Query handling systems (the “smart RAG” layer)

This layer decides how to ask for

information before asking what to

retrieve.

Query handling systems. Smart rag layer.

Big idea one sentence. Instead of

sending a raw question to retrieval, you

analyze, expand, split, and orchestrate

queries so retrieval has the best

possible chance to succeed. One, why

smart rag exists at all. Users ask

questions like humans vague overloaded

ambiguous multi-art vector stores expect

focused queries clear intent scoped

retrieval. So do is about the translator

in the middle two query expansion say

more than the user did what it is you

take the user's query and expand it with

related terms or synonyms to improve

recall. Example, user asks open quote

leave policy for parents. Close quote.

Expanded internally to parental leave,

maternity leave, paternity leave, career

leave, paid parental leave. All of these

are used for retrieval even though the

user never typed them. Why AWS likes

this? Increases recall. Reduces nothing

found cases. Helps with short or vague

queries. Exam signal improve recall.

Broaden search synonyms. Query

expansion. Query decomposition split

complex questions. What it is? You break

a multi-part question into multiple

focused subqueries.

Example, user asks, open quote, what are

the tax benefits and eligibility rules

for first home buyers in NSW after 2023?

Close quote decomposed into one, tax

benefits for first home buyers in NSW

after 2023. Two, eligibility rules for

first home buyers in NSW after 2023.

Each subquery retrieves its own

evidence. Why this matters? Each sub

question is clearer. Retrieval quality

improves. Final answer is more complete.

Exam signal. Multi-art question. Complex

query. Query decomposition. Natour.

Query transformation. Change the shape,

not the meaning. What it is. You rewrite

the query into a form better suited for

retrieval. Not for humans. Examples.

Turn a question into a statement. Remove

conversational fluff. Normalized tense

and wording user open quote can you tell

me if I'm allowed to close quote

transformed open quote eligibility

criteria for close quote why this helps

vector and keyword search work better on

concise factual phrasing exam signal

rewrite for retrieval normalized queries

query transformation five orchestration

with step functions this is the AWS part

all of the above steps are not one LLM

call they're a pipeline typical flow You

can see the code in our conversation

history. This orchestration is often

done with AWS Step Functions because it

gives you branching, retries, parallel

retrieval, observability, exam signal,

multi-step query handling,

orchestration, branching, step

functions. NAS chapter 6 AWS static plus

one query handling addition static

expansion rules decomposition logic

transformation templates orchestration

workflow plus one the user query you

design the system once each query flows

through it that's static one again seven

MCP clients concept retrieval access

patterns this is subtle and very exammy

what MCP clients means conceptually

think of standardized clients that know

how to talk to retrieval systems in a

consistent way. They request context,

specify scopes, define access patterns,

avoid ad hoc retrieval logic everywhere.

Why this exists? Without this, every

agent or service implements retrieval

differently, breaks consistency, causes

security and quality issues. With MCP

style clients, retrieval is

standardized, access is controlled,

context is predictable. You don't need

protocol level details for the exam,

just the concept.

Clients that standardize how models and

agents access retrieval context

eight realistic endto-end example user

query compare parental leave and

careleaf policies for NSW employees

smart rag handling one detect complexity

needs comparison two decompose parental

leave policy NSW care relief policy NSW

three expand synonyms and policy names

four transform concise factual phrasing

five retrieve possibly in parallel Six,

aggregate. Seven, generate answer.

Without this layer, one fuzzy query,

partial retrieval, weak answer. Nine,

classic exam traps. Watch carefully.

Send raw query directly to Vector DB.

LLM magically fixes retrieval. Query

expansion replaces metadata filters.

Decomposition is only for prompts. AWS

wants systems thinking, not prompt

optimism. One memory story. Lock it in.

The research assistant expansion. What

else could this mean? Decomposition. Is

this actually multiple questions?

Transformation. How do I phrase this for

the library? Orchestration. Let me

coordinate the research. MCP client. One

official request form for all

departments. The assistant prepares the

question before going to the library.

Shack exam compression rules. Memorize

vague query expand complex query

decompose. Chatty query transform,

multi-step logic, step functions,

consistent retrieval access, MCP style

clients. If the answer treats retrieval

as a single call, it's likely wrong.

What AWS is really testing, they're

testing whether you understand that. The

quality of answers depends more on how

you ask than what you ask.

Smart rag systems shape queries before

retrieval ever happens. Only really

sticks once you see how a smart rag

layer behaves in the wild. Below are

three real production style examples

that AWS exam questions are clearly

inspired by. Real examples day 12 query

handling systems. Example one HR policy

assistant query expansion user query

raw. What's the leave policy for

parents? This is too vague for

retrieval. Smart rag behavior.

Step one, query expansion. The system

expands the query internally to include

parental leave, maternity leave,

paternity leave, car leave, paid

parental leave. The user never sees this

have retrieval. Each expanded term is

searched in the vector store, often via

open search or a managed KB result.

Instead of no results, you get all

relevant policy sections. Why AWS loves

this? Improves recall. Reduces false

negatives. No prompt trickery. Exam

takeaway short or vague queries query

expansion.

Example two legal research assistant

querition user query. What are the

eligibility rules and tax benefits for

first home buyers in NSW after 2023.

This is two questions pretending to be

one. SN smart ride behavior. Step one

detect complexity. The system detects

multiple topics conjunction decompose.

It splits into one eligibility rules for

first home buyers in NSW after 2023.

Two, tax benefits for first home buyers

in NSW after 2023.

Orchestration. These subqueries are

retrieved in parallel using AWS step

functions. Each query applies metadata

filters. NSW year 2023 uses hybrid

search may use reranking day 11.

Aggregation results are combined into a

single structured context for the LLM.

Why AWS loves this? Better coverage,

cleaner retrieval, easier to debug. Exam

takeaway multi-art question query

decomposition.

Example three. Customer support chatbot

query transformation. User query chatty.

Hey, can you tell me if I'm allowed to

cancel my subscription and maybe get a

refund?

Humans talk like this. Search engines

hate it. Smart rag behavior. Step one,

query transformation. The system

rewrites it to subscription cancellation

policy. Refund eligibility criteria. No

new meaning is added only clarity

retrieval. Short factual queries perform

better in vector search. Keyword search

hybrid systems. Why AWS loves this?

Smaller prompts, more deterministic

retrieval, less noise. Exam takeaway.

Conversational input. Query

transformation. Example four. Full smart

rag orchestration all combined. User

query compare parental leave and carele

policies for NSW employees hired after

2022. End to end flow. One analyze query

detects comparisons multiple topics.

Two, decompose parental leave policy NSW

after 2022. Car leave policy NSW after

2022. Three, expand. Add synonyms and

official policy names. Four, transform,

concise, factual phrasing. Five,

retrieve hybrid search filters, possibly

in parallel. Six, aggregate merge

evidence. Seven, answer all

orchestration is handled by step

functions, not one giant prompt. Example

five, MCP style client retrieval access

pattern. Problem without MCP style

clients, every agent queries the vector

store differently. Different filters,

different scopes, inconsistent security.

With MCP style client, conceptually you

create a standard retrieval client that

accepts query, tenant ID, topic, max

documents, always applies metadata

filters, hybrid search rules, reranking

policy. All agents use this client. CHA:

Why AWS cares? Consistency, security,

predictability, easier auditing. Exam

takeaway: MCP clients standardized

retrieval access, not ad hoc queries.

Statics one real world lockin static

expansion rules decomposition logic

transformation templates orchestration

workflow retrieval client interface plus

one user query system logic stays fixed

queries flow through it one memory story

never forget the research coordinator

expansion what else could this mean

decomposition is this more than one

question transformation how should I

phrase this for the archive step

functions let me coordinate the research

MCP client. Everyone uses the same

request form. Good answers come from

better questions, not bigger models.

Ultrashort exam cheat sheet. Vague,

expand, complex, decompose, chatty,

transform, multi-step, upstep functions,

consistent retrieval, MCP style client.