AWS Certified Generative AI Developer - Professional: “Input enhancement” & intent/entity extraction

making input smarter before it ever

reaches a prompt or workflow. Think of

it as giving your system context glasses

so it understands what the user wants

and what the data actually means. Day

eight, input enhancement and intent

entity extraction.

Big idea, one sentence. Before you

generate anything, you understand the

input, what the user intends, which

entities matter, and how to normalize

the messy real world into clean machine

ready facts. Now one what input

enhancement really means plain language

users don't speak in APIs they say

things like book me with Dr. Sam next

Tuesday morning. Call me on 0412LE5.

I moved last week. New addresses. Input

enhancement means understand intent what

they want. Extract entities, dates,

names, phone numbers, locations,

normalize formats, turn chaos into

structure. Use that to route workflows

or enrich prompts. AWS expects you to do

this before the LM reasons. Howard two,

intent versus entity. The exam

distinction. Intent equals the goal.

Examples: book appointment. Cancel

booking. Update contact details. General

question. Entities. The details.

Examples. Date next Tuesday. Provider

Dr. Sam. Phone 04125513.

Address 39 Lemington RD. AWS exam

questions often hinge on this

separation. Number three, where AWS

fits. Entity and intent extraction. AWS

uses Amazon comprehend to detect

entities, names, dates, locations,

numbers, classify intent via custom

models. Important comprehend does not

generate text. It labels and classifies

text. That's why it belongs before

prompts. Number four, static plus2 input

enhancement edition. This pattern shows

up again. Static intent, taxonomy, book,

cancel, update. Entity types you care

about. Normalization rules phone date

address plus one. Raw user input. Text

transcript. OCR output. Two. Extracted

plus normalized result. Intent equals

book appointment. Entities date provider

phone. Rules stay fixed. Input changes.

Structured understanding emerges. And a

five. Real. Example. Number one. Voice

booking assistant. Very exam. Real raw

input. Open quote. Hey, I want to book

with Dr. Sam next 2 morning. My number

is 041255123.

Close quote. Step A, intent detection.

Intent, book appointment. Step B, entity

extraction. Entities detected. Provider

Dr. Sam. Date next to morning.

Phone041255123.

Step C normalization. Date 20261209

plus 6141255123.

Provider SAMC scorecore dentistry ID42.

Hatch result structured input. You can

see the code in our conversation

history. Now prompt becomes simpler.

Workflow routing is deterministic.

Errors drop sharply. Number six. Real

example number two. Routing workflows by

intent. Detected intents. Book

appointment. Booking workflow. Cancel

booking. Cancellation workflow. General

question. LLM answer. Complaint. Human

escalation. This routing happens before

prompts.

Exam signal. If the question says route

requests determine next action avoid

unnecessary LLM calls intentbased

routing. NAR 7 normalization. This is

exam gold. Phone numbers 0415555123

plus 6141255123

0412

all become 6141255123.

Dates next Tuesday tomorrow morning

21126.

all become ISO date time with time zone

addresses normalize abbreviations

standardized casing resolve local

differences local handling NA versus N

US date formats address rules ddup same

phone entered twice same address with

small variations normalization reduces

prompt length ambiguity downstream bugs

number eight why AWS wants this before

prompts because LLMs are expensive LLMs

hallucinate LLMs are bad at strict

formatting

Input enhancement makes prompts smaller,

improves accuracy, enables deterministic

systems, supports compliance. If an exam

answer sends raw input straight to the

model, it's usually wrong. One memory

story, lock it in. The receptionist

desk. Imagine a receptionist. They do

not answer medical questions. They

listen, identify what you want, write

down names, dates, phone numbers, format

everything neatly, send you to the right

department. That receptionist is intent

entity extraction. The doctor LM comes

later. Nine exam compression rules.

Memorize these. Intent equals what to

do. Entities equals details. Normalize

before reasoning. Route workflows by

intent. Reduce prompt size with

structure. If you see understand user

request, route workflow, extract

details, reduce ambiguity, comprehend

normalization before prompts. What AWS

is really testing. They're testing

whether you know that

understanding comes before reasoning.

Smart systems don't ask models to guess

what could have been extracted

deterministically.