what is the best way to search embeddings with metadata

The best way is usually:

1. Store embeddings in a vector database
   • Examples: Pinecone, Weaviate, Milvus, Qdrant, Elasticsearch/OpenSearch, Postgres + pgvector.
2. Store metadata alongside each vector
   • Things like type, date, author, tenant_id, category, language, etc.
3. Use metadata filtering before or during vector search
   • Example: “search only documents where tenant_id = 123 and category = legal.”
4. Then rank by vector similarity
   • Typically cosine similarity, dot product, or L2 depending on your embeddings.

Best practice

Use a system that supports hybrid retrieval:

• Metadata filters for hard constraints
• Vector similarity for semantic matching
• Optionally keyword/BM25 search for exact terms

This is better than searching embeddings alone because metadata helps narrow the search space and improves precision.

Common patterns

1. Filter first, then vector search

Good when metadata is strict:

• tenant, permissions, language, document type, date range

Example:

• “Find similar passages, but only in English and only for customer A.”

2. Vector search first, then filter/rerank

Useful when metadata is loose or the vector index is small.

3. Hybrid search

Best for most production systems:

• combine semantic score + keyword score + metadata constraints

Example query

“Find contracts similar to this clause, only from 2024, only for legal docs.”

You’d do:

• metadata filter: year = 2024 AND doc_type = contract AND department = legal
• vector search: similarity against the clause embedding

If you’re building from scratch

A good setup is:

• Postgres + pgvector for simpler systems
• Qdrant/Weaviate/Pinecone for larger-scale or more advanced filtering
• Elasticsearch/OpenSearch if you need strong text search + vector + filters

Rule of thumb

• Small/simple: Postgres + pgvector
• Medium/large with rich metadata filters: Qdrant or Weaviate
• Search-heavy enterprise: Elasticsearch/OpenSearch
• Managed scaling: Pinecone

If you want, I can also recommend the best approach for your specific use case (RAG, product search, document search, multi-tenant app, etc.).