Vector Search & LLM Integration

app/Config/services.php

@@ -22,6 +22,18 @@
     // Callback URL for social authentication methods
     'callback_url' => env('APP_URL', false),
 
+    // LLM Service
+    // Options: openai
+    'llm' => env('LLM_SERVICE', ''),
+
+    // OpenAI API-compatible service details
+    'openai' => [
+        'endpoint' => env('OPENAI_ENDPOINT', 'https://api.openai.com'),
+        'key' => env('OPENAI_KEY', ''),
+        'embedding_model' => env('OPENAI_EMBEDDING_MODEL', 'text-embedding-3-small'),
+        'query_model' => env('OPENAI_QUERY_MODEL', 'gpt-4o'),
+    ],
+
     'github'   => [
         'client_id'     => env('GITHUB_APP_ID', false),
         'client_secret' => env('GITHUB_APP_SECRET', false),

app/Console/Commands/RegenerateVectorsCommand.php

@@ -0,0 +1,46 @@
+<?php
+
+namespace BookStack\Console\Commands;
+
+use BookStack\Entities\EntityProvider;
+use BookStack\Entities\Models\Entity;
+use BookStack\Search\Vectors\SearchVector;
+use BookStack\Search\Vectors\StoreEntityVectorsJob;
+use Illuminate\Console\Command;
+
+class RegenerateVectorsCommand extends Command
+{
+    /**
+     * The name and signature of the console command.
+     *
+     * @var string
+     */
+    protected $signature = 'bookstack:regenerate-vectors';
+
+    /**
+     * The console command description.
+     *
+     * @var string
+     */
+    protected $description = 'Re-index vectors for all content in the system';
+
+    /**
+     * Execute the console command.
+     */
+    public function handle(EntityProvider $entityProvider)
+    {
+        // TODO - Add confirmation before run regarding deletion/time/effort/api-cost etc...
+        SearchVector::query()->delete();
+
+        $types = $entityProvider->all();
+        foreach ($types as $type => $typeInstance) {
+            $this->info("Creating jobs to store vectors for {$type} data...");
+            /** @var Entity[] $entities  */
+            $typeInstance->newQuery()->chunkById(100, function ($entities) {
+                foreach ($entities as $entity) {
+                    dispatch(new StoreEntityVectorsJob($entity));
+                }
+            });
+        }
+    }
+}

app/Search/SearchController.php

@@ -6,6 +6,7 @@
 use BookStack\Entities\Queries\QueryPopular;
 use BookStack\Entities\Tools\SiblingFetcher;
 use BookStack\Http\Controller;
+use BookStack\Search\Vectors\VectorSearchRunner;
 use Illuminate\Http\Request;
 
 class SearchController extends Controller
@@ -139,4 +140,19 @@ public function searchSiblings(Request $request, SiblingFetcher $siblingFetcher)
 
         return view('entities.list-basic', ['entities' => $entities, 'style' => 'compact']);
     }
+
+    public function searchQuery(Request $request, VectorSearchRunner $runner)
+    {
+        $query = $request->get('query', '');
+
+        if ($query) {
+            $results = $runner->run($query);
+        } else {
+            $results = null;
+        }
+
+        return view('search.query', [
+            'results' => $results,
+        ]);
+    }
 }

app/Search/SearchIndex.php

@@ -6,6 +6,8 @@
 use BookStack\Entities\EntityProvider;
 use BookStack\Entities\Models\Entity;
 use BookStack\Entities\Models\Page;
+use BookStack\Search\Vectors\StoreEntityVectorsJob;
+use BookStack\Search\Vectors\VectorQueryServiceProvider;
 use BookStack\Util\HtmlDocument;
 use DOMNode;
 use Illuminate\Database\Eloquent\Builder;
@@ -25,7 +27,7 @@ class SearchIndex
     public static string $softDelimiters = ".-";
 
     public function __construct(
-        protected EntityProvider $entityProvider
+        protected EntityProvider $entityProvider,
     ) {
     }
 
@@ -37,6 +39,10 @@ public function indexEntity(Entity $entity): void
         $this->deleteEntityTerms($entity);
         $terms = $this->entityToTermDataArray($entity);
         $this->insertTerms($terms);
+
+        if (VectorQueryServiceProvider::isEnabled()) {
+            dispatch(new StoreEntityVectorsJob($entity));
+        }
     }
 
     /**
@@ -47,9 +53,15 @@ public function indexEntity(Entity $entity): void
     public function indexEntities(array $entities): void
     {
         $terms = [];
+        $vectorQueryEnabled = VectorQueryServiceProvider::isEnabled();
+
         foreach ($entities as $entity) {
             $entityTerms = $this->entityToTermDataArray($entity);
             array_push($terms, ...$entityTerms);
+
+            if ($vectorQueryEnabled) {
+                dispatch(new StoreEntityVectorsJob($entity));
+            }
         }
 
         $this->insertTerms($terms);

app/Search/Vectors/EntityVectorGenerator.php

@@ -0,0 +1,84 @@
+<?php
+
+namespace BookStack\Search\Vectors;
+
+use BookStack\Entities\Models\Entity;
+use BookStack\Search\Vectors\Services\VectorQueryService;
+use Illuminate\Support\Facades\DB;
+
+class EntityVectorGenerator
+{
+    public function __construct(
+        protected VectorQueryServiceProvider $vectorQueryServiceProvider
+    ) {
+    }
+
+    public function generateAndStore(Entity $entity): void
+    {
+        $vectorService = $this->vectorQueryServiceProvider->get();
+
+        $text = $this->entityToPlainText($entity);
+        $chunks = $this->chunkText($text);
+        $embeddings = $this->chunksToEmbeddings($chunks, $vectorService);
+
+        $this->deleteExistingEmbeddingsForEntity($entity);
+        $this->storeEmbeddings($embeddings, $chunks, $entity);
+    }
+
+    protected function deleteExistingEmbeddingsForEntity(Entity $entity): void
+    {
+        SearchVector::query()
+            ->where('entity_type', '=', $entity->getMorphClass())
+            ->where('entity_id', '=', $entity->id)
+            ->delete();
+    }
+
+    protected function storeEmbeddings(array $embeddings, array $textChunks, Entity $entity): void
+    {
+        $toInsert = [];
+
+        foreach ($embeddings as $index => $embedding) {
+            $text = $textChunks[$index];
+            $toInsert[] = [
+                'entity_id' => $entity->id,
+                'entity_type' => $entity->getMorphClass(),
+                'embedding' => DB::raw('VEC_FROMTEXT("[' . implode(',', $embedding) . ']")'),
+                'text' => $text,
+            ];
+        }
+
+        // TODO - Chunk inserts
+        SearchVector::query()->insert($toInsert);
+    }
+
+    /**
+     * @param string[] $chunks
+     * @return float[] array
+     */
+    protected function chunksToEmbeddings(array $chunks, VectorQueryService $vectorQueryService): array
+    {
+        $embeddings = [];
+        foreach ($chunks as $index => $chunk) {
+            $embeddings[$index] = $vectorQueryService->generateEmbeddings($chunk);
+        }
+        return $embeddings;
+    }
+
+    /**
+     * @return string[]
+     */
+    protected function chunkText(string $text): array
+    {
+        // TODO - Join adjacent smaller chunks up
+        return array_filter(array_map(function (string $section): string {
+            return trim($section);
+        }, explode("\n", $text)));
+    }
+
+    protected function entityToPlainText(Entity $entity): string
+    {
+        $text = $entity->name . "\n\n" . $entity->{$entity->textField};
+        // TODO - Add tags
+        return $text;
+    }
+}

app/Search/Vectors/SearchVector.php

@@ -0,0 +1,16 @@
+<?php
+
+namespace BookStack\Search\Vectors;
+
+use Illuminate\Database\Eloquent\Model;
+
+/**
+ * @property string $entity_type
+ * @property int $entity_id
+ * @property string $text
+ * @property string $embedding
+ */
+class SearchVector extends Model
+{
+    public $timestamps = false;
+}

app/Search/Vectors/Services/OpenAiVectorQueryService.php

@@ -0,0 +1,66 @@
+<?php
+
+namespace BookStack\Search\Vectors\Services;
+
+use BookStack\Http\HttpRequestService;
+
+class OpenAiVectorQueryService implements VectorQueryService
+{
+    protected string $key;
+    protected string $endpoint;
+    protected string $embeddingModel;
+    protected string $queryModel;
+
+    public function __construct(
+        protected array $options,
+        protected HttpRequestService $http,
+    ) {
+        // TODO - Some kind of validation of options
+        $this->key = $this->options['key'] ?? '';
+        $this->endpoint = $this->options['endpoint'] ?? '';
+        $this->embeddingModel = $this->options['embedding_model'] ?? '';
+        $this->queryModel = $this->options['query_model'] ?? '';
+    }
+
+    protected function jsonRequest(string $method, string $uri, array $data): array
+    {
+        $fullUrl = rtrim($this->endpoint, '/') . '/' . ltrim($uri, '/');
+        $client = $this->http->buildClient(30);
+        $request = $this->http->jsonRequest($method, $fullUrl, $data)
+            ->withHeader('Authorization', 'Bearer ' . $this->key);
+
+        $response = $client->sendRequest($request);
+        return json_decode($response->getBody()->getContents(), true);
+    }
+
+    public function generateEmbeddings(string $text): array
+    {
+        $response = $this->jsonRequest('POST', 'v1/embeddings', [
+            'input' => $text,
+            'model' => $this->embeddingModel,
+        ]);
+
+        return $response['data'][0]['embedding'];
+    }
+
+    public function query(string $input, array $context): string
+    {
+        $formattedContext = implode("\n", $context);
+
+        $response = $this->jsonRequest('POST', 'v1/chat/completions', [
+            'model' => $this->queryModel,
+            'messages' => [
+                [
+                    'role' => 'developer',
+                    'content' => 'You are a helpful assistant providing search query responses. Be specific, factual and to-the-point in response. Don\'t try to converse or continue the conversation.'
+                ],
+                [
+                    'role' => 'user',
+                    'content' => "Provide a response to the below given QUERY using the below given CONTEXT. The CONTEXT is split into parts via lines. Ignore any nonsensical lines of CONTEXT.\nQUERY: {$input}\n\nCONTEXT: {$formattedContext}",
+                ]
+            ],
+        ]);
+
+        return $response['choices'][0]['message']['content'] ?? '';
+    }
+}

app/Search/Vectors/Services/VectorQueryService.php

@@ -0,0 +1,21 @@
+<?php
+
+namespace BookStack\Search\Vectors\Services;
+
+interface VectorQueryService
+{
+    /**
+     * Generate embedding vectors from the given chunk of text.
+     * @return float[]
+     */
+    public function generateEmbeddings(string $text): array;
+
+    /**
+     * Query the LLM service using the given user input, and
+     * relevant context text retrieved locally via a vector search.
+     * Returns the response output text from the LLM.
+     *
+     * @param string[] $context
+     */
+    public function query(string $input, array $context): string;
+}

app/Search/Vectors/StoreEntityVectorsJob.php

@@ -0,0 +1,28 @@
+<?php
+
+namespace BookStack\Search\Vectors;
+
+use BookStack\Entities\Models\Entity;
+use Illuminate\Contracts\Queue\ShouldQueue;
+use Illuminate\Foundation\Queue\Queueable;
+
+class StoreEntityVectorsJob implements ShouldQueue
+{
+    use Queueable;
+
+    /**
+     * Create a new job instance.
+     */
+    public function __construct(
+        protected Entity $entity
+    ) {
+    }
+
+    /**
+     * Execute the job.
+     */
+    public function handle(EntityVectorGenerator $generator): void
+    {
+        $generator->generateAndStore($this->entity);
+    }
+}

app/Search/Vectors/VectorQueryServiceProvider.php

@@ -0,0 +1,36 @@
+<?php
+
+namespace BookStack\Search\Vectors;
+
+use BookStack\Http\HttpRequestService;
+use BookStack\Search\Vectors\Services\OpenAiVectorQueryService;
+use BookStack\Search\Vectors\Services\VectorQueryService;
+
+class VectorQueryServiceProvider
+{
+    public function __construct(
+        protected HttpRequestService $http,
+    ) {
+    }
+
+    public function get(): VectorQueryService
+    {
+        $service = $this->getServiceName();
+
+        if ($service === 'openai') {
+            return new OpenAiVectorQueryService(config('services.openai'), $this->http);
+        }
+
+        throw new \Exception("No '{$service}' LLM service found");
+    }
+
+    protected static function getServiceName(): string
+    {
+        return strtolower(config('services.llm'));
+    }
+
+    public static function isEnabled(): bool
+    {
+        return !empty(static::getServiceName());
+    }
+}