Laravel Vector Search

Laravel 13 Vector Search (pgvector)

Agent Workflow (MANDATORY)

Before ANY implementation, spawn 3 agents in parallel, one Agent call each with a name:

1. fuse-ai-pilot:explore-codebase - Check current DB driver (must be PostgreSQL) and existing embedding columns
2. fuse-ai-pilot:research-expert - Verify pgvector extension version and HNSW vs IVFFlat tradeoffs
3. mcp__context7__query-docs - Pull laravel.com/docs/13.x/search + queries examples

After implementation, run fuse-ai-pilot:sniper for validation.

Overview

Critical Rules

1. Use PostgreSQL - Vector clauses ONLY work on pgsql connections - no fallback to MySQL/SQLite
2. Create an HNSW index - Without an index, queries do full table scans; > 10k rows means seconds-to-minutes latency
3. Match dimensions exactly - Insert-time and query-time embedding models MUST share the same dimensions
4. Cache embeddings - Regenerating embeddings on every request is the #1 cost driver; persist them
5. Lock the embedding model - Changing the model invalidates ALL stored embeddings; treat the model as a schema field

Architecture

database/migrations/
└── XXXX_create_documents_table.php   # Schema::ensureVectorExtensionExists(), vector(1536) col, HNSW index

app/Models/
└── Document.php                       # casts embedding to array, uses whereVectorSimilarTo

app/Ai/Services/
└── VectorSearchService.php            # encapsulates query + threshold logic

→ See Document-model.php.md for full example

Reference Guide

Templates

Quick Reference

Migration

Schema::ensureVectorExtensionExists();

Schema::create('documents', function (Blueprint $table) {
    $table->id();
    $table->text('content');
    $table->vector('embedding', 1536);
    $table->timestamps();
    $table->vectorIndex('embedding', algorithm: 'hnsw');
});

Query

$documents = Document::query()
    ->whereVectorSimilarTo('embedding', 'best wineries in Napa Valley', minSimilarity: 0.4)
    ->limit(10)
    ->get();

→ See VectorSearchService.php.md for complete example

Best Practices

DO

• Create an HNSW index BEFORE inserting bulk data - faster total ingest
• Store the embedding model name alongside the vector to detect drift
• Use minSimilarity 0.3-0.5 as a starting threshold; tune empirically
• Combine vector search with classic where() for hybrid filtering (date ranges, tenancy)

DON'T

• Don't run vector queries without an index past a few thousand rows - it becomes a full table scan
• Don't mix embedding models in the same column - distances become meaningless
• Don't generate query embeddings inside loops - batch them via Embeddings::for([...])
• Don't store embeddings as JSON strings - use the native vector column type for index support