LlamaIndex Bedrock Embeddings Integration

Common errors

• botocore.exceptions.NoRegionError: You must specify a region.

  cause The AWS region was not specified in the `BedrockEmbedding` constructor or via environment variables (`AWS_REGION_NAME`).
  fix Ensure `region_name` is provided to `BedrockEmbedding` (e.g., `region_name="us-east-1"`) or set the `AWS_REGION_NAME` environment variable. Also check other AWS credential configurations.

• An error occurred (ValidationException) when calling the InvokeModel operation: Input is too long for requested model.

  cause The input text provided to the Bedrock embedding model (especially Cohere models) exceeded its maximum token or character limit.
  fix Reduce the size of the text chunks being embedded. For LlamaIndex, adjust `chunk_size` and `chunk_overlap` settings in your `Settings` (or `ServiceContext` for older versions) object to ensure chunks adhere to the model's limits.

• Vector dimension 1536 does not match the dimension of the index 1024

  cause The embedding model used to generate vectors (e.g., OpenAI's ADA with 1536 dimensions by default in LlamaIndex) does not match the dimensionality of the vector store index (e.g., 1024 for AWS Titan).
  fix Ensure consistency in embedding model dimensions. If you intend to use AWS Bedrock embeddings, configure LlamaIndex to use `BedrockEmbedding` for all indexing and querying operations, and re-index your data if the dimensions are mismatched. Explicitly set the embedding model in LlamaIndex's global settings or `ServiceContext`.

• Unable to install llama-index-embeddings-bedrock (due to `llama-index-core` version conflict)

  cause Specific versions of `llama-index-embeddings-bedrock` might have strict dependencies on `llama-index-core`, leading to conflicts if `llama-index-core` is already installed at an incompatible version.
  fix Upgrade `llama-index-embeddings-bedrock` to the latest version, which typically has broader `llama-index-core` compatibility. If issues persist, try upgrading `llama-index-core` to its latest version or, as a last resort, downgrade `llama-index-core` to a version compatible with your `llama-index-embeddings-bedrock` (e.g., `pip install llama-index-core==0.10.0`).

Warnings

• gotcha Cohere embedding models on Bedrock have strict input token limits (e.g., 512 tokens or ~2048 characters for `cohere.embed-english-v3`). Exceeding this limit will result in a `ValidationException` or 'Input too long' error, even with small LlamaIndex chunk sizes if the underlying text is too long.
  Fix: Ensure that text chunks passed to the embedding model are within the model's maximum context length. Adjust LlamaIndex's chunking strategy (e.g., `chunk_size` and `chunk_overlap`) and consider model-specific limitations.
• gotcha When using an `application_inference_profile_arn` with `BedrockEmbedding`, the `model_name` argument *must* still match the underlying model referenced by the profile. The integration does not validate this, and mismatched values lead to undefined behavior or errors.
  Fix: Always ensure the `model_name` passed to `BedrockEmbedding` accurately reflects the model configured in your AWS Application Inference Profile.
• gotcha Failing to configure AWS credentials (e.g., `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, `AWS_REGION_NAME`) or specify `region_name` during `BedrockEmbedding` initialization will lead to `botocore.exceptions.NoRegionError` or authentication failures.
  Fix: Set AWS credentials as environment variables, provide them directly to `BedrockEmbedding` during initialization, or configure an AWS profile, and always specify the `region_name`.
• gotcha Mixing embedding models with different output vector dimensions (e.g., default OpenAI `text-embedding-ada-002` (1536 dims) with AWS Titan (1024 dims)) when using a vector store can lead to `Vector dimension does not match the dimension of the index` errors.
  Fix: Ensure that the embedding model used for generating vectors matches the expected dimension of your vector store index. Re-index your data if you switch embedding models.

Install

• pip install llama-index-embeddings-bedrock Install latest version

Imports

• BedrockEmbedding

  from llama_index.embeddings.bedrock import BedrockEmbedding

  This is the primary class for Bedrock embedding models.

Quickstart

This quickstart demonstrates how to initialize the `BedrockEmbedding` class, configure AWS credentials and region, and generate an embedding for a given text. It highlights using environment variables for sensitive information.

import os
from llama_index.embeddings.bedrock import BedrockEmbedding

# Configure AWS credentials and region via environment variables or explicitly
# os.environ['AWS_ACCESS_KEY_ID'] = 'YOUR_ACCESS_KEY'
# os.environ['AWS_SECRET_ACCESS_KEY'] = 'YOUR_SECRET_KEY'
# os.environ['AWS_REGION_NAME'] = 'us-east-1'

# Initialize the embedding model
embed_model = BedrockEmbedding(
    model_name="cohere.embed-english-v3", # Example model, choose from supported models
    region_name=os.environ.get('AWS_REGION_NAME', 'us-east-1'),
    # Optionally, specify credentials directly or via profile_name
    aws_access_key_id=os.environ.get('AWS_ACCESS_KEY_ID'),
    aws_secret_access_key=os.environ.get('AWS_SECRET_ACCESS_KEY'),
    # profile_name='my-aws-profile'
)

# Get a single embedding
text = "Hello, world! This is a test document."
embedding = embed_model.get_text_embedding(text)
print(f"Embedding length: {len(embedding)}")
print(f"First 5 embedding values: {embedding[:5]}")

# List supported models
# supported_models = BedrockEmbedding.list_supported_models()
# print("Supported models:", supported_models)