Enhancements in Amazon Bedrock Knowledge Bases
Amazon Bedrock Knowledge Bases has recently expanded its capabilities to support hybrid search for knowledge bases created with Amazon Aurora PostgreSQL and MongoDB Atlas vector stores. This enhancement, designed to elevate the relevance of search results, was previously limited to Opensearch Serverless and Opensearch Managed Clusters.
At the heart of this innovation lies retrieval augmented generation (RAG) applications, which leverage semantic search techniques based on vectors to navigate unstructured text. These vectors, crafted using foundation models, encapsulate contextual and linguistic nuances within data, enabling the system to respond to inquiries in a manner akin to human conversation.
Hybrid search represents a significant advancement by integrating both semantic and full-text search methodologies. This dual-query approach amalgamates results from both strategies, thereby enhancing the relevance of retrieved documents. It allows for the identification of documents that either align conceptually through semantic search or contain specific keywords as identified in full-text searches. The broader search scope this method provides significantly improves the quality of results, particularly for queries focused on keywords.
To enable hybrid search, users can utilize the Knowledge Base APIs or access the feature via the Bedrock console. Within the console, users have the option to select hybrid search as their preferred method or to default to semantic search alone.
Hybrid search functionality with Aurora PostgreSQL is accessible across all AWS Regions where Bedrock Knowledge Bases operates, with the exception of the Europe (Zurich) and GovCloud (US) Regions. For those utilizing MongoDB Atlas, hybrid search is available in the US West (Oregon) and US East (N. Virginia) AWS Regions.
For further details, interested parties can consult the Bedrock Knowledge Bases documentation. To embark on this enhanced search experience, visit the Amazon Bedrock console to get started.
