AI Systems · Experiment: Event notes · LEAP 2026 · Short read
LEAP 2026: Managed RAG and Orchestration Trade-offs
By Obada Alsehli
An AWS AI Training Academy reflection on choosing ownership boundaries for retrieval, orchestration, and sensitive data instead of treating managed versus custom as a binary.
A familiar pipeline, a different ownership question
Days 3–4 of LEAP 2026 took me to the AWS AI Training Academy. Much of the material was a recap of ideas I had been working with in retrieval-augmented generation and agentic systems. I came to compare that mental model with how AWS packages the capabilities.
The useful question was not simply whether a managed service can perform retrieval. It was which parts of the system I want to own, and which responsibilities a service can take on.
Retrieval is not the whole application
A retrieval pipeline connects data ingestion, parsing or chunking, representations, indexing, and retrieval to the application's response. Amazon Bedrock Knowledge Bases documents managed infrastructure for these capabilities, as well as customer-managed options with different configuration responsibilities.
That does not remove the need to define the application's behavior. What evidence should support an answer? How will retrieval quality be assessed? How does the surrounding application use conversation state? What should happen when the available evidence is insufficient?
The agentic loop discussed at the event—reason, act, observe, and iterate—raised a related question about orchestration. A framework or managed service can provide mechanisms, but the application still needs explicit boundaries around actions and state.
Build or buy is a set of decisions
Managed and custom are not two fixed packages of all-or-nothing control. The relevant comparison is the responsibility boundary: storage, retrieval configuration, orchestration, observability, access, and data location.
For a sensitive workload, I would make those requirements explicit before choosing a service or building the components myself. I would also test the behavior that matters rather than treating a product description as a result.
I still lean toward owning more of the stack in sensitive domains. The Academy made that preference something to examine as a decision, rather than leave as a default. This is a reflection on the sessions, not a report of an AWS deployment or a current-service setup guide.
Cite this work
Obada Alsehli··n.d.
@article{alsehliwritingleap2026managedragorchestration,
author = {Alsehli, Obada},
title = {{LEAP 2026: Managed RAG and Orchestration Trade-offs}},
howpublished = {\url{https://obadaalsehli.com/writing/leap-2026-managed-rag-orchestration}}
}Academic attribution & reproducible reference format
BIBTEX