https://clip.local/embed' => Http::response([
'embedding' => [3.0, 4.0],
], 200),
'https://files.local/*' => Http::response('fake-image-bytes', 200),
'https://vision.local/vectors/upsert/file' => Http::response([
'https://vision.local/vectors/upsert/vector' => Http::response([
'status' => 'ok',
], 200),
]);
@@ -82,7 +81,7 @@ it('persists a normalized embedding and upserts the artwork to the vector gatewa
return str_contains($request->url(), 'vision.local/vectors/upsert');
});
Http::assertSentCount(3);
Http::assertSentCount(2);
});
it('keeps the local embedding when vector upsert fails', function () {
@@ -93,15 +92,14 @@ it('keeps the local embedding when vector upsert fails', function () {
config()->set('vision.vector_gateway.enabled', true);
config()->set('vision.vector_gateway.base_url', 'https://vision.local');
config()->set('vision.vector_gateway.api_key', 'test-key');
config()->set('vision.vector_gateway.upsert_file_endpoint', '/vectors/upsert/file');
config()->set('vision.vector_gateway.upsert_vector_endpoint', '/vectors/upsert/vector');
config()->set('cdn.files_url', 'https://files.local');
Http::fake([
'https://clip.local/embed' => Http::response([
'embedding' => [1.0, 2.0, 2.0],
], 200),
'https://files.local/*' => Http::response('fake-image-bytes', 200),
'https://vision.local/vectors/upsert/file' => Http::response([
'https://vision.local/vectors/upsert/vector' => Http::response([
'message' => 'gateway error',
], 500),
]);
@@ -122,5 +120,286 @@ it('keeps the local embedding when vector upsert fails', function () {
expect(ArtworkEmbedding::query()->where('artwork_id', $artwork->id)->exists())->toBeTrue()
->and($artwork->last_vector_indexed_at)->toBeNull();
Http::assertSentCount(3);
Http::assertSentCount(2);
});
it('reuses the stored embedding when vector indexing is retried', function () {
config()->set('recommendations.embedding.enabled', true);
config()->set('recommendations.embedding.endpoint', '/embed');
config()->set('recommendations.embedding.min_dim', 2);
config()->set('vision.clip.base_url', 'https://clip.local');
config()->set('vision.clip.api_key', 'clip-key');
config()->set('vision.vector_gateway.enabled', true);
config()->set('vision.vector_gateway.base_url', 'https://vision.local');
config()->set('vision.vector_gateway.api_key', 'test-key');
config()->set(
'vision.vector_gateway.upsert_vector_endpoint',
'/vectors/upsert/vector'
);
config()->set('cdn.files_url', 'https://files.local');
$clipCalls = 0;
$vectorCalls = 0;
Http::fake(function (
\Illuminate\Http\Client\Request $request
) use (&$clipCalls, &$vectorCalls) {
if (
$request->url()
=== 'https://clip.local/embed'
) {
$clipCalls++;
return Http::response([
'vector' => [3.0, 4.0],
], 200);
}
if (
$request->url()
=== 'https://vision.local/vectors/upsert/vector'
) {
$vectorCalls++;
if ($vectorCalls === 1) {
return Http::response([
'message' => 'temporary gateway error',
], 500);
}
return Http::response([
'id' => 123,
'collection' => 'images',
'dim' => 2,
], 200);
}
return Http::response([
'error' => 'unexpected request',
], 500);
});
$artwork = Artwork::factory()->create([
'hash' => '9988776655443322',
'thumb_ext' => 'webp',
'is_public' => true,
'is_approved' => true,
'published_at' => now()->subDay(),
]);
$first = new GenerateArtworkEmbeddingJob(
$artwork->id,
'9988776655443322'
);
app()->call([
$first,
'handle',
]);
$artwork->refresh();
expect(
ArtworkEmbedding::query()
->where('artwork_id', $artwork->id)
->exists()
)->toBeTrue()
->and($artwork->last_vector_indexed_at)
->toBeNull()
->and($clipCalls)
->toBe(1)
->and($vectorCalls)
->toBe(1);
$second = new GenerateArtworkEmbeddingJob(
$artwork->id,
'9988776655443322'
);
app()->call([
$second,
'handle',
]);
$artwork->refresh();
expect($artwork->last_vector_indexed_at)
->not->toBeNull()
->and($clipCalls)
->toBe(1)
->and($vectorCalls)
->toBe(2);
Http::assertSent(function (
\Illuminate\Http\Client\Request $request
): bool {
if (
$request->url()
!== 'https://vision.local/vectors/upsert/vector'
) {
return false;
}
$payload = $request->data();
return isset($payload['vector'])
&& count($payload['vector']) === 2
&& round(
(float) $payload['vector'][0],
4
) === 0.6
&& round(
(float) $payload['vector'][1],
4
) === 0.8;
});
});
it('recovers when the vector marker predates the stored embedding', function () {
config()->set('recommendations.embedding.enabled', true);
config()->set('recommendations.embedding.endpoint', '/embed');
config()->set('recommendations.embedding.min_dim', 2);
config()->set(
'vision.clip.base_url',
'https://clip.local'
);
config()->set(
'vision.clip.api_key',
'clip-key'
);
config()->set(
'vision.vector_gateway.enabled',
true
);
config()->set(
'vision.vector_gateway.base_url',
'https://vision.local'
);
config()->set(
'vision.vector_gateway.api_key',
'test-key'
);
config()->set(
'vision.vector_gateway.upsert_vector_endpoint',
'/vectors/upsert/vector'
);
$clipCalls = 0;
$vectorCalls = 0;
Http::fake(function (
\Illuminate\Http\Client\Request $request
) use (&$clipCalls, &$vectorCalls) {
if (
$request->url()
=== 'https://clip.local/embed'
) {
$clipCalls++;
return Http::response([
'vector' => [9.0, 9.0],
], 200);
}
if (
$request->url()
=== 'https://vision.local/vectors/upsert/vector'
) {
$vectorCalls++;
return Http::response([
'id' => 123,
'collection' => 'images',
'dim' => 2,
], 200);
}
return Http::response([
'error' => 'unexpected request',
], 500);
});
$oldMarker = now()->subMinutes(10);
$embeddingGeneratedAt = now()->subMinute();
$artwork = Artwork::factory()->create([
'hash' => 'aabbccddeeff00112233445566778899aabbccddeeff00112233445566778899',
'thumb_ext' => 'webp',
'is_public' => true,
'is_approved' => true,
'published_at' => now()->subDay(),
'last_vector_indexed_at' => $oldMarker,
]);
ArtworkEmbedding::query()->create([
'artwork_id' => $artwork->id,
'model' => 'clip',
'model_version' => 'v1',
'algo_version' => 'clip-cosine-v1',
'dim' => 2,
'embedding_json' => json_encode(
[0.6, 0.8],
JSON_THROW_ON_ERROR
),
'source_hash' => 'aabbccddeeff00112233445566778899aabbccddeeff00112233445566778899',
'is_normalized' => true,
'generated_at' => $embeddingGeneratedAt,
'meta' => ['source' => 'clip'],
]);
$job = new GenerateArtworkEmbeddingJob(
$artwork->id,
'aabbccddeeff00112233445566778899aabbccddeeff00112233445566778899'
);
app()->call([
$job,
'handle',
]);
$artwork->refresh();
expect($clipCalls)
->toBe(0)
->and($vectorCalls)
->toBe(1)
->and($artwork->last_vector_indexed_at)
->not->toBeNull()
->and(
$artwork->last_vector_indexed_at
->greaterThan($oldMarker)
)
->toBeTrue();
Http::assertSent(function (
\Illuminate\Http\Client\Request $request
): bool {
if (
$request->url()
!== 'https://vision.local/vectors/upsert/vector'
) {
return false;
}
$payload = $request->data();
return isset($payload['vector'])
&& count($payload['vector']) === 2
&& round(
(float) $payload['vector'][0],
4
) === 0.6
&& round(
(float) $payload['vector'][1],
4
) === 0.8;
});
});
@@ -56,6 +56,7 @@ function fakeSuccessfulUrlSearch(Artwork $source, Artwork $match): array
{
return [
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
@@ -70,6 +71,7 @@ it('does not call URL search when file search succeeds and leaves the circuit cl
Http::fake([
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search/file' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
@@ -143,6 +145,7 @@ it('opens the circuit when both search strategies fail transiently', function ()
Http::fake([
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search/file' => Http::response(['error' => 'down'], 503),
'https://vision.klevze.net/vectors/search' => Http::response(['error' => 'down'], 503),
]);
@@ -353,6 +356,7 @@ it('serves similar-ai from cache without contacting the gateway', function (): v
Http::fake([
'https://files.skinbase.org/*' => Http::response('image-bytes', 200, ['Content-Type' => 'image/webp']),
'https://vision.klevze.net/vectors/search/point' => Http::response(['detail' => 'Point not found'], 404),
'https://vision.klevze.net/vectors/search/file' => Http::response([
'results' => [
['id' => $source->id, 'score' => 1.0],
@@ -38,6 +38,9 @@ function fakeFileSearchOk(): void
{
Http::fake(function ($request) {
$url = $request->url();
if (str_contains($url, '/vectors/search/point')) {
return Http::response(['detail' => 'Point not found'], 404);
}
if (str_contains($url, '/vectors/search')) {
return Http::response(['results' => []], 200);
}