Description

Identify Amazon Bedrock knowledge bases configured to use a foundation embedding model in the LEGACY lifecycle state. This is a preventive retirement check of the model used to create vector embeddings, rather than an assessment of every model listed in the regional catalog or the model used to generate answers.

Rationale

Migrating embedding workloads before model retirement protects document ingestion and retrieval. A Legacy model may remain usable, but AWS recommends moving to an Active model before its end-of-life date. Lifecycle states and dates can differ by AWS Region; the configured model's regional record is the relevant source.

Impact

Changing embedding models can require a new knowledge base, compatible vector storage, and re-embedding and re-indexing the source documents. Migration can incur ingestion costs and affect retrieval quality, dimensions, data types, and application integrations. Validate the replacement before switching traffic.

Audit

This policy marks an AWS Bedrock Knowledge Base as INCOMPLIANT if its configured Foundation Embedding Model ARN resolves to a present AWS Bedrock Foundation Model whose Model Lifecycle: Status is LEGACY. It is COMPLIANT when that status is ACTIVE. The related Model ARN must exactly match the configured ARN, including its AWS Region. Legacy means deprecated and scheduled for retirement; it does not mean that the model is already unavailable or that paid extended access applies.

A knowledge base with SQL configuration type REDSHIFT, a native Amazon Kendra index ARN, or status DELETING is INAPPLICABLE, unless conflicting foundation embedding configuration is detected first. Other documented statuses (CREATING, ACTIVE, UPDATING, FAILED, DELETE_UNSUCCESSFUL, and UPDATE_UNSUCCESSFUL) are evaluated against their configured embedding model. This does not establish whether the knowledge base is serving requests. Missing or unrecognized knowledge base statuses produce UNDETERMINED after SQL and Kendra exclusions.

Missing, unresolved, or disappeared foundation model relationships remain UNDETERMINED when a native embedding ARN positively requires the target. A populated foundation relationship with a missing native embedding ARN, an ARN that does not identify a foundation model, conflicting SQL, native Kendra, or SageMaker configuration together with a foundation embedding model ARN, mismatched model ARNs or Regions, missing required model identity, and missing or unrecognized lifecycle values also produce UNDETERMINED. Required native status, related model ARN, and lifecycle values use canonical extract diagnostics.

SageMaker endpoint embedding ARNs in SageMaker Embedding Model ARN or Foundation Embedding Model ARN are INAPPLICABLE. Knowledge bases with neither a native foundation embedding ARN nor its relationship helper are also INAPPLICABLE: their variant cannot be established reliably from the exported fields. This excludes service-managed embeddings and uncollected managed/custom embedding configurations; it is not evidence of support or non-use. A native Kendra index ARN establishes Kendra configuration even when the related index is not collected; a helper Kendra lookup ID alone is insufficient. The engine automatically marks disappeared knowledge bases as DISAPPEARED.