What should I use for hybrid search over structured and unstructured data?

For hybrid search over structured + unstructured data, the best choice usually depends on how much of each you have and how tightly you need to combine them. In practice, these are the common options:

Good architecture options

1. Search engine with both keyword + vector + filters

Best when you want one system to handle:

• full-text search over documents
• semantic/vector search
• structured filters/facets on metadata

Examples:

• Elasticsearch / OpenSearch
• Azure AI Search
• Meilisearch or Typesense for simpler setups

Use this if:
you want a single retrieval layer and your structured data can be represented as filters/fields alongside text.

---

2. Vector database + SQL/OLAP database

Best when structured data is truly relational or analytical.

Pattern:

• store documents/embeddings in a vector DB like:
  • Pinecone
  • Weaviate
  • Milvus
  • pgvector
• store structured entities in Postgres / MySQL / Snowflake / BigQuery
• query both, then merge/rerank

Use this if:
your structured data has joins, transactions, or complex filtering logic.

---

3. Postgres with pgvector

Best for smaller to medium systems or when simplicity matters.

You can do:

• SQL filters on structured columns
• vector similarity with pgvector
• keyword search with Postgres full-text search

Use this if:
you want a simpler stack and don’t need massive scale.

---

What I’d recommend

If you want the most practical “hybrid search” stack:

• OpenSearch / Elasticsearch if search relevance is the top priority
• Postgres + pgvector if you want simplicity and already use Postgres
• Weaviate / Azure AI Search if you want built-in hybrid retrieval with less engineering

---

Rule of thumb

• Mostly search / documents / metadata filters → use Elasticsearch/OpenSearch
• Mostly app data / relational structured records → use Postgres + pgvector
• Need both strong semantic and structured retrieval with minimal plumbing → use Weaviate or Azure AI Search

If you tell me:

1. your data types,
2. expected scale, and
3. whether you need joins/filters/facets,

I can suggest a specific stack.