Vector Stores

LarAI Kit supports multiple vector store backends. Switch between them via .env — no code changes.

Choosing a Vector Store

Pinecone

Works with MySQL or PostgreSQL. Vectors stored in Pinecone's cloud.

Setup

1. Create account at pinecone.io
2. Create an index: Dimensions 1536, Metric cosine
3. Add to .env:

LARAI_VECTOR_STORE=pinecone
PINECONE_API_KEY=pcsk_your-key
PINECONE_INDEX_HOST=https://your-index-abc123.svc.pinecone.io

Features

• Batch upsert — up to 100 vectors per API call via upsertMany()
• Retry with backoff — automatic retry on 429/5xx with exponential backoff and jitter
• Scoped search — scope parameters translate to Pinecone metadata $eq filters for multi-tenant isolation
• Zero dependencies — uses Laravel's built-in Http facade

Free tier

100K vectors, 1 index — enough for development and small production apps.

pgvector

Self-hosted vector search using PostgreSQL's pgvector extension.

DB_CONNECTION=pgsql
LARAI_VECTOR_STORE=pgvector

Scoped search

Scope parameters translate to WHERE source_meta->key = value JSON queries.

None (disable RAG)

LARAI_VECTOR_STORE=none

Chat still works but without document context. The NullVectorStore silently no-ops all operations.

Multi-Tenant Scoping

Both Pinecone and pgvector support scoped queries to prevent data leaks between tenants:

// Ingest with scope
$ingestion->ingestText($content, scope: ['chatbot_id' => 42]);

// Retrieve with same scope — only returns that tenant's documents
$retrieval->retrieve($query, scope: ['chatbot_id' => 42]);

// ChatService with scope
$chat->sendMessage($message, scope: ['chatbot_id' => 42]);

Without scope, queries are global. Always pass scope in multi-tenant applications.

Custom Vector Store

Implement the VectorStore contract:

use LarAIgent\AiKit\Contracts\VectorStore;
use Illuminate\Support\Collection;

class WeaviateVectorStore implements VectorStore
{
    public function upsert(int $chunkId, array $embedding, array $metadata = []): void { ... }
    public function upsertMany(array $items): void { ... }
    public function search(array $embedding, int $limit = 5, float $threshold = 0.4, array $scope = []): Collection { ... }
    public function delete(array $chunkIds): void { ... }
}

Bind in a service provider:

$this->app->singleton(VectorStore::class, WeaviateVectorStore::class);

Batch Operations

Both upsertMany() and embedMany() support batched operations:

// Embedding: batch all texts in one API call
$embedder->embedMany(['text 1', 'text 2', 'text 3']);

// Vector store: batch upsert (Pinecone does 100/request)
$store->upsertMany([
    ['chunk_id' => 1, 'embedding' => [...], 'metadata' => [...]],
    ['chunk_id' => 2, 'embedding' => [...], 'metadata' => [...]],
]);

The ingestion pipeline uses both automatically — you don't need to call these directly.