Bedrock's managed knowledge base now retrieves through an agent, and the reference CDK stack ships seven layers of telemetry with it
AspireAI community · AWS Machine Learning Blog (Chandrasekaran & Sahu)
Classic RAG embeds a question, fetches the nearest chunks from one index and generates once. That fails in three recognizable ways: a question spanning two corpora gets half an answer, a question whose right source depends on the question has no mechanism to choose, and there is nowhere for a second look after the first result. Amazon Bedrock's Managed Knowledge Base plus AgentCore turn retrieval into an agentic capability - the agent decides whether to retrieve, which knowledge base to consult and whether to go back for more before composing a grounded, cited answer.
The AgenticRetrieveStream API does the multi-turn planning and internal reranking as a service feature rather than something you assemble yourself, and the Managed Knowledge Base owns ingestion, storage, indexing and retrieval, so the OpenSearch collection and index that a 2024 version of this pattern provisioned are no longer something you create, scale or pay for. The reference CDK application routes an agent across two knowledge bases through an AgentCore Gateway and ships two CloudWatch dashboards spanning seven layers of telemetry, plus both on-demand and continuous evaluation, from the same deployment.
The limitation is scope: the post does not date the Managed Knowledge Base launch separately from its own publication, and roughly half of what the stack provisions exists to tell you what the agent did rather than to make it better. That is the point - a looping agent is not something you can reason about the way you reason about a straight-line pipeline.
https://aws.amazon.com/blogs/machine-learning/build-agentic-retrieval-on-a-managed-amazon-bedrock-knowledge-base-with-cdk
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.