AWS Bedrock Embedding Models Lists, Pricing – Complete Guide

AWS Bedrock embedding models convert text, images, audio, and video into numerical vectors that applications can use for semantic search, Retrieval-Augmented Generation, recommendations, clustering, and classification.

Choosing an embedding model is not simply a matter of selecting the newest option in the Amazon Bedrock model catalog.

The correct choice depends on your content type, languages, retrieval workload, vector database, required dimensions, AWS Region, latency target, and indexing budget.

For most text-based RAG applications, AWS bedrock titan embedding models V2 is a practical starting point.

Cohere Embed Multilingual is useful when users and documents span multiple languages.

Cohere Embed v4, Amazon Nova Multimodal Embeddings, and Amazon Titan Multimodal Embeddings are better candidates when visual information matters.

The best model, however, is the one that retrieves the correct evidence from your own documents.

A small evaluation set is more useful than relying only on public benchmarks.

What Are AWS Bedrock Embedding Models?

An embedding model represents the meaning of content as a vector, allowing systems to compare items by semantic similarity instead of exact keyword matches.

When text is sent to an embedding model, the model returns an array of numbers.

Content with similar meaning should produce vectors located near one another in the vector space.

For example, these queries may retrieve similar documents even though they use different words:

• “How do I reset my account password?”
• “I cannot access my login.”
• “Where can I change my credentials?”

The embedding model creates the vectors, but it does not perform the search itself.

Similarity calculations, filtering, indexing, and retrieval are handled by a vector database or search engine.

AWS also recommends splitting documents into logical sections for retrieval instead of embedding every long document as one vector.

Common use cases include:

• RAG applications
• Enterprise document search
• Product recommendations
• Duplicate-content detection
• Customer-support routing
• Intent classification
• Document clustering
• Image similarity search
• Video and audio retrieval
• Multimodal knowledge bases

AWS Bedrock Embedding Models List

AWS embedding models list now includes text-only, multilingual, image, document, audio, and video embedding options.

AWS Bedrock available embedding models are listed below.

Exact catalog available to an account can change by Region, endpoint, service tier, and model lifecycle.

AWS now directs customers to its model catalog and regional availability pages for the current source of truth.

Amazon Titan Text Embeddings V2

Amazon Titan Text Embeddings V2 is usually the strongest default for AWS-native, text-based RAG applications that prioritize cost, broad regional support, and configurable dimensions.

Titan Text Embeddings V2 accepts up to 8,192 tokens or 50,000 characters.

It returns 1,024 dimensions by default, with 512-dimensional and 256-dimensional options available.

It supports floating-point and binary vectors in supported Bedrock Knowledge Bases configurations.

Its model ID is:

amazon.titan-embed-text-v2:0

Titan V2 works well for:

• Internal document search
• Support knowledge bases
• Policy and procedure retrieval
• Product documentation
• Semantic similarity
• Content classification
• English-first RAG systems

Smaller vector dimensions reduce storage, transfer, and search costs.

They can also improve vector-search latency. However, reducing dimensions may lower retrieval quality for difficult datasets.

Do not select 256 dimensions only because it costs less.

Test 256, 512, and 1,024 dimensions against the same retrieval evaluation set.

Titan V2 supports many languages, but AWS describes it as optimized for English.

Cross-language retrieval can produce weaker results, particularly when documents and queries use different languages.

Cohere Embed Multilingual may therefore be the safer starting point for cross-language applications.

Cohere Embed English and Multilingual v3

Cohere Embed v3 is a strong choice when retrieval quality depends on explicit document and query modes or when multilingual search is a core requirement.

Cohere provides two v3 text embedding models in Bedrock:

• cohere.embed-english-v3
• cohere.embed-multilingual-v3

Both produce 1,024-dimensional vectors and support floating-point and binary vector types in Bedrock Knowledge Bases. Cohere Embed Multilingual supports more than 100 languages for multilingual and cross-language retrieval.

Cohere models use an input_type parameter.

This distinction matters in production.

Use:

• search_document when indexing document chunks
• search_query when embedding user queries
• classification for classification tasks
• clustering when grouping similar records

Using the same generic input mode for every request can reduce retrieval quality.

Your ingestion pipeline and query pipeline should explicitly send the correct input type.

The v3 models accept shorter text inputs than Titan V2.

AWS documentation lists a maximum of 512 tokens for each text.

This makes chunk-size control important.

Cohere Embed v4

Cohere Embed v4 is one of the best AWS Bedrock embedding models for mixed text-and-image documents, long-context inputs, and configurable vector formats.

Cohere Embed v4 is a unified multimodal model that can process text, images, and interleaved text-image content.

It supports configurable dimensions from 256 to 1,536 and multiple vector types, including float, int8, uint8, binary, and unsigned binary.

Its current programmatic model ID is:

cohere.embed-v4:0

It is suitable for:

• Visually rich PDFs
• Product catalogs
• Slide decks
• Scanned business documents
• Technical manuals with diagrams
• Image-to-document search
• Multimodal RAG
• Long-document retrieval

Model supports a context length of approximately 128,000 tokens per document.

That does not mean a 100,000-token document should become one embedding.

Smaller, meaningfully structured chunks usually produce better retrieval because each vector represents a narrower idea.

A long context window solves input limitations.

It does not solve chunking strategy.

Amazon Titan Multimodal Embeddings G1

Amazon Titan Multimodal Embeddings G1 is designed for placing text and images in the same semantic space.

Model can retrieve images using text, find visually similar images, or compare combinations of text and images.

Its model ID is:

amazon.titan-embed-image-v1

Titan Multimodal accepts up to 256 text tokens and images up to 25 MB with a maximum resolution of 2,048 by 2,048 pixels.

It supports output vectors of 1,024, 384, or 256 dimensions.

Good use cases include:

• E-commerce visual search
• Real-estate image discovery
• Brand-asset management
• Product recommendations
• Similar-image detection
• Image classification
• Text-to-image retrieval

This model is not a replacement for a text embedding model when the corpus consists mainly of long articles, policies, or support documents.

Its value comes from sharing a vector space across visual and textual content.

Amazon Nova Multimodal Embeddings

Amazon Nova Multimodal Embeddings is the AWS-native option for retrieval across text, images, audio, and video.

Amazon Nova Multimodal Embeddings converts multiple content types into vectors for search and retrieval.

Its model ID is:

amazon.nova-2-multimodal-embeddings-v1:0

It is relevant when one application must search across several media formats, such as:

• Text documents
• Product images
• Recorded calls
• Training videos
• Presentations
• Marketing assets
• Support recordings

AWS currently documents an output dimension of 1,024 for Nova Multimodal Embeddings in Knowledge Bases support information.

Regional access is more limited than established text models, so confirm availability in the Region where the workload and data must remain.

TwelveLabs Marengo Embed 3.0

Marengo Embed 3.0 is the specialized Bedrock option for search across video, audio, images, and related text.

Marengo can create embeddings from text, images, audio, video, or combined inputs.

Version 3.0 produces 512-dimensional vectors and supports video or audio files of up to four hours and 6 GB through supported asynchronous workflows.

This model is appropriate for:

• Searching training-video libraries
• Finding scenes within media archives
• Sports-video analysis
• Matching text queries to video clips
• Searching call or meeting recordings
• Media-content recommendations

It is unnecessary for a standard text RAG chatbot.

Use it when the media itself, not just a transcript. Its part of the retrieval problem.

Which AWS Bedrock Embedding Model Is Best?

Best model depends on your retrieval task, not the model release date or maximum context window.

Use this starting framework:

• Text-only English RAG: Amazon Titan Text Embeddings V2
• Multilingual text search: Cohere Embed Multilingual v3
• English retrieval with Cohere search modes: Cohere Embed English v3
• Text and image documents: Cohere Embed v4
• AWS-native image search: Amazon Titan Multimodal Embeddings G1
• Text, image, audio, and video retrieval: Amazon Nova Multimodal Embeddings
• Specialized video search: TwelveLabs Marengo Embed 3.0
• Existing Titan V1 index: Keep V1 temporarily and plan a measured migration

Do not evaluate models by checking whether the top result “looks reasonable.”

Create a labelled dataset of real queries and expected evidence.

Measure:

• Recall at K
• Precision at K
• Mean Reciprocal Rank
• Normalized Discounted Cumulative Gain
• Empty-result rate
• Irrelevant-result rate
• Retrieval latency
• Cost per indexed document
• Cost per query
• Final answer accuracy
• Human correction rate

1% improvement on a public benchmark may have no value if the model misses your product names, medical terminology, legal clauses, acronyms, or internal language.

AWS Bedrock Embedding Models Pricing

Embedding inference is often inexpensive, but total RAG cost also includes re-indexing, vector storage, parsing, retrieval, reranking, and generation.

AWS has published the following on-demand examples:

• Amazon Titan Text Embeddings V2: approximately $0.00002 per 1,000 input tokens
• Cohere Embed English or Multilingual v3: approximately $0.0001 per 1,000 input tokens

At those example rates:

• Embedding 10 million tokens with Titan V2 costs about $0.20
• Embedding 10 million tokens with Cohere v3 costs about $1.00

These figures cover embedding inference only.

Your full cost calculation should include:

1. Initial corpus embedding
2. Embeddings generated during updates
3. Query embedding volume
4. Vector database storage
5. Vector search requests
6. Document parsing
7. Knowledge Bases retrieval
8. Reranking
9. Generation-model tokens
10. Monitoring and data transfer

Pricing can differ by model, modality, Region, inference method, and service tier.

Verify current AWS Bedrock embedding models pricing before approving a production budget.

AWS Bedrock Models by Region

Model availability is regional, so select the deployment Region before finalizing the embedding model and vector architecture.

Amazon Titan Text Embeddings V2 has broad support across US, European, Asia-Pacific, Canadian, South American, and AWS GovCloud Regions. Cohere v3 is also available in multiple US, European, Asia-Pacific, Canadian, and South American Regions.

Newer multimodal models may have narrower in-Region availability or depend on geographic and global inference profiles.

Check availability with the AWS CLI:

aws bedrock list-foundation-models –by-output-modality EMBEDDING –region us-east-1

To inspect one model:

aws bedrock get-foundation-model –model-identifier amazon.titan-embed-text-v2:0 –region us-east-1

Choose the Region based on:

• Data-residency requirements
• Bedrock model availability
• Knowledge Bases availability
• Vector-store availability
• Application latency
• Compliance controls
• Disaster-recovery design
• Cross-Region data movement

Cross-Region inference can improve throughput, but AWS warns that data may be shared across participating Regions.

This must be reviewed before using it for regulated or residency-sensitive information.

AWS Bedrock Model IDs and ARNs

Use a model ID for most runtime calls, while some Bedrock configurations require the full foundation-model ARN.

A model ID looks like this:

amazon.titan-embed-text-v2:0

A foundation-model ARN follows this pattern:

arn:aws:bedrock:{region}::foundation-model/{model-id}

Example:

arn:aws:bedrock:us-east-1::foundation-model/amazon.titan-embed-text-v2:0

AWS-managed foundation-model ARNs do not include your account ID between the Region and resource type.

API accepts either supported model identifiers or ARN formats depending on the operation.

Do not confuse a base model ID with:

• A foundation-model ARN
• A provisioned-model ARN
• A custom-model ARN
• An inference-profile ID
• A marketplace endpoint
• A Knowledge Base ID

Inference-profile IDs may start with geographic prefixes such as us., eu., or global..

For example, Cohere Embed v4 supports geographic and global profile identifiers in addition to its in-Region model ID.

AWS Bedrock Knowledge Base Embedding Models

Knowledge Base embedding model determines vector dimensions, retrieval behavior, and how the entire corpus must be indexed.

Amazon Bedrock Knowledge Bases supports text and multimodal embedding options, including Titan, Cohere, and newer Amazon multimodal models, subject to feature and regional availability.

Before creating an index, confirm:

• Model ID or ARN
• Output dimensions
• Floating-point or binary vectors
• Vector-store dimension
• Distance metric
• Chunking strategy
• Metadata design
• Region compatibility

Vector index dimension must match the embedding output.

1,024-dimensional index cannot directly accept 512-dimensional vectors.

Changing models later is not a configuration-only update.

You normally need to:

1. Create a new index with the correct dimensions.
2. Re-embed the full corpus.
3. Validate retrieval against the old system.
4. Update application configuration.
5. Switch traffic.
6. Retain a rollback path.
7. Remove the old index after validation.

Store the model ID, model version, dimensions, chunking version, and preprocessing version with each indexing job.

This makes migrations and retrieval failures easier to investigate.

How to Implement Bedrock Embeddings Correctly

Retrieval quality depends as much on chunking, metadata, query handling, and evaluation as it does on the embedding model.

Start with a representative evaluation set before embedding the complete corpus.

A reliable implementation should include:

• Separate document and query embedding flows
• Structure-aware chunking
• Useful metadata
• Duplicate detection
• Incremental re-indexing
• Input-length validation
• Retry and throttling controls
• Retrieval-quality monitoring
• Model and index versioning
• Access-control filtering
• A migration strategy

Do not remove punctuation, numbers, product codes, medical terminology, or formatting automatically.

Aggressive text cleaning can destroy the information users actually search for.

Do not use one chunk size for every file type.

A support article, contract, table, product record, and slide deck have different structures.

Do not assume embeddings alone will solve every query.

Exact product numbers, names, dates, error codes, and identifiers often benefit from keyword or hybrid search.

A reranker can then reorder the combined results.

AWS notes that Bedrock embedding workloads are governed by request-per-minute quotas rather than token-per-minute quotas.

Throughput planning should therefore include request batching, concurrency, indexing schedules, and quota increases.

Common AWS Bedrock Supported Embedding Models Mistakes

Most poor RAG results come from weak retrieval design rather than an incapable generation model.

Common mistakes include:

• Selecting a model without a retrieval benchmark
• Embedding entire long documents as one vector
• Mixing vectors from different embedding models
• Using mismatched index dimensions
• Ignoring document-versus-query input types
• Removing important terms during preprocessing
• Using vector search for exact identifiers
• Changing models without re-indexing
• Testing only simple questions
• Ignoring multilingual query patterns
• Skipping metadata filters
• Measuring final answers without measuring retrieval
• Treating similarity scores as universal confidence scores

Similarity-score ranges vary by model, vector type, normalization, index, and distance metric.

A threshold copied from another project is not a reliable production threshold.