Question about RAG document updates in Typesense

[ololoDevs]

Hi! I'm working with the Neuron AI RAG system and have a question about document management in the vector store.
I've created an Article Laravel model where I store RAG content in markdown format in a content field. I'm using the following approach to generate embeddings and sync to Typesense:

// Get documents from article content
$documents = StringDataLoader::for($article->content)->getDocuments();

// Send chunks → embeddings → Typesense
IngestRAG::make($typesense, $collectionName, $embeddingProvider)
    ->addDocuments($documents);

The Problem:

Every time I update an article, a new document is created in the vector store instead of updating the existing one. This leads to duplicate entries for the same article.

What I need:

When I update an article's content, I want the existing document in Typesense to be updated/replaced rather than creating a new one

Ideally, old chunks from the previous version should be removed and replaced with new chunks from the updated content

Questions:

1. Does NeuronAI\RAG\RAG have a method to update/replace existing documents instead of just adding new ones?
2. Should I manually delete old documents before adding new ones? If so, what's the recommended approach?
3. Is there a built-in way to track and manage document versions in the RAG system?

Any guidance on the best practice for handling content updates in the RAG pipeline would be greatly appreciated!

My IngestRAG class:

<?php

namespace App\Neuron\RAG;

use NeuronAI\Providers\AIProviderInterface;
use NeuronAI\Providers\HttpClientOptions;
use NeuronAI\RAG\Embeddings\EmbeddingsProviderInterface;
use NeuronAI\RAG\RAG;
use NeuronAI\RAG\VectorStore\VectorStoreInterface;
use NeuronAI\RAG\VectorStore\TypesenseVectorStore;
use NeuronAI\RAG\Embeddings\GeminiEmbeddingsProvider;
use Typesense\Client;
use NeuronAI\Providers\Gemini\Gemini;

class IngestRAG extends RAG
{
    public function __construct(
        protected Client $typesenseClient,
        protected string $collectionName,
        protected string $embeddingProvider,
    ) {}

    protected function provider(): AIProviderInterface
    {
        return match ($this->embeddingProvider) {
            'gemini' => new Gemini(
                key: config('services.gemini.key'),
                model: config('services.gemini.model'),
                parameters: [], // Add custom params (temperature, logprobs, etc)
                httpOptions: new HttpClientOptions(timeout: 30),
            ),
            default => throw new \Exception("Unknown provider: {$this->embeddingProvider}")
        };
    }

    protected function vectorStore(): \NeuronAI\RAG\VectorStore\VectorStoreInterface
    {
        return new TypesenseVectorStore(
            client: $this->typesenseClient,
            collection: $this->collectionName,
            vectorDimension: 3072
        );
    }

    protected function embeddings(): EmbeddingsProviderInterface
    {
        return match ($this->embeddingProvider) {
            'gemini' => new GeminiEmbeddingsProvider(
                key: config('services.gemini.key'),
                model: config('services.gemini.embeddings_model'),
            ),
            default => throw new \Exception("Unknown provider: {$this->embeddingProvider}")
        };
    }
}

[ilvalerione]

The RAG component automatically manages this scenario using the reindex method: https://docs.neuron-ai.dev/rag/data-loader#reindex-knowledge-source

But, there is an additional step needed in your case. Since you are creating chunks using the StringDataLoader the Documents it creates have the sourceType and sourceName fields set to manual by default. So, during reindex there is no way to differentiate the source of the documents. To be able to make "reindexBySource()" works you need to manually set these fields on documents generated out of each article.

In you case it could be:

$store = []
foreach ($articles as $article) {
    // Get documents from article content
    $documents = StringDataLoader::for($article->content)->getDocuments();
    // Add references
    foreach ($documents as $document) {
        $document->sourceType = 'database';
        $document->sourceName = $article->title;
        $store[] = $document;
    }
}

And then use reindexBySource to load documents:

// Upsert chunks → embeddings → Typesense
IngestRAG::make($typesense, $collectionName, $embeddingProvider)
    ->reindexBySource($store);

If you were importing files for example, it's automatically managed by the FileDataLoader component, since it already knows the file they come from: https://github.com/neuron-core/neuron-ai/blob/2.x/src/RAG/DataLoader/FileDataLoader.php#L111

[ilvalerione]

As described in the documentation

If sourceType and sourceName of the Documents are already present into the vector store, they will be deleted and the Documents of the new version will be saved. Other documents will be stored as usual into the vector database.

[ololoDevs]

You didn’t fully answer my main question. I would like to clarify how access to the vector store actually works.

Does every RAG agent have access to the entire vector database, as long as it points to the same Typesense collection?

Or does each RAG agent only see the documents that were uploaded through that specific agent?

From what I understand, the visibility of documents depends on the collection in the vector store, not on the agent itself.
So if two different agents use the same collection, they will both see the same documents.
And to completely separate knowledge sources, we need to create separate Typesense collections.

Could you please confirm if this understanding is correct?

[ololoDevs]

And I have one more question.

I noticed that the Document class has this method:

public function addMetadata(string $key, string|int $value): Document
{
    $this->metadata[$key] = $value;
    return $this;
}

How can this help in practice?
Can I add custom tags or labels to each document when I create it?
For example: category = serviceA, level = basic, type = faq, etc.

If I add metadata to documents, how can this metadata be used later during retrieval?
Is it included in the vector store so I can filter or prioritize documents by these tags, or is metadata only used internally?

Could you explain how metadata is intended to be used within the RAG workflow?

[ilvalerione]

You didn’t fully answer my main question. I would like to clarify how access to the vector store actually works.

Does every RAG agent have access to the entire vector database, as long as it points to the same Typesense collection?

Or does each RAG agent only see the documents that were uploaded through that specific agent?

From what I understand, the visibility of documents depends on the collection in the vector store, not on the agent itself. So if two different agents use the same collection, they will both see the same documents. And to completely separate knowledge sources, we need to create separate Typesense collections.

Could you please confirm if this understanding is correct?

Yes, it clearly depends by the vecotr store instance you use. If you give the same vector store to multiple agents they will operate on the same database.

[ilvalerione]

And I have one more question.

I noticed that the Document class has this method:

public function addMetadata(string $key, string|int $value): Document
{
    $this->metadata[$key] = $value;
    return $this;
}

How can this help in practice? Can I add custom tags or labels to each document when I create it? For example: category = serviceA, level = basic, type = faq, etc.

If I add metadata to documents, how can this metadata be used later during retrieval? Is it included in the vector store so I can filter or prioritize documents by these tags, or is metadata only used internally?

Could you explain how metadata is intended to be used within the RAG workflow?

Again in the documentation: https://docs.neuron-ai.dev/rag/data-loader#document-meta-data