Quick Answer
Amazon Nova Multimodal Embeddings is generally available in AWS GovCloud (US-West) for cross-modal retrieval across text, documents, images, video, and audio through one embedding model. AWS’s August 12, 2026 announcement lists text inputs up to 8K tokens, video or audio segments up to 30 seconds, segmentation for larger files, multiple output dimensions, and synchronous and asynchronous APIs.
Availability in one GovCloud Region is not a blanket authorization to process regulated or sensitive data. Confirm the exact account partition, Region, service and model access, data classification, agency or contract requirements, network path, keys, logging, retention, human review, and every dependent service.
Workload and API Decision
| Decision | Candidate path | Acceptance boundary |
|---|---|---|
| Interactive text or image retrieval | Synchronous API | p95 latency, throttling, dimensions, relevance, and cost |
| Larger files or batch archives | Asynchronous API | job lifecycle, queue, retry, output location, encryption, and deletion |
| Video or audio longer than 30 seconds | Segment before embedding | segment policy, overlap, timestamps, context loss, and recombination |
| Unified cross-modal search | One shared embedding space | evaluate each modality pair, not only aggregate quality |
| Lower storage or compute | Smaller output dimension | measure quality loss and downstream index behavior |
Cross-Modal Evaluation
Build a labeled set for every needed direction: text-to-image, text-to-video, text-to-audio, document-to-image, and any other actual query. Include hard negatives that share names, colors, speakers, scenes, or document topics but are not relevant.
Measure retrieval quality at the top-k shown to the user, latency, failure, throughput, segment coverage, freshness, storage, index, API, transfer, and review cost. Keep model identifier, dimension, preprocessing, segment policy, and evaluation set version fixed while comparing results.
GovCloud Governance Gate
Document:
- the GovCloud account and US-West Region where inference occurs;
- data origin, classification, residency, allowed modalities, and prohibited content;
- IAM roles, separation of duties, network endpoints, encryption, keys, and rotation;
- synchronous request or asynchronous job artifacts, temporary objects, logs, backups, and deletion;
- downstream vector store, application, export, and cross-partition data flow;
- audit, incident, model change, access revocation, and rollback procedures.
Use the Nova Multimodal Embeddings GovCloud readiness tool before uploading production material.
Frequently Asked Questions
Which GovCloud Region has Nova Multimodal Embeddings?
AWS announced general availability in AWS GovCloud (US-West). Recheck current Region and model access before deployment.
Which modalities are supported?
The launch lists text, documents, images, video, and audio through one embedding model.
What are the launch input limits?
AWS lists up to 8K tokens for text and up to 30-second video or audio segments, with segmentation available for larger files. Verify the current API documentation.
Does GovCloud availability prove compliance?
No. Compliance and authorization depend on the workload, data, account, configuration, services, agreements, controls, evidence, and responsible organization.
Official Sources
- Nova Multimodal Embeddings in GovCloud announcement
- Amazon Nova user guide
- Amazon Bedrock in GovCloud
Source check: August 15, 2026. Recheck model ID, access, Region, limits, API mode, pricing, security, and compliance documentation.