feat: Bedrock Multimodal Embeddings 💬🖼️

[JGalego]

Adds support for multimodal embeddings (TEXT, IMAGE or both) via Amazon Titan for Multimodal Embeddings (Titan MM) and model-specific inference parameters via model_kwargs. Notice that Titan MM expects a base64-encoded image.

Example:

"""
Semantic Router powered by Amazon Titan Multimodal Embeddings
https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-titan-embed-mm.html
"""

# Standard imports
import base64

# Library imports
from datasets import load_dataset
from semantic_router.encoders import BedrockEncoder

# Load NFSW/SFW dataset and preprocess images
# https://huggingface.co/datasets/aurelio-ai/shrek-detection
data = load_dataset("aurelio-ai/shrek-detection", split="train", trust_remote_code=True)
data = data.to_list()

def process_image(example):
    """Turns images into base64-encoded strings."""
    im_bytes = example['image']['bytes']
    example['image'] = base64.b64encode(im_bytes).decode('utf8')
    return example

data = [process_image(example) for example in data]

# Initialize encoder
encoder = BedrockEncoder("amazon.titan-embed-image-v1", region="us-east-1")

# TEXT + IMAGE
embeddings = encoder(data[:5])

# TEXT only
embeddings = encoder([example['text'] for example in data[:5]])

# IMAGE only
embeddings = encoder([{'image': example['image']} for example in data[:5]])

# Change embedding dimension
embeddings = encoder(data[:5], model_kwargs={'embeddingConfig': {'outputEmbeddingLength': 256}})

[JGalego]

@jamescalam pinging you as top contributor for a review

[jamescalam]

all good - thanks for the PR @JGalego!