refactor: Migrate Cohere to V2

[vblagoje]

Why:

• Migrates our Cohere integration to V2 (embedder(s) and ranker).
• Embedders can now use embedding_types to choose the size of the embeddings, providing more control over memory usage and performance.
• (Chat)Generator migration is done in a separate PR.

What:

• Introduced EmbeddingTypes enum to define and manage supported embedding types:
  • Allowed values:
    • float: Default float embeddings (valid for all models).
    • int8: Signed int8 embeddings (valid for v3 models only).
    • uint8: Unsigned int8 embeddings (valid for v3 models only).
    • binary: Signed binary embeddings (valid for v3 models only).
    • ubinary: Unsigned binary embeddings (valid for v3 models only).
• Updated CohereDocumentEmbedder and CohereTextEmbedder to handle multiple embedding types.
• Updated CohereRanker to use Cohere V2 API, added max_tokens_per_doc for token-level control, and deprecated the use of max_chunks_per_doc (now ignored).

How can it be used:

Users can specify the desired embedding type and initialize the embedders or ranker:

from haystack_integrations.components.embedders.cohere import CohereDocumentEmbedder
from haystack_integrations.components.embedders.cohere.embedding_types import EmbeddingTypes
from haystack_integrations.components.rankers.cohere import CohereRanker

embedder = CohereDocumentEmbedder(
    api_key="your_api_key",
    model="embed-english-v3.0",
    embedding_type=EmbeddingTypes.INT8,
)

ranker = CohereRanker(
    api_key="your_api_key",
    model="rerank-english-v2.0",
    max_tokens_per_doc=500,
)

How did you test it:

• Updated and added unit tests for CohereDocumentEmbedder and CohereTextEmbedder to validate functionality with multiple embedding types.
• Verified correct handling of embedding type combinations (e.g., int8, binary).
• Updated CohereRanker tests

Notes for the reviewer:

• Check the handling of embedding_types in the embedders to ensure that only valid types are accepted for v3 models.
• Verify that max_tokens_per_doc in CohereRanker works as intended and aligns with the Cohere V2 API.
• Ensure the integration supports seamless switching between embedding types and document configurations.
• To be merged in tandem with PR and then a new major version of cohere-haystack integration should be released.

[vblagoje]

@davidsbatista please review ranker changes. This is for the next week @anakin87 @davidsbatista

[anakin87]

I left some comments, but in general I agree with the direction of this PR.

Just one question: should we use the opportunity of this refactoring to move the client definition to __init__?

[davidsbatista]

LGTM, I would just suggest to replace the None by 4096 as this is the default value from Chohere API V2 (https://docs.cohere.com/v2/reference/rerank)

[davidsbatista]

LGTM from my side

[anakin87]

Looks good, but please answer this question:

should we use the opportunity of this refactoring to move the client definition to __init__?

[davidsbatista]

@anakin87 so that we can then simply do, for instance:

from haystack_integrations.amazon_bedrock import AmazonBedrockDocumentEmbedder

or any other component?

I would vote for this simplification in importing

[anakin87]

@davidsbatista no no, sorry if I did not explain myself better (we have discussed this with Vladimir in the past).

Currently, the Cohere client is recreated every time the run method is called.
In other components, we usually create the client at initialization and reuse it, so I propose to consider this refactoring as an opportunity to introduce this improvement.

[vblagoje]

Ah right, @anakin87 apologies - too many context switches. Yes, let's fix that one in a separate PR to ease the cognitive load - as IIRC we agreed privately?

[anakin87]

Yes, let's fix that one in a separate PR to ease the cognitive load - as IIRC we agreed privately?

OK