Question regarding input_type designation for cohere embeddings

Description

I’m currently investigating some discrepancies in behaviors between Weaviate’s implementation of a pure vector similiarity search using HNSW + cosine distance metrics (e.g. the defaults according to docs) and a rival database (Redis) using a 3rd party integration library (LangChain).

I just wanted to confirm Weaviate’s baked-in behavior utilizing the text2VecCohere integrations described here for managing Cohere embedding API calls.

Could someone confirm that input_type is correctly being populated as search_document when vectorizing new objects in the database and search_query when a nearText search is being performed?

I’ve dug through all the Cohere-related integration docs for Weaviate I could find as well as the client library but didn’t see any indications that I could specify input types for search/addition embedding operations. If there’s any integration docs I missed that explain this my apologies!

Server Setup Information

  • Weaviate Server Version: 1.26.1
  • Deployment Method: k8s
  • Number of Running Nodes: 3
  • Client Language and Version: Typescript 3.1.4

Any additional Information

Hello,

you can check the code here:

Vectorize (new objects) is using search_document
VectorizeQuery (nearText, right below) is using search_query