Skip to content

Cohere Embed v4 returns one vector for several images when input_type is image, against the documented one-vector-per-input promise #198

Description

@JGoutin

Observed while working on #126/#193, in code that shipped before this milestone.

On cohere.embed-v4, sending N images with input_type="image" returns one vector, while the same N images with input_type="search_document" return N. The gateway forwards input_type faithfully, so this may well match Cohere upstream — but docs/api_cohere_embed.md promises one vector per input without qualification, and that promise does not hold in this combination.

Two things to settle:

  • Whether Bedrock/Cohere really behave this way for that parameter pair, or whether the gateway is mis-shaping the request. Verify against the Cohere reference and a live call before changing anything.
  • If it is upstream behaviour, the page must say so — a documented limitation is acceptable here, an unqualified promise that is false is not.

Ref: docs/api_cohere_embed.md, stdapi/models/embedding/cohere_embed.py, https://docs.cohere.com/reference/embed

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    api embeddingsEmbeddings and vector store APIsbugSomething isn't working

    Type

    No type

    Projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions