Because its output is probabilistic Amazon Nova Multimodal Embeddings may result in inaccurate or inappropriate content. Customers should evaluate outputs for accuracy and appropriateness for their use case, especially if they will be directly surfaced to end users. Additionally, if Amazon Nova Multimodal Embeddings is used in customer workflows that produce consequential decisions, customers must evaluate the potential risks of their use case and implement appropriate human oversight, testing, and other use case-specific safeguards to mitigate such risks. For more information, see the AWS Responsible AI Policy.

The model supports input texts of up to 8,172 tokens, images of up to 1 image per pair, video of up to 30s, and audio of up to 30s. It accepts inputs via both sync and async APIs (with optionality for batch inference) and supports four output embedding lengths (with 3072-dimension as default), 1024, 256, 384), see Amazon Nova User Guide.

We curated data from over 200 languages for Amazon Nova Multimodal Embeddings, It accepts inputs via both sync and async APIs (with optionality for batch inference) and supports four output embedding lengths (with 3072-dimension as default, 1024, 256, 384), see Amazon Nova User Guide.

Amazon Nova Multimodal Embeddings is not trained on programming languages.

Amazon Nova Multimodal Embeddings is trained on data up until the model release date of October 28, 2025. The model will not be aware of new information beyond this date.

Amazon Nova Multimodal Embeddings converts model inputs into embedding using machine learning, specifically, using neural networks built on a transformer architecture. At a high level, the core model works by taking in the inputs and encoding them into numerical vectors. The vectors (optionally) pass through a dimension reduction layer to output the embedding depending on the users selected dimension length. Amazon Nova Multimodal Embeddings is available pursuant to the AWS Customer Agreement or other relevant agreements with AWS.

We say that an Amazon Nova Multimodal Embeddings model exhibits a particular "behavior" when it generates the same kind of output for the same kinds of prompts and configuration (for example, segmentation size and embedding length). For a given model architecture, the control levers that we have over the behaviors are primarily the unlabeled pre-training data corpus and the contrastive learning, meaning that the model can understand when information is either similar or exactly the same. Our development process exercises these control levers as follows: 1/ We pre-train Amazon Nova Multimodal Embeddings using curated data from a variety of sources, which may include licensed and proprietary data, open-source datasets, and publicly available data where appropriate. By constructing a well-balanced data mixture, the model can learn generalizable patterns; 2/ During the contrastive learning stage, we shape the embedding space so that the embeddings of semantically similar or dissimilar embeddings are correspondingly nearer or further apart.

Intrinsic and confounding variation differ between customer applications. This means that performance will also differ between applications, even if they support the same use case. Consider two applications A and B. With Application A, a media application first uses an LLM to generate detailed text captions for a video, then converts those captions into a text embedding. Application A must address multiple issues such as captioning errors, variations in text quality and level of detail and text embedding accuracy. With Application B, a media application generates native video embeddings. Application B must contend with shorter context lengths and missed events. Because A and B have differing kinds of inputs, they will likely have different embedding qualities even if each application is deployed perfectly. As a result, the overall utility of Amazon Nova Multimodal Embeddings will depend both on the model and on workflows it enables. Performance results depend on a variety of factors including Amazon Nova Multimodal Embeddings itself, the customer workflow, and the evaluation dataset, we recommend that customers test Amazon Nova Multimodal Embeddings using their own content.

We use multiple datasets and human teams to evaluate the performance of Amazon Nova Multimodal Embeddings models. No single evaluation dataset suffices to completely capture performance. This is because evaluation datasets vary based on use case, intrinsic and confounding variation, the quality of ground truth available, and other factors. Our development testing involves automated testing against publicly available and proprietary datasets, benchmarking against proxies for anticipated customer use cases, human evaluation of generations against proprietary datasets, manual red-teaming, and more. Our development process examines Amazon Nova Multimodal Embeddings's performance using all of these tests, takes steps to improve the model and the suite of evaluation datasets, and then iterates. In this Service Card, we provide examples of test results to illustrate our methodology.

Automated Evaluations provide a quantitative mechanism for evaluating the model's performance on tasks such as retrieval against candidate models in a like-for-like manner. Automated assessments can take several forms. One form is to see if the numerical representation of a text caption is sufficiently close to the numerical representation of the input item to be retrieved. One industry standard evaluation metric is net discounted cumulative gain (NDGC@k).

Safety is a shared responsibility between AWS and our customers. Our goal for safety is to mitigate key risks of concern to our enterprise customers, and to society more broadly. We align the behaviors of our foundation models with internal design policies and our commitment to responsible AI development practices. Amazon is committed to producing generative AI services that keep child safety at the forefront of development, deployment and operation, and conduct testing and implement mitigations to prevent Amazon Nova Multimodal Embeddings from accepting image inputs of inappropriate content related to children. In the case where a customer tries to retrieve an image using an embedding generated with Amazon Nova Multimodal Embeddings if no safety filters are triggered, it returns an image. In the case where the model cannot complete a prompt due to the possibility of generating an undesired output, it will not generate an embedding and return an error message.

Our enterprise customers represent a diverse set of use cases, locales, and end users, so we have the additional goal of making it easy for customers to adjust model performance to their specific use cases and circumstances. AWS offers services and tools to help customers identify and mitigate safety risks, such as Amazon Bedrock Guardrails and Amazon Bedrock Model Evaluations. Customers are responsible for end-to-end testing of their applications on datasets representative of their use cases and any additional safety mitigations, and deciding if test results meet their specific expectations of safety, fairness, and other properties, as well as overall effectiveness.

Amazon Nova Multimodal Embeddings utilizes Amazon Bedrock Abuse Detection solution, which uses hash matching or classifiers to detect potential CSAM. If Amazon Bedrock detects apparent CSAM in Amazon Nova Multimodal Embeddings user image inputs it will block the request, display an automated error message and may also file a report with the National Center for Missing and Exploited Children (NCMEC) or a relevant authority. We take CSAM commitments seriously and will continue to update our detection, blocking, and reporting mechanisms.

Chemical, Biological, Radiological, and Nuclear (CBRN): Compared to information available via internet searches, science articles, and paid experts, we see no indications that Amazon Nova Multimodal Embeddings increases access to information about chemical, biological, radiological or nuclear (CBRN) threats. We continue to assess for CBRN risk, and engage with other third-party researchers or vendors to share, learn about, and mitigate possible CBRN threats and vulnerabilities.

Despite the challenges in building training datasets that capture every real-world scenario, we aimed to combat societal bias and cultural appropriation. We tested Amazon Nova Multimodal Embeddings' ability to moderate these outcomes using a proprietary dataset of aggregated red teaming iterations that depict bias, stereotyping, and hate against individuals and groups. When users provide no guidance about the desired attributes of an object or person, it is unclear how to judge search query embeddings. For example, for the prompt "basketball players", some users might prefer a team with similar demographic attributes and other might want a distribution of attributes (for example, gender) matching some distribution they have in mind. Given this ambiguity, when there is no information included in the prompt, Amazon Nova Multimodal Embeddings is designed to return diverse results, but without specifying a desired distribution. Amazon Nova Multimodal Embeddings is designed with controls to help address this. For example, when retrieving images with the phrase “engineer” or “nurse”, the controls will return a more balance gender distribution than before. Given the intrinsic ambiguity of generating embeddings with associations to people, groups of people, or occupations without guidance, we recommend that customers consider specifying attributes in the search query. We acknowledge that while the system is designed for fair results, responsible development is a constant evolution and we welcome feedback on fairness findings that can be used to improve the overall system.

Amazon Nova Multimodal Embeddings is designed to accept text, images, document images, video, and audio as inputs. The output is limited to embeddings. Amazon Nova Multimodal Embeddings is designed to perform consistently well across a wide range of use cases. We evaluated the embedding quality on each modality and also across modalities. We measured model robustness using the following benchmarks: 1/ Text: Massive Multilingual Text Embedding Benchmark (MMTEB) (Multilingual, v1); 2/ Image: M-BEIR (MSCOCO Flikr30k); 3/ Document image: ViDoRe v2 benchmark; 4/ Video: ActivityNet, DiDeMo, MSRVTT; 5/ Audio: AudioCaps.

Customers wanting to check whether similar inputs are positioned closely in the embedding space can measure cosine similarities between the output vectors.

Amazon Nova Multimodal Embeddings is available in Amazon Bedrock. Amazon Bedrock is a managed service and does not store or review customer prompts or customer image generations, and prompts and generations are never shared between customers, or with Amazon Bedrock third party model providers. AWS does not use inputs or outputs generated through the Amazon Bedrock service to train Amazon Bedrock models, including Amazon Nova Multimodal Embeddings. For more information, see Section 50.3 of the AWS Service Terms and the AWS Data Privacy FAQs. For service-specific privacy information, see Security in the Amazon Bedrock FAQs. Amazon Nova models are designed to avoid completing prompts or generating outputs that would violate our content policies (for example, prompts seeking private information). If a user is concerned that their personal information has been included in an Amazon Nova model output, the user should contact us here.

All Amazon Bedrock models, including Amazon Nova Multimodal Embeddings, come with enterprise security that enables customers to build generative AI applications that support common data security and compliance standards, including GDPR and HIPAA. Customers can use AWS PrivateLink to establish private connectivity between customized Amazon Nova models and on-premises networks without exposing customer traffic to the internet. Customer data is always encrypted in transit and at rest, and customers can use their own keys to encrypt the data, for example, using AWS Key Management Service (AWS KMS). Customers can use AWS Identity and Access Management (IAM) to securely control access to Amazon Bedrock resources. Also, Amazon Bedrock offers comprehensive monitoring and logging capabilities that can support customer governance and audit requirements. For example, Amazon CloudWatch; can help track usage metrics that are required for audit purposes, and AWS CloudTrail can help monitor API activity and troubleshoot issues as Amazon Nova Multimodal Embeddings is integrated with other AWS systems. Customers can also choose to store the metadata, and embedding output in their vector store of choice and in their own encrypted Amazon Simple Storage Service (Amazon S3) bucket. For more information, see Amazon Bedrock Security.

AWS offers uncapped intellectual property (IP) indemnity coverage for outputs of generally available Amazon Nova Multimodal Embeddings models (see Section 50.10 of the AWS Service Terms). This means that customers are protected from third-party claims alleging IP infringement or misappropriation (including copyright claims) by the outputs generated by these Amazon Nova models. In addition, our standard IP indemnity for use of the Services protects customers from third-party claims alleging IP infringement (including copyright claims) by the Services (including Amazon Nova models) and the data used to train them.

Amazon Nova Multimodal Embeddings provides information to customers in the following locations: this Service Card, AWS documentation, AWS educational channels (for example, blogs, developer classes), and the AWS Console. We accept feedback through customer support mechanisms such as account managers. Where appropriate for their use case, customers who incorporate Amazon Nova Multimodal Embeddings in their workflow should consider disclosing their use of ML to end users and other individuals impacted by the application, and customers should give their end users the ability to provide feedback to improve workflows. In their documentation, customers can also reference this Service Card.

We have rigorous methodologies to build our AWS AI services responsibly, including a working backwards product development process that incorporates Responsible AI at the design phase, design consultations, and implementation assessments by dedicated Responsible AI science and data experts, routine testing, reviews with customers, best practice development, dissemination, and training.

The performance of any application using Amazon Nova Multimodal Embeddings depends on the design of the customer workflow, including the factors discussed below: