Amazon Bedrock Knowledge Base
Amazon Bedrock Knowledge Base resource replication with Arpio
Amazon Bedrock Knowledge Bases give foundation models access to your company's data through retrieval-augmented generation (RAG). Bedrock ingests documents from a data source, converts them to vector embeddings, and stores the embeddings in a vector database that it queries at inference time.
A knowledge base is meaningless without the vector store that holds its embeddings, the IAM role it runs as, and the data sources it ingests from. Arpio treats these resources as a connected graph and pulls the dependencies into recovery points automatically.
Arpio replicates the following resource types from the Amazon Bedrock Knowledge Bases service.
Knowledge Base
Arpio replicates selected Knowledge Base resources, deploying them in the recovery environment with the same embedding model, vector store configuration, and data sources.
When a knowledge base is backed by Amazon Aurora PostgreSQL (pgvector) or an Amazon OpenSearch Service domain, Arpio also replicates the underlying store together with its data, so the recovered knowledge base serves the same embeddings as the primary environment.
When a knowledge base uses an external SaaS vector store (Pinecone, Redis Enterprise Cloud, or MongoDB Atlas), the recovered knowledge base connects to the same external store, which an AWS region failure does not affect.
OpenSearch Serverless collections are recovered with their vector index structure but not the vector data; see the note below.
The following attributes are translated during replication:
|
Attribute |
Translation |
|---|---|
|
Service Role ARN |
Translated to the corresponding IAM role that Arpio manages in the recovery environment. |
|
Embedding Model ARN |
Translated to the same foundation model in the recovery region. The embedding model used by the knowledge base must be available in the recovery region. |
|
Vector Store Configuration (Amazon Aurora / RDS) |
The cluster ARN is translated to the corresponding Aurora cluster that Arpio manages in the recovery environment. Database name, table name, and field mappings are replicated without translation. |
|
Vector Store Configuration (Amazon OpenSearch Service managed cluster) |
The domain ARN and domain endpoint are translated to the corresponding OpenSearch domain that Arpio manages in the recovery environment. |
|
Vector Store Configuration (Amazon OpenSearch Serverless) |
The collection ARN is translated to the corresponding OpenSearch Serverless collection that Arpio manages in the recovery environment. |
|
Vector Store Configuration (Pinecone, Redis Enterprise Cloud, MongoDB Atlas) |
Connection endpoints and field mappings are replicated without translation. The recovered knowledge base connects to the same external vector store as the primary environment. |
|
Vector Store Credentials Secret ARN |
Translated to the corresponding Secrets Manager secret that Arpio manages in the recovery environment. |
The following resources are automatically selected into recovery points when a Knowledge Base is selected:
- The IAM service role the knowledge base runs as
- The Aurora (RDS) cluster hosting the vector store, when the knowledge base uses an Aurora vector store
- The OpenSearch Service domain or OpenSearch Serverless collection hosting the vector store, when the knowledge base uses an OpenSearch vector store
- The Secrets Manager secret holding the vector store credentials
- All Amazon S3 Data Sources belonging to the knowledge base
Limitations
- Arpio does not currently support backing up and restoring the data within OpenSearch Serverless Collections. This means the vectors from the primary environment are not directly restored in the recovery environment. Instead, Arpio triggers a sync job for each data source to handle re-populating vectors in the recovery collection as part of the restore process.
-
Knowledge bases backed by Neptune Analytics or S3 Vectors vector stores are not currently supported and will not appear for selection in the Arpio application. Kendra and structured (SQL) knowledge bases, and knowledge bases using custom or imported embedding models, are also not currently supported. Please contact Arpio support if you depend on these configurations.
Data Source
Arpio replicates Amazon S3 data sources along with their ingestion configuration.
When the source documents live in an S3 bucket in your protected AWS account, Arpio replicates the bucket and points the recovered data source at the replicated copy. When the bucket is in another account or is a public bucket, the recovered data source keeps reading from the original bucket.
The following attributes are translated during replication:
|
Attribute |
Translation |
|---|---|
|
Knowledge Base |
Translated to the corresponding knowledge base that Arpio manages in the recovery environment. |
|
S3 Bucket ARN |
Translated to the corresponding S3 bucket that Arpio manages in the recovery environment. |
The following resources are automatically selected into recovery points when a Data Source is selected:
- The Knowledge Base the data source belongs to, along with the knowledge base's own dependencies
- The Amazon S3 bucket holding the source documents, when the bucket is in the protected AWS account
Limitations
- Only Amazon S3 data sources are replicated. Data sources of other types (Confluence, Salesforce, SharePoint, web crawler, and custom) are not currently supported and will be missing from the recovered knowledge base, so their content will not be searchable there.