How to Index and Query Multiple Vectors per Document (Search Service)?

Couchbase Server
,
1

I’m trying to implement semantic search over multiple text fragments within a single document, and I’m unclear whether Couchbase’s vector search supports searching across a list of embedding vectors stored in one field.

“curious_fact_embeddings”: [ [0.123, 0.456, 0.789, …], [0.234, 0.567, 0.890, …], [0.345, 0.678, 0.901, …], [0.456, 0.789, 0.012, …]]

I want to search across all individual vectors within curious_fact_embeddings to find the document with the most semantically similar fact.

My search code:

vector_query = VectorQuery.create(
    "curious_fact_embeddings",
    query_vector,  # 768-dimensional query vector
    num_candidates=20
)
vector_search = VectorSearch.from_vector_query(vector_query)
results = collection.search(vector_search)

I tested by embedding a single string, and the results are a lot better. Is the semantic search averaging results over all embeddings in the list? Depending on the string order, the results vary.

Does Couchbase vector search support indexing and searching a field containing multiple vectors (list of lists)? If yes, how should I configure the Search index to handle this structure?
If no, what’s the recommended pattern?

Any guidance would be greatly appreciated!

3

Hi Ramiro,

Since v8.0, Couchbase supports three types of vector indexes

  1. Hyperscale Vector Index (powered by Query & GSI service)
  2. Composite Vector Index (powered by Query & GSI service)
  3. Search Vector Index (powered by Search service) – was also available since v7.6

You can read more about different kinds of indexes here

Now, coming to your question on the supported data models:
None of the 3 vector indexes currently support list of lists as shown in your example.

Workaround:
Suggested data model is one vector per field. This is supported by all 3 vector indexes above.


In your case, it will look like: 

{
"doc_id": 1
“curious_fact_embeddings”: [0.123, 0.456, 0.789, …]
}

{
"doc_id": 2
“curious_fact_embeddings”: [0.111, 0.222, 0.333, …]
}

We are also exploring other data models like list of objects where embedding is one of the field inside each object. This might come in a future release.

Hope that answers your question!

1 Like
4

Great, that’s the approach I took, denormalizing the data. It would be nice to have a vector search for a list of embeddings, reducing the need to duplicate metadata when applying hybrid filters. Thanks for the response!

5

Thanks for the feedback and glad you found the workaround.!

6

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.