Add PhpVectorStore — zero-dependency local vector store

[MauricioPerera]

Summary

Adds PhpVectorStore as a new VectorStoreInterface implementation — the first zero-dependency local vector store option for Neuron AI.

Uses php-vector-store, a pure PHP vector database with:

• Binary Float32/Int8 storage — 392 bytes per 384d vector (4x smaller than JSON)
• Matryoshka multi-stage search — 3-5x faster via progressive 128d→256d→384d stages
• No external services — no Pinecone, no Qdrant, no Docker, no C extensions
• 100% recall on real embeddings (benchmarked with EmbeddingGemma-300m)
• Scales to ~500K vectors on a VPS with 512MB RAM

Why

All current vector store implementations require either external services (Pinecone, Qdrant, Chroma, Elasticsearch) or are ephemeral (MemoryVectorStore). PhpVectorStore fills the gap: persistent, local, zero-config.

Store	Dependencies	Persistent	Local	Storage/vec (384d)
PhpVectorStore	None	Yes	Yes	392 B
MemoryVectorStore	None	No	Yes	~3 KB (lost on restart)
FileVectorStore	None	Yes	Yes	~7 KB (JSON)
PineconeVectorStore	API key + network	Yes	No	Cloud
QdrantVectorStore	Docker/server	Yes	No	Server

Usage

class MyRAG extends RAG
{
    protected function vectorStore(): VectorStoreInterface
    {
        return new PhpVectorStore(
            directory:  __DIR__ . '/vectors',
            dimensions: 384,           // Matryoshka: truncate to 384d
            quantized:  true,          // Int8: 4x smaller
            matryoshka: true,          // Multi-stage search: 3-5x faster
        );
    }
}

Installation

composer require mauricioperera/php-vector-store

Test plan

• Verify addDocument / addDocuments store vectors to disk
• Verify similaritySearch returns Documents with correct scores
• Verify deleteBy filters by sourceType/sourceName
• Test with existing RAG examples (swap MemoryVectorStore for PhpVectorStore)
• Verify persistence: restart PHP, vectors still searchable

🤖 Generated with Claude Code

[ilvalerione]

I think this should be a dedicated composer package published on https://neuronhub.dev so we can also share it with the community using our communication channels.

Or since the Neuron integration requires no more than one class, you could include it in the default lib in a dedicated namespace, adding Neuron ^2|^3 as a dev and suggested dependency.