Add PostgresHybridStore for hybrid search with RRF

Combines pgvector semantic search with PostgreSQL Full-Text Search
using Reciprocal Rank Fusion (RRF), following Supabase approach.

Features:
- Configurable semantic/keyword ratio (0.0 to 1.0)
- RRF fusion with customizable k parameter
- Multilingual FTS support (default: 'simple')
- Optional relevance filtering with defaultMaxScore
- All pgvector distance metrics supported

Author: ahmed-bhs

src/store/src/Bridge/Postgres/PostgresHybridStore.php

@@ -0,0 +1,348 @@
+<?php
+
+/*
+ * This file is part of the Symfony package.
+ *
+ * (c) Fabien Potencier <fabien@symfony.com>
+ *
+ * For the full copyright and license information, please view the LICENSE
+ * file that was distributed with this source code.
+ */
+
+namespace Symfony\AI\Store\Bridge\Postgres;
+
+use Symfony\AI\Platform\Vector\Vector;
+use Symfony\AI\Platform\Vector\VectorInterface;
+use Symfony\AI\Store\Document\Metadata;
+use Symfony\AI\Store\Document\VectorDocument;
+use Symfony\AI\Store\Exception\InvalidArgumentException;
+use Symfony\AI\Store\ManagedStoreInterface;
+use Symfony\AI\Store\StoreInterface;
+use Symfony\Component\Uid\Uuid;
+
+/**
+ * Hybrid Search Store for PostgreSQL/Supabase
+ * Combines pgvector (semantic) + PostgreSQL Full-Text Search (ts_rank_cd) using RRF.
+ *
+ * Uses Reciprocal Rank Fusion (RRF) to combine vector similarity and full-text search,
+ * following the same approach as Supabase hybrid search implementation.
+ *
+ * Requirements:
+ * - PostgreSQL with pgvector extension
+ * - A 'content' text field for full-text search
+ *
+ * @see https://supabase.com/docs/guides/ai/hybrid-search
+ *
+ * @author Ahmed EBEN HASSINE <ahmedbhs123@æmail.com>
+ */
+final readonly class PostgresHybridStore implements ManagedStoreInterface, StoreInterface
+{
+    /**
+     * @param string     $vectorFieldName  Name of the vector field
+     * @param string     $contentFieldName Name of the text field for FTS
+     * @param float      $semanticRatio    Ratio between semantic (vector) and keyword (FTS) search (0.0 to 1.0)
+     *                                     - 0.0 = 100% keyword search (FTS)
+     *                                     - 0.5 = balanced hybrid search
+     *                                     - 1.0 = 100% semantic search (vector only) - default
+     * @param Distance   $distance         Distance metric for vector similarity
+     * @param string     $language         PostgreSQL text search configuration (default: 'simple')
+     *                                     - 'simple': Works for ALL languages, no stemming (recommended for multilingual content)
+     *                                     - 'english', 'french', 'spanish', etc.: Language-specific stemming/stopwords
+     * @param int        $rrfK             RRF (Reciprocal Rank Fusion) constant for hybrid search (default: 60)
+     *                                     Higher values = more equal weighting between results
+     * @param float|null $defaultMaxScore  Default maximum distance threshold for vector search (default: null = no filter)
+     *                                     Only applies to pure vector search (semanticRatio = 1.0)
+     *                                     Prevents returning irrelevant results with high distance scores
+     *                                     Example: 0.8 means only return documents with distance < 0.8
+     */
+    public function __construct(
+        private \PDO $connection,
+        private string $tableName,
+        private string $vectorFieldName = 'embedding',
+        private string $contentFieldName = 'content',
+        private float $semanticRatio = 1.0,
+        private Distance $distance = Distance::L2,
+        private string $language = 'simple',
+        private int $rrfK = 60,
+        private ?float $defaultMaxScore = null,
+    ) {
+        if ($semanticRatio < 0.0 || $semanticRatio > 1.0) {
+            throw new InvalidArgumentException(\sprintf('The semantic ratio must be between 0.0 and 1.0, "%s" given.', $semanticRatio));
+        }
+    }
+
+    public function setup(array $options = []): void
+    {
+        // Enable pgvector extension
+        $this->connection->exec('CREATE EXTENSION IF NOT EXISTS vector');
+
+        // Create table with vector field, content field for FTS, and tsvector field
+        $this->connection->exec(
+            \sprintf(
+                'CREATE TABLE IF NOT EXISTS %s (
+                    id UUID PRIMARY KEY,
+                    metadata JSONB,
+                    %s TEXT NOT NULL,
+                    %s %s(%d) NOT NULL,
+                    content_tsv tsvector GENERATED ALWAYS AS (to_tsvector(\'%s\', %s)) STORED
+                )',
+                $this->tableName,
+                $this->contentFieldName,
+                $this->vectorFieldName,
+                $options['vector_type'] ?? 'vector',
+                $options['vector_size'] ?? 1536,
+                $this->language,
+                $this->contentFieldName,
+            ),
+        );
+
+        // Create vector index
+        $this->connection->exec(
+            \sprintf(
+                'CREATE INDEX IF NOT EXISTS %s_%s_idx ON %s USING %s (%s %s)',
+                $this->tableName,
+                $this->vectorFieldName,
+                $this->tableName,
+                $options['index_method'] ?? 'ivfflat',
+                $this->vectorFieldName,
+                $options['index_opclass'] ?? 'vector_cosine_ops',
+            ),
+        );
+
+        // Create GIN index for full-text search
+        $this->connection->exec(
+            \sprintf(
+                'CREATE INDEX IF NOT EXISTS %s_content_tsv_idx ON %s USING gin(content_tsv)',
+                $this->tableName,
+                $this->tableName,
+            ),
+        );
+    }
+
+    public function drop(): void
+    {
+        $this->connection->exec(\sprintf('DROP TABLE IF EXISTS %s', $this->tableName));
+    }
+
+    public function add(VectorDocument ...$documents): void
+    {
+        $statement = $this->connection->prepare(
+            \sprintf(
+                'INSERT INTO %1$s (id, metadata, %2$s, %3$s)
+                VALUES (:id, :metadata, :content, :vector)
+                ON CONFLICT (id) DO UPDATE SET
+                    metadata = EXCLUDED.metadata,
+                    %2$s = EXCLUDED.%2$s,
+                    %3$s = EXCLUDED.%3$s',
+                $this->tableName,
+                $this->contentFieldName,
+                $this->vectorFieldName,
+            ),
+        );
+
+        foreach ($documents as $document) {
+            $operation = [
+                'id' => $document->id->toRfc4122(),
+                'metadata' => json_encode($document->metadata->getArrayCopy(), \JSON_THROW_ON_ERROR),
+                'content' => $document->metadata->getText() ?? '',
+                'vector' => $this->toPgvector($document->vector),
+            ];
+
+            $statement->execute($operation);
+        }
+    }
+
+    /**
+     * Hybrid search combining vector similarity and full-text search.
+     *
+     * @param array{
+     *   q?: string,
+     *   semanticRatio?: float,
+     *   limit?: int,
+     *   where?: string,
+     *   params?: array,
+     *   maxScore?: float
+     * } $options
+     */
+    public function query(Vector $vector, array $options = []): array
+    {
+        $semanticRatio = $options['semanticRatio'] ?? $this->semanticRatio;
+
+        if ($semanticRatio < 0.0 || $semanticRatio > 1.0) {
+            throw new InvalidArgumentException(\sprintf('The semantic ratio must be between 0.0 and 1.0, "%s" given.', $semanticRatio));
+        }
+
+        $queryText = $options['q'] ?? '';
+        $limit = $options['limit'] ?? 5;
+
+        // Build WHERE clause
+        $where = [];
+        $params = [];
+
+        // Only add embedding param if we're doing vector search
+        if ($semanticRatio > 0.0) {
+            $params['embedding'] = $this->toPgvector($vector);
+        }
+
+        // Use maxScore from options, or defaultMaxScore if configured
+        $maxScore = $options['maxScore'] ?? $this->defaultMaxScore;
+
+        if (null !== $maxScore) {
+            $where[] = "({$this->vectorFieldName} {$this->distance->getComparisonSign()} :embedding) <= :maxScore";
+            $params['maxScore'] = $maxScore;
+            // Ensure embedding is available if maxScore is used
+            if (!isset($params['embedding'])) {
+                $params['embedding'] = $this->toPgvector($vector);
+            }
+        }
+
+        if ($options['where'] ?? false) {
+            $where[] = '('.$options['where'].')';
+        }
+
+        $whereClause = $where ? 'WHERE '.implode(' AND ', $where) : '';
+
+        // Choose query strategy based on semanticRatio and query text
+        if (1.0 === $semanticRatio || empty($queryText)) {
+            // Pure vector search
+            $sql = $this->buildVectorOnlyQuery($whereClause, $limit);
+        } elseif (0.0 === $semanticRatio) {
+            // Pure full-text search
+            $sql = $this->buildFtsOnlyQuery($whereClause, $limit);
+            $params['query'] = $queryText;
+        } else {
+            // Hybrid search with weighted combination
+            $sql = $this->buildHybridQuery($whereClause, $limit, $semanticRatio);
+            $params['query'] = $queryText;
+        }
+
+        $statement = $this->connection->prepare($sql);
+        $statement->execute([...$params, ...($options['params'] ?? [])]);
+
+        $documents = [];
+        foreach ($statement->fetchAll(\PDO::FETCH_ASSOC) as $result) {
+            $documents[] = new VectorDocument(
+                id: Uuid::fromString($result['id']),
+                vector: new Vector($this->fromPgvector($result['embedding'])),
+                metadata: new Metadata(json_decode($result['metadata'] ?? '{}', true, 512, \JSON_THROW_ON_ERROR)),
+                score: $result['score'],
+            );
+        }
+
+        return $documents;
+    }
+
+    private function buildVectorOnlyQuery(string $whereClause, int $limit): string
+    {
+        return \sprintf(<<<SQL
+            SELECT id, %s AS embedding, metadata, (%s %s :embedding) AS score
+            FROM %s
+            %s
+            ORDER BY score ASC
+            LIMIT %d
+            SQL,
+            $this->vectorFieldName,
+            $this->vectorFieldName,
+            $this->distance->getComparisonSign(),
+            $this->tableName,
+            $whereClause,
+            $limit,
+        );
+    }
+
+    private function buildFtsOnlyQuery(string $whereClause, int $limit): string
+    {
+        // Add FTS match filter to ensure only relevant documents are returned
+        $ftsFilter = \sprintf("content_tsv @@ websearch_to_tsquery('%s', :query)", $this->language);
+
+        if ($whereClause) {
+            // Combine existing WHERE clause with FTS filter
+            $whereClause = str_replace('WHERE ', "WHERE $ftsFilter AND ", $whereClause);
+        } else {
+            $whereClause = "WHERE $ftsFilter";
+        }
+
+        return \sprintf(<<<SQL
+            SELECT id, %s AS embedding, metadata,
+                   (1.0 / (1.0 + ts_rank_cd(content_tsv, websearch_to_tsquery('%s', :query)))) AS score
+            FROM %s
+            %s
+            ORDER BY score ASC
+            LIMIT %d
+            SQL,
+            $this->vectorFieldName,
+            $this->language,
+            $this->tableName,
+            $whereClause,
+            $limit,
+        );
+    }
+
+    private function buildHybridQuery(string $whereClause, int $limit, float $semanticRatio): string
+    {
+        // Add FTS filter for the fts_scores CTE
+        $ftsWhereClause = $whereClause;
+        $ftsFilter = \sprintf("content_tsv @@ websearch_to_tsquery('%s', :query)", $this->language);
+
+        if ($whereClause) {
+            $ftsWhereClause = str_replace('WHERE ', "WHERE $ftsFilter AND ", $whereClause);
+        } else {
+            $ftsWhereClause = "WHERE $ftsFilter";
+        }
+
+        // RRF (Reciprocal Rank Fusion) - Same approach as Supabase
+        // Formula: COALESCE(1.0 / (k + rank), 0.0) * weight
+        // Lower score is better (like distance)
+        return \sprintf(<<<SQL
+            WITH vector_scores AS (
+                SELECT id, %s AS embedding, metadata,
+                       ROW_NUMBER() OVER (ORDER BY %s %s :embedding) AS rank_ix
+                FROM %s
+                %s
+            ),
+            fts_scores AS (
+                SELECT id,
+                       ROW_NUMBER() OVER (ORDER BY ts_rank_cd(content_tsv, websearch_to_tsquery('%s', :query)) DESC) AS rank_ix
+                FROM %s
+                %s
+            )
+            SELECT v.id, v.embedding, v.metadata,
+                   (
+                       COALESCE(1.0 / (%d + v.rank_ix), 0.0) * %f +
+                       COALESCE(1.0 / (%d + f.rank_ix), 0.0) * %f
+                   ) AS score
+            FROM vector_scores v
+            FULL OUTER JOIN fts_scores f ON v.id = f.id
+            WHERE v.id IS NOT NULL OR f.id IS NOT NULL
+            ORDER BY score DESC
+            LIMIT %d
+            SQL,
+            $this->vectorFieldName,
+            $this->vectorFieldName,
+            $this->distance->getComparisonSign(),
+            $this->tableName,
+            $whereClause,
+            $this->language,
+            $this->tableName,
+            $ftsWhereClause,
+            $this->rrfK,
+            $semanticRatio,
+            $this->rrfK,
+            1.0 - $semanticRatio,
+            $limit,
+        );
+    }
+
+    private function toPgvector(VectorInterface $vector): string
+    {
+        return '['.implode(',', $vector->getData()).']';
+    }
+
+    /**
+     * @return float[]
+     */
+    private function fromPgvector(string $vector): array
+    {
+        return json_decode($vector, true, 512, \JSON_THROW_ON_ERROR);
+    }
+}