405 lines
13 KiB
PHP
405 lines
13 KiB
PHP
<?php
|
|
|
|
namespace App\Services;
|
|
|
|
use App\Models\EmbeddingEntry;
|
|
use Illuminate\Support\Collection;
|
|
use Illuminate\Support\Facades\DB;
|
|
use Illuminate\Support\Facades\Http;
|
|
use InvalidArgumentException;
|
|
use Laravel\Ai\Ai;
|
|
use Laravel\Ai\Embeddings;
|
|
use Laravel\Ai\Enums\Lab;
|
|
use RuntimeException;
|
|
use Throwable;
|
|
|
|
class EmbeddingWorkbench
|
|
{
|
|
/**
|
|
* @var Collection<int, string>|null
|
|
*/
|
|
protected ?Collection $ollamaModelNames = null;
|
|
|
|
public function providerName(): string
|
|
{
|
|
$provider = config('ai.default_for_embeddings', 'ollama');
|
|
|
|
return $provider instanceof Lab ? $provider->value : (string) $provider;
|
|
}
|
|
|
|
public function modelName(?string $model = null): string
|
|
{
|
|
$model = trim((string) $model);
|
|
|
|
if ($model !== '') {
|
|
return $this->validateModel($model);
|
|
}
|
|
|
|
$default = Ai::embeddingProvider($this->providerName())->defaultEmbeddingsModel();
|
|
$availableModels = $this->availableModelNames();
|
|
|
|
if ($availableModels->contains($default)) {
|
|
return $default;
|
|
}
|
|
|
|
$firstAvailableModel = $availableModels->first();
|
|
|
|
if ($firstAvailableModel === null) {
|
|
throw new InvalidArgumentException('No embedding models are available.');
|
|
}
|
|
|
|
return $firstAvailableModel;
|
|
}
|
|
|
|
public function dimensions(?string $model = null): int
|
|
{
|
|
$this->modelName($model);
|
|
|
|
return Ai::embeddingProvider($this->providerName())->defaultEmbeddingsDimensions();
|
|
}
|
|
|
|
/**
|
|
* @return array<int, array{name: string, dimensions: int, default: bool}>
|
|
*/
|
|
public function modelOptions(): array
|
|
{
|
|
$default = Ai::embeddingProvider($this->providerName())->defaultEmbeddingsModel();
|
|
|
|
return $this->availableModelNames()
|
|
->map(fn (string $model): array => [
|
|
'name' => $model,
|
|
'dimensions' => $this->dimensions($model),
|
|
'default' => $model === $default,
|
|
])
|
|
->all();
|
|
}
|
|
|
|
/**
|
|
* @param list<string> $phrases
|
|
* @return Collection<int, EmbeddingEntry>
|
|
*/
|
|
public function embed(array $phrases, ?string $model = null): Collection
|
|
{
|
|
$phrases = $this->uniquePhrases($phrases);
|
|
|
|
if ($phrases === []) {
|
|
return collect();
|
|
}
|
|
|
|
$provider = $this->providerName();
|
|
$model = $this->modelName($model);
|
|
$dimensions = $this->dimensions($model);
|
|
$hashes = collect($phrases)->mapWithKeys(fn (string $phrase): array => [
|
|
EmbeddingVector::hashPhrase($phrase) => $phrase,
|
|
]);
|
|
|
|
$entries = EmbeddingEntry::query()
|
|
->where('provider', $provider)
|
|
->where('model', $model)
|
|
->where('embedding_dimensions', $dimensions)
|
|
->whereIn('content_hash', $hashes->keys())
|
|
->get()
|
|
->keyBy('content_hash');
|
|
|
|
$missingPhrases = $hashes
|
|
->reject(fn (string $phrase, string $hash): bool => $entries->has($hash))
|
|
->values()
|
|
->all();
|
|
|
|
if ($missingPhrases !== []) {
|
|
$response = Embeddings::for($missingPhrases)
|
|
->dimensions($dimensions)
|
|
->cache()
|
|
->timeout(60)
|
|
->generate($provider, $model);
|
|
|
|
foreach ($missingPhrases as $index => $phrase) {
|
|
$embedding = array_map('floatval', $response->embeddings[$index] ?? []);
|
|
|
|
$this->ensureExpectedDimensions($embedding, $dimensions);
|
|
|
|
$entry = EmbeddingEntry::query()->updateOrCreate(
|
|
[
|
|
'content_hash' => EmbeddingVector::hashPhrase($phrase),
|
|
'provider' => $response->meta->provider,
|
|
'model' => $response->meta->model,
|
|
'embedding_dimensions' => $dimensions,
|
|
],
|
|
[
|
|
'source_text' => $phrase,
|
|
'embedding' => $embedding,
|
|
'tokens' => $response->tokens,
|
|
]
|
|
);
|
|
|
|
$entries->put($entry->content_hash, $entry);
|
|
}
|
|
}
|
|
|
|
return collect($phrases)
|
|
->map(fn (string $phrase): ?EmbeddingEntry => $entries->get(EmbeddingVector::hashPhrase($phrase)))
|
|
->filter()
|
|
->values();
|
|
}
|
|
|
|
/**
|
|
* @return Collection<int, array{entry: EmbeddingEntry, similarity: float, distance: float}>
|
|
*/
|
|
public function search(string $query, int $limit = 10, float $minimumSimilarity = 0.3, ?string $model = null): Collection
|
|
{
|
|
$query = EmbeddingVector::cleanPhrase($query);
|
|
|
|
if ($query === '') {
|
|
return collect();
|
|
}
|
|
|
|
$limit = max(1, min(50, $limit));
|
|
$minimumSimilarity = max(0.0, min(1.0, $minimumSimilarity));
|
|
$provider = $this->providerName();
|
|
$model = $this->modelName($model);
|
|
$dimensions = $this->dimensions($model);
|
|
$queryEmbedding = $this->generateEmbedding($query, $model);
|
|
|
|
if ($this->canUseVectorQueries()) {
|
|
return EmbeddingEntry::query()
|
|
->select('embedding_entries.*')
|
|
->selectVectorDistance('embedding', $queryEmbedding, as: 'distance')
|
|
->where('provider', $provider)
|
|
->where('model', $model)
|
|
->where('embedding_dimensions', $dimensions)
|
|
->whereVectorSimilarTo('embedding', $queryEmbedding, minSimilarity: $minimumSimilarity)
|
|
->limit($limit)
|
|
->get()
|
|
->map(fn (EmbeddingEntry $entry): array => [
|
|
'entry' => $entry,
|
|
'similarity' => round(1 - (float) $entry->distance, 6),
|
|
'distance' => round((float) $entry->distance, 6),
|
|
]);
|
|
}
|
|
|
|
return EmbeddingEntry::query()
|
|
->where('provider', $provider)
|
|
->where('model', $model)
|
|
->where('embedding_dimensions', $dimensions)
|
|
->get()
|
|
->map(function (EmbeddingEntry $entry) use ($queryEmbedding): array {
|
|
$similarity = EmbeddingVector::cosineSimilarity($queryEmbedding, $entry->embedding);
|
|
|
|
return [
|
|
'entry' => $entry,
|
|
'similarity' => round($similarity, 6),
|
|
'distance' => round(1 - $similarity, 6),
|
|
];
|
|
})
|
|
->filter(fn (array $result): bool => $result['similarity'] >= $minimumSimilarity)
|
|
->sortByDesc('similarity')
|
|
->take($limit)
|
|
->values();
|
|
}
|
|
|
|
/**
|
|
* @param list<string> $phrases
|
|
* @return array{entries: Collection<int, EmbeddingEntry>, matrix: array<int, array{entry: EmbeddingEntry, scores: array<int, float>}>, closest_pair: array{first: EmbeddingEntry, second: EmbeddingEntry, similarity: float}|null}
|
|
*/
|
|
public function compare(array $phrases, ?string $model = null): array
|
|
{
|
|
$entries = $this->embed($phrases, $model)->values();
|
|
$matrix = [];
|
|
$closestPair = null;
|
|
|
|
foreach ($entries as $rowIndex => $entry) {
|
|
$scores = [];
|
|
|
|
foreach ($entries as $columnIndex => $comparedEntry) {
|
|
$similarity = round(EmbeddingVector::cosineSimilarity($entry->embedding, $comparedEntry->embedding), 6);
|
|
$scores[] = $similarity;
|
|
|
|
if ($rowIndex < $columnIndex && (
|
|
$closestPair === null || $similarity > $closestPair['similarity']
|
|
)) {
|
|
$closestPair = [
|
|
'first' => $entry,
|
|
'second' => $comparedEntry,
|
|
'similarity' => $similarity,
|
|
];
|
|
}
|
|
}
|
|
|
|
$matrix[] = [
|
|
'entry' => $entry,
|
|
'scores' => $scores,
|
|
];
|
|
}
|
|
|
|
return [
|
|
'entries' => $entries,
|
|
'matrix' => $matrix,
|
|
'closest_pair' => $closestPair,
|
|
];
|
|
}
|
|
|
|
/**
|
|
* @return array{id: int, source_text: string, provider: string, model: string, dimensions: int, tokens: int|null, vector_preview: array<int, float>, vector_json: string}
|
|
*/
|
|
public function presentEntry(EmbeddingEntry $entry): array
|
|
{
|
|
return [
|
|
'id' => $entry->id,
|
|
'source_text' => $entry->source_text,
|
|
'provider' => $entry->provider,
|
|
'model' => $entry->model,
|
|
'dimensions' => $entry->embedding_dimensions,
|
|
'tokens' => $entry->tokens,
|
|
'vector_preview' => EmbeddingVector::preview($entry->embedding),
|
|
'vector_json' => json_encode(EmbeddingVector::rounded($entry->embedding), JSON_PRETTY_PRINT | JSON_THROW_ON_ERROR),
|
|
];
|
|
}
|
|
|
|
/**
|
|
* @param array<int, string> $phrases
|
|
* @return list<string>
|
|
*/
|
|
protected function uniquePhrases(array $phrases): array
|
|
{
|
|
$unique = [];
|
|
|
|
foreach ($phrases as $phrase) {
|
|
$phrase = EmbeddingVector::cleanPhrase($phrase);
|
|
|
|
if ($phrase === '') {
|
|
continue;
|
|
}
|
|
|
|
$unique[EmbeddingVector::hashPhrase($phrase)] = $phrase;
|
|
}
|
|
|
|
return array_values($unique);
|
|
}
|
|
|
|
/**
|
|
* @return array<int, float>
|
|
*/
|
|
protected function generateEmbedding(string $text, ?string $model = null): array
|
|
{
|
|
$model = $this->modelName($model);
|
|
$dimensions = $this->dimensions($model);
|
|
$embedding = array_map(
|
|
'floatval',
|
|
Embeddings::for([$text])
|
|
->dimensions($dimensions)
|
|
->cache()
|
|
->timeout(60)
|
|
->generate($this->providerName(), $model)
|
|
->first()
|
|
);
|
|
|
|
$this->ensureExpectedDimensions($embedding, $dimensions);
|
|
|
|
return $embedding;
|
|
}
|
|
|
|
/**
|
|
* @param array<int, float> $embedding
|
|
*/
|
|
protected function ensureExpectedDimensions(array $embedding, int $dimensions): void
|
|
{
|
|
if (count($embedding) !== $dimensions) {
|
|
throw new RuntimeException('The embedding model returned '.count($embedding)." dimensions, but the database is configured for {$dimensions}.");
|
|
}
|
|
}
|
|
|
|
protected function canUseVectorQueries(): bool
|
|
{
|
|
return DB::connection()->getDriverName() === 'pgsql';
|
|
}
|
|
|
|
/**
|
|
* @return Collection<int, string>
|
|
*/
|
|
protected function availableModelNames(): Collection
|
|
{
|
|
if ($this->providerName() === 'ollama') {
|
|
$ollamaModels = $this->ollamaModelNames();
|
|
|
|
if ($ollamaModels->isNotEmpty()) {
|
|
return $ollamaModels;
|
|
}
|
|
}
|
|
|
|
return $this->configuredModelNames();
|
|
}
|
|
|
|
/**
|
|
* @return Collection<int, string>
|
|
*/
|
|
protected function configuredModelNames(): Collection
|
|
{
|
|
$models = config('ai.providers.'.$this->providerName().'.models.embeddings.available', []);
|
|
|
|
if (! is_array($models)) {
|
|
$models = [];
|
|
}
|
|
|
|
return collect($models)
|
|
->push(Ai::embeddingProvider($this->providerName())->defaultEmbeddingsModel())
|
|
->map(fn (mixed $model): string => trim((string) $model))
|
|
->filter()
|
|
->unique()
|
|
->values();
|
|
}
|
|
|
|
/**
|
|
* @return Collection<int, string>
|
|
*/
|
|
protected function ollamaModelNames(): Collection
|
|
{
|
|
if ($this->ollamaModelNames !== null) {
|
|
return $this->ollamaModelNames;
|
|
}
|
|
|
|
try {
|
|
$models = Http::baseUrl($this->ollamaUrl())
|
|
->acceptJson()
|
|
->connectTimeout(2)
|
|
->timeout(5)
|
|
->get('api/tags')
|
|
->throw()
|
|
->json('models', []);
|
|
} catch (Throwable) {
|
|
return $this->ollamaModelNames = collect();
|
|
}
|
|
|
|
if (! is_array($models)) {
|
|
return $this->ollamaModelNames = collect();
|
|
}
|
|
|
|
return $this->ollamaModelNames = collect($models)
|
|
->map(fn (mixed $model): string => is_array($model) ? trim((string) ($model['name'] ?? '')) : '')
|
|
->filter()
|
|
->unique()
|
|
->values();
|
|
}
|
|
|
|
protected function ollamaUrl(): string
|
|
{
|
|
return rtrim((string) config('ai.providers.ollama.url', 'http://localhost:11434'), '/');
|
|
}
|
|
|
|
protected function validateModel(string $model): string
|
|
{
|
|
$model = trim($model);
|
|
|
|
if ($model === '') {
|
|
throw new InvalidArgumentException('Choose an embedding model.');
|
|
}
|
|
|
|
$availableModels = $this->availableModelNames();
|
|
|
|
if ($availableModels->doesntContain($model)) {
|
|
throw new InvalidArgumentException("The embedding model [{$model}] is not available.");
|
|
}
|
|
|
|
return $model;
|
|
}
|
|
}
|