Exploring Semantic Search on Amazon DynamoDB Vector Search with Amazon Bedrock Embeddings

Disclaimer: This project is intended for learning and exploration to evaluate Amazon DynamoDB’s vector search capabilities in an AWS environment. It is a proof-of-concept designed to help developers and architects understand the integration patterns behind embedding generation, vector storage, and similarity-based retrieval. It is not a production-ready implementation or a benchmark for a specific managed vector indexing solution.

• The solution is implemented as a Spring Boot application and follows a straightforward ingestion and query workflow. During initialization, the application creates the DynamoDB table if it does not already exist, downloads a sample arXiv dataset, and ingests each paper record. For each document, the title and abstract are combined into a single text payload, passed to Amazon Bedrock Titan Embed Text V2 to generate a vector embedding, and stored in DynamoDB with the paper metadata.
• At query time, the user submits a natural-language prompt. That prompt is embedded using the same model, compared against the stored vectors using cosine similarity, and ranked to return the most relevant results. This is a useful pattern for research discovery, enterprise search, document retrieval, and AI-assisted knowledge exploration. It demonstrates how semantic relationships can be used to improve information discovery, even when the exact wording differs between the query and the source content.
• From an architectural perspective, this project highlights a few important design concerns. First, it shows how embeddings generated by Amazon Bedrock can be stored in DynamoDB alongside metadata for retrieval and ranking. Second, it demonstrates a simple similarity-search pattern using vector comparison in application code. Third, it helps teams reason about the tradeoffs between a lightweight exploratory architecture and a production-ready vector search implementation that may use a managed index or other purpose-built search services for larger scale, lower latency, and more advanced operational characteristics.
• This project is intentionally small and readable so that teams can prototype quickly and understand the mechanics of semantic search on AWS without introducing unnecessary infrastructure complexity early in the design process. It is especially useful for developers, architects, and builders evaluating whether vector search is a fit for their use case and for understanding how embedding-based retrieval can be integrated into an AWS-native application.
• By showcasing how AWS services can work together to convert unstructured content into searchable, AI-ready data, this project provides a practical starting point for exploring next-generation search experiences. It is a useful learning artifact for teams interested in semantic search, knowledge retrieval, and AI-powered data access patterns in the AWS ecosystem.