I need a practical recommendation for storing embeddings from text, images, and product data in one system.

A practical setup is:

Recommended approach

Use one vector database for all embeddings, with separate namespaces/collections or a shared collection with metadata for:

• text embeddings
• image embeddings
• product/data embeddings

This gives you:

• one retrieval system
• unified search across modalities
• simpler ops than running separate stores

Best-practice design

1) Store each item with:

• id
• embedding
• modality = text | image | product
• source_id or document_id
• metadata such as:
  • title
  • tags
  • category
  • timestamp
  • language
  • product attributes
  • image URL / file reference

2) Keep raw data outside the vector DB

Store the original:

• text in your document store / object storage
• images in S3/GCS/Azure Blob
• product records in SQL/NoSQL

The vector DB should hold the embedding + lookup metadata, not be your system of record.

3) Use a schema that supports filtering

Example filterable fields:

• modality
• brand
• category
• language
• price_range
• availability
• tenant_id

That lets you do:

• “search only images”
• “search only products in electronics”
• “search text and products together”

Strong practical choices

If you want the simplest production option:

• Pinecone, Weaviate, or Qdrant
  • all support metadata filtering and scalable vector search

If you want one open-source stack:

• Qdrant is a very practical default
  • easy to run
  • good metadata filtering
  • works well for multimodal embeddings

If you already use Postgres:

• pgvector can work well for smaller to medium systems
  • especially if your scale is modest
  • but it’s usually less convenient for heavy multimodal search at large scale

Important implementation detail

If your text, image, and product embeddings come from different models, that’s fine, but:

• keep them in separate collections if dimensions or distance metrics differ
• or store them in one collection only if they are comparable and you normalize the schema carefully

In practice, separate collections are often cleaner:

• text_vectors
• image_vectors
• product_vectors

Then search across them at the application layer if needed.

My recommendation

If you want a balanced, practical choice:

Use Qdrant with separate collections by modality and shared metadata fields.

• raw text/images/products stay in their own storage
• embeddings go into Qdrant
• your app queries one or more collections depending on the use case

If you want, I can give you:

1. a concrete schema example, or
2. a decision matrix comparing Qdrant vs Pinecone vs pgvector for your case.