Series
The Exam Room
Exploring AWS, one service or situation at a time. Each track below follows one certification; once a track is complete, its cheat sheet sorts every post into the exam's scored domains.
Every post, oldest first
Exam Room · Advanced GenAI
Picking a Bedrock Model for High-Volume RAG
A million LLM requests a day, peaking at thirty per second, split across US and EU customers, with a P99 first-token target under 1.5 seconds and real reasoning over retrieved context. Bedrock has seven model families and four ways to buy capacity. Most of the landscape falls away once you name what actually decides it, and the real trick is what you do *after* you've picked the model.
Read article
Exam Room · Advanced GenAI
How to Build a Citations-Required RAG Over 50K Internal Documents
Fifty thousand internal documents, five gigabytes of text, weekly churn, a three-second latency budget, per-user access control, and a citation in every single answer. The RAG landscape on Bedrock is bigger than one product and the interesting part of the design is what falls away once you name the five things that actually decide it.
Read article
Exam Room · Advanced GenAI
Combining RAG and Fine-Tuning for a Legal Contract Assistant
A legal-tech team wants a contract review assistant that understands two hundred thousand past matters, speaks in the firm's voice with clause-by-section citations, and refuses anything off-domain. A hundred thousand dollars, three months. RAG, fine-tuning, and continued pre-training each solve a different half of that sentence; the interesting answer is which two to pick, not which one.
Read article
Exam Room · Advanced GenAI
Designing Short-Term and Long-Term Memory for a Bedrock Chat Assistant
A customer-support assistant where the average conversation runs fifteen turns before it resolves, and returning users pick up two weeks later expecting the bot to remember they've been waiting on a refund. Two memory problems in one product (what's live in the current conversation and what persists across visits) and four plausible ways to build it. Bedrock Agents' built-in memory handles one half cleanly; the other half is where teams reach for DynamoDB or a knowledge base and get it wrong.
Read article
Exam Room · Advanced GenAI
Configuring Bedrock Guardrails for PII, Topics, and Grounding
A consumer-facing chatbot on Bedrock has passed every red-team round on the obvious harms (no weapons, no hate, no CSAM) and is still shipping embarrassments: a card number pasted by one user echoing back in a reply, the bot cheerfully comparing the company's product with a named competitor, and a hallucinated policy line that nobody in the building wrote. Five different filter jobs wrap the same Bedrock invocation, and Guardrails is the one surface that does all five without five Lambdas.
Read article
Exam Room · Advanced GenAI
Spreading Bedrock Load with Cross-Region Inference Profiles
A Bedrock-backed SaaS serving US, EU and APAC customers is hitting regional quota in us-east-1 during peak while the same model sits idle in eu-west-1. The team wants to spread load without fracturing the product into three regional deployments. Three letters on the front of the model ID do the job, provided the model supports it and the geography fits the customer.
Read article
Exam Room · Advanced GenAI
Building RAG When the Source Documents Change Daily
A support assistant that has to answer from a product manual which the product team edits weekly, a pricing sheet that changes at month-end, and an operational runbook that mutates hourly. The base model doesn't know any of it, and fine-tuning won't keep up. Retrieval is the answer; the question is how much of the retrieval plumbing we want to own, and Bedrock Knowledge Bases, a LangChain stack, and a hand-rolled pipeline each put the lines in different places.
Read article
Exam Room · Advanced GenAI
How to Wire an LLM to Side-Effecting Actions with Bedrock Agents
An assistant that has to look up a customer's subscription, pause it, refund a charge, and email confirmation. Not just answer, act. The glue between a language model and the rest of our systems is a solved problem three different ways: Bedrock Agents, a LangChain agent loop, or a hand-written tool router. Each of them handles tool definition, invocation, and error recovery, but they put the guardrails in very different places.
Read article
Exam Room · Advanced GenAI
Picking a Vector Store for Bedrock RAG
Twelve million embedding vectors, a 50ms retrieval budget, hybrid queries that mix keyword and semantic, and a bill that should not double the Bedrock spend on its own. OpenSearch Serverless, Aurora with pgvector, and Pinecone Serverless all serve the same shape of query, but their pricing curves, operational shapes, and query surfaces diverge the moment the corpus grows beyond demo scale.
Read article
Exam Room · Advanced GenAI
How to Manage Prompts Across Thirty Services on Bedrock
One prompt scattered across thirty services, no versioning, no tests, drift between the copy in the code and the copy in the docs, a silent regression when somebody changed 'concise' to 'brief' and retention on one response tanked. Prompt engineering at a hundred callers isn't prose discipline, it's configuration management. Bedrock Prompt Management, Git-backed templates, and parameterised prompts each solve a slice of the same problem.
Read article
Exam Room · Advanced GenAI
Evaluating LLM Output With Bedrock Eval Jobs
Two thousand historical support tickets, a summarisation prompt, a new model candidate, and a product manager asking whether switching would hurt quality. Bedrock evaluation jobs offer automated scoring, human review through Ground Truth workflows, and model comparison side by side, but they answer different questions, and getting the correct number out of the correct job matters more than running more of them.
Read article
Exam Room · Advanced GenAI
How to Build a Multi-Modal Bedrock Assistant for Insurance Claims
A claims-processing assistant that reads a scanned invoice, listens to a voicemail, answers the customer's question in plain text, and, if asked, reads it back. Four modalities, one conversation. The model choice, the orchestration shape, and the ways different inputs fail each push the architecture in different directions, and the naive 'just use a multi-modal model' misses half of where the real work is.
Read article
Exam Room · Advanced GenAI
How to Cut a Bedrock Bill Without Hurting Quality
A Bedrock bill that doubled in two months, a product roadmap that blames the retrieval service, a finance partner who'd like a straight answer. The cheapest token is the one you don't send; the next cheapest is the one you send to the correct model. Model routing, prompt compression, cached retrieval, and provisioned throughput each solve a different slice of the cost problem, and none of them is the silver bullet.
Read article
Exam Room · Advanced GenAI
Importing Custom Weights into Bedrock
A research team has fine-tuned an open-weights model for medical-notes summarisation on a private SageMaker cluster. The resulting weights live in S3; the production runtime wants Bedrock's ergonomics. Custom model import bridges that gap, but it only works for certain base architectures, comes with throughput minimums, and quietly changes the cost model compared to on-demand foundation models.
Read article
Exam Room · Advanced GenAI
Streaming Responses to Cut First-Token Latency
A chat interface where users wait four seconds for any response on long generations, abandon rate creeping up, product asking why we can't do the typing-animation thing that every other assistant does. Streaming isn't just a UX polish, it changes how the entire response path has to work, from the SDK call through API Gateway to the browser, and each hop has its own way of getting it wrong.
Read article
Exam Room · Advanced GenAI
When a Document Won't Fit the Context Window
A 400-page contract, a 200-page policy manual, and a legal team asking 'what clauses govern refund disputes across both?' A 200,000-token context window sounds like enough until you realise what goes in with the documents. Chunking, map-reduce, hierarchical summarisation, and sliding context windows each answer a different question, and getting the boundaries correct is most of the battle.
Read article
Exam Room · Advanced GenAI
Picking an Embedding Model for Retrieval
An index with 20 million chunks, queries that need to work in English, Spanish, Portuguese, and Japanese, and a budget that won't bear re-embedding every six months when someone decides the new model is better. The embedding model quietly caps what a retrieval system can ever do. Titan, Cohere, and a self-hosted model all trade different things, and the comparison is messier than the marketing suggests.
Read article
Exam Room · Advanced GenAI
Making a Bedrock App Audit-Ready
A support assistant built on Bedrock has run cleanly for months, and now a compliance review has landed on it. The auditors want to know who invoked the model, what prompts and completions flowed through it, which model version answered, who approved the deployment, and proof that no customer PII leaked into the logs. The app works. It just can't answer a single one of those questions yet.
Read article
Exam Room · Advanced GenAI
Checking a Bedrock Feature for Bias and Explainability
A Bedrock-powered feature drafts summaries of job applicants and triages people-facing cases, so its outputs shape decisions about real people, and it is up for a responsible-AI review before launch. The team already knows how to stop the model saying toxic or off-topic things; now they are being asked to show the feature is fair across groups and that any given output can be explained and justified. Content safety was the easy part; 'is it fair, and can you explain it?' is a different set of jobs.
Read article
Exam Room · Advanced GenAI
How to Match Bedrock Pricing to Workload Rhythm
Standard on-demand Bedrock is priced per token, throttled per minute, and latency-variable in a way that product hates. The Reserved tier buys guaranteed tokens-per-minute at a fixed monthly price, Priority buys faster responses per request, Flex discounts patient traffic, and the Batch API halves the rate for anything that can wait. The right pick changes with the shape of the workload: a 24/7 assistant and a burst-heavy report generator land on opposite sides of the same line.
Read article
Exam Room · Advanced GenAI
Keeping PII Out of LLM Prompts and Logs
A claims assistant that has to answer questions about a customer's claim while keeping the customer's name, address, policy number, and medical details out of training data, out of logs, and out of anything a subpoena could touch later. PII redaction isn't one knob, it's four or five, in different places, each covering a different leak path. Comprehend, Bedrock Guardrails, custom redaction, and the prompt itself each handle a different shape of the problem.
Read article
Exam Room · Advanced GenAI
Evaluating a RAG Pipeline End to End
A citations-required assistant over internal documents answers most questions well and a few of them wrong, and the team cannot tell why. Was the right passage never retrieved, or was it retrieved and then ignored? Those are two different failures with two different fixes, and an evaluation that reports one score for the whole pipeline hides which one you have. This walks through separating the retrieval half from the generation half so the regression harness points at the part that actually broke.
Read article
Exam Room · Advanced GenAI
Hybrid Search and Reranking for Bedrock RAG
A support assistant on Bedrock retrieves by pure semantic search and keeps whiffing on exact-term queries: product codes, error codes, part numbers, acronyms, proper names. The embedding of 'ERR-4021' sits close to a hundred other error codes, so a plausible-but-wrong chunk outranks the one document that actually names it. The fix is two levers, hybrid retrieval and a reranker, and knowing which one the precision problem actually needs.
Read article
Exam Room · Advanced GenAI
Prompt Engineering Techniques That Move the Needle
A dozen prompt-engineering tricks all promise better output, but half of them cost tokens without moving quality and a couple quietly widen a security hole. Zero-shot, few-shot, chain-of-thought, ReAct, schema prompting, system framing, and delimiters each earn their keep on a specific task shape and hurt on the wrong one. The job is matching the technique to the task rather than stacking all of them and hoping.
Read article
Exam Room · Advanced GenAI
Choosing a Model From the Bedrock Catalogue
The Bedrock catalogue lists models from a dozen providers across text, images, video, and embeddings, and the reflex is to reach for the biggest, cleverest one for everything. That reflex triples the bill and slows every call, because most workloads never needed the flagship. The job is matching the model class to the shape of the work, then picking the smallest model that clears the quality bar.
Read article
Exam Room · Advanced GenAI
Choosing an Inference Option for a GenAI Workload
A summarisation feature, a nightly re-scoring job, and a customer-facing chat surface all need to serve a model, but they have nothing else in common. One wants low latency on steady traffic, one runs offline overnight on millions of records, and one spikes at odd hours and sits idle the rest of the time. Bedrock and SageMaker hosting each offer several ways to serve inference, and the right one falls out of the traffic shape and the cost model more than the model itself.
Read article
Exam Room · Advanced GenAI
Fine-Tuning, Continued Pre-Training, or Distillation
Prompting stopped closing the gap and retrieval only goes so far, so the model itself has to change. But three training routes wear the same word: fine-tuning on labelled pairs, continued pre-training on raw domain text, and distillation into a smaller student. They need different data, chase different goals, and cost wildly different amounts to run and to serve.
Read article
Exam Room · Advanced GenAI
Managing Prompts With Bedrock Prompt Management
The prompt is the product, and yet it usually lives as a string literal glued together in application code, edited by whoever last touched the deploy, with no version and no way to roll back a bad wording change. Amazon Bedrock Prompt Management makes the prompt a first-class resource with input variables, an attached model and inference configuration, and immutable versions an application references by identifier. The decision is whether that managed store earns its place over a prompt kept in your own source control, and it turns on how many services share the wording, who needs to edit it, and whether prompts feed Bedrock Flows.
Read article
Exam Room · Advanced GenAI
Prompt Caching Versus Response Caching on Bedrock
Two things called caching sit in front of a Bedrock app and they are not the same lever. Prompt caching still calls the model but pays a fraction for a shared prefix; response caching may skip the model entirely and risks serving a stale or wrong-match answer. Knowing which repeated part you are actually paying for, the input prefix or the whole request, decides which one you reach for, and whether you stack both.
Read article
Exam Room · Advanced GenAI
Getting Documents Into a Bedrock Knowledge Base
A Bedrock Knowledge Base is only as good as what gets into it, and the ingestion pipeline is where most retrieval quality is won or lost. Documents pass through a source connector, a parser, a chunker, and an embedding model before they ever reach the vector index, and each stage has a choice with real cost and quality consequences. The job is picking the connector, the parser, and the sync mode that fit the documents you actually have.
Read article
Exam Room · Advanced GenAI
Choosing a Chunking Strategy for Bedrock Knowledge Bases
A mixed corpus is going into an Amazon Bedrock Knowledge Base: hundred-page PDFs, structured policy documents with numbered sections and tables, and a pile of short FAQ entries. The default fixed-size chunking retrieves badly because answers land across a chunk boundary or the chunks are so large they dilute relevance, and the team needs a chunking strategy that fits the shape of the documents rather than fighting it.
Read article
Exam Room · Advanced GenAI
Which AWS Store Can Do Vector Search
A team already runs OpenSearch for logs, Aurora for the transactional data, DynamoDB for the session store, and a Mongo-compatible cluster for the catalogue. Now a knowledge base needs vector search, and the first question is not which vector database to buy. It is which of the stores already humming in the account can do the job once the right feature is switched on, which one you enable differently than you expect, and which one looks eligible but is a trap.
Read article
Exam Room · Advanced GenAI
Picking the Right Tool to Check and Govern GenAI Data
Every night a knowledge base pulls in support tickets, product docs, and a database export, and every night some fraction of that arrives malformed, half-empty, duplicated, or carrying customer PII into a place it should never reach. The model is only as trustworthy as the corpus behind it, so the pipeline needs a gate: something that ingests the raw data, checks it against rules, and governs who can see what before any of it reaches an embedding. AWS offers at least five tools that each claim a piece of that job, and they are easy to mix up.
Read article
Exam Room · Advanced GenAI
Retrieval Over Structured Data With Text-to-SQL
A finance team wants natural-language answers to questions like total revenue by region last quarter, and the obvious move is to embed every row and reach for vector search. It falls apart the moment the answer is a sum instead of a sentence. The right pattern turns the question into SQL, runs it against a known schema, and returns the computed result. This walks the axes that decide when retrieval belongs in a query engine, not a vector index, and how to run generated SQL safely.
Read article
Exam Room · Advanced GenAI
Defending a Bedrock App Against Prompt Injection
A support assistant on Bedrock reads user questions, pulls answers from a knowledge base, and can raise refunds through a tool. Every one of those surfaces is a way in for an attacker who wants to override the instructions, jailbreak the guardrails, or turn the refund tool against the business. No single control stops all of it, so the real question is how the layers stack.
Read article
Exam Room · Advanced GenAI
Securing a Bedrock App: IAM, PrivateLink, and Keys
A production assistant on Amazon Bedrock has four separate ways to leak, and each one is a different control plane. Who is allowed to call the model, how the traffic reaches Bedrock, who holds the encryption keys, and where the data lives are four questions, not one. Getting one right does nothing for the other three; they are layers you stack, not options you pick between.
Read article
Exam Room · Advanced GenAI
Tuning Fine-Tuning: Epochs, Learning Rate, and Batch Size
A fine-tuning job on Amazon Bedrock exposes a short list of hyperparameters, and the two failure modes sit at opposite ends of the same dial: too little training and the model never picks up the behaviour you paid for, too much and it memorises the examples and forgets how to do everything else. Epochs, learning rate, and batch size are the knobs, and the training and validation loss curves are how you read whether you have turned them too far. The job is matching the amount of training to the amount of data, then trusting a held-out evaluation over the loss number.
Read article
Exam Room · Advanced GenAI
Orchestrating Multiple Bedrock Agents
A support assistant that once answered one kind of question now has to look up billing, check delivery windows, raise a ticket, and draft a reply, and each of those wants different tools and different judgement. One agent can reason its way through all of it, a supervisor agent can farm the pieces out to specialists, and a Flow can run the whole thing as a fixed pipeline. The trade is flexibility against predictability, cost, and latency, and the right shape depends on whether the model or the designer decides what happens next.
Read article
Exam Room · Advanced GenAI
Running Agents in Production With Bedrock AgentCore
A prototype agent built on an open-source framework works on a laptop, then has to become something real subscribers hit all day, with sessions that must not bleed into each other, memory that survives a reconnect, credentials that let it touch internal systems without over-granting, and traces you can read when a run goes wrong. Bedrock AgentCore is the set of building blocks for that operational layer, framework-agnostic and model-agnostic, so the question is which pieces you actually need and where the managed line should sit. The trade is control against operational burden, and it turns on how much of the runtime, memory, identity, and observability you want to own yourself.
Read article
Exam Room · Advanced GenAI
Buy or Build: Amazon Quick Versus a Custom RAG App
An enterprise wants a chat assistant over its own documents, and the choice is Amazon Quick off the shelf or a retrieval app built on Bedrock Knowledge Bases. Quick connects to the data sources, does the retrieval, and respects each user's permissions out of the box; a custom build hands you every knob and every operational bill. The deciding question is rarely which is more powerful, it is how bespoke the requirement is and whether per-user document access control has to be enforced.
Read article
Exam Room · Advanced GenAI
When a Purpose-Built AI Service Beats a Foundation Model
Reaching for a foundation model on every AI task is a reflex that gets expensive fast, because a lot of what teams build is a well-defined single-purpose job that a managed AWS service already does cheaper, faster, and deterministically. Textract, Comprehend, Transcribe, Translate, Rekognition, Polly, Lex, Personalize, and Kendra each own a task shape, and the win is knowing when one of them beats a prompt. The interesting cases are the pipelines where a purpose-built service feeds a foundation model rather than competing with it.
Read article
Exam Room · Advanced GenAI
Proving Where AI Content Came From
A generative feature raises four different questions at once: can you prove an output came from your model, can a reader see what the service is and is not meant to do, can you stop bad behaviour as it happens, and can you show how the model was built and evaluated. Each one has a different AWS control, and reaching for the wrong one leaves the real question unanswered. The job is matching the responsibility to the artefact that actually discharges it.
Read article
Exam Room · Advanced GenAI
Lab: Invoke a Foundation Model From Lambda
A hands-on lab, the first in the series. You are handed a Lambda function with an execution role, a model id in an environment variable, and one gap: the Bedrock call itself. Five lines of Python, deployed with CloudFormation, proven with a script, torn down when you are done. Every later lab builds on this function, so this is where the track starts.
Read article
Exam Room · Advanced GenAI
Where Humans Belong in a GenAI Pipeline
A generative-AI system needs human judgement in three different places, and it is easy to reach for the wrong service because all three sound like putting a person in the loop. Labelling training data, reviewing live predictions, and judging model quality are separate jobs with separate tools on AWS. Getting the placement right keeps people on the decisions only people can make and keeps them off the ones a threshold or a metric already handles.
Read article
Exam Room · Advanced GenAI
Searching Images and Text With Multimodal Embeddings
Searching a product catalogue by typing a description, or by dropping in a lookalike photo, sounds like a job for a vision model, but a chat model that reads an image reasons about it rather than retrieving against a library. The tool that powers cross-modal search is a multimodal embedding model, which places images and text in one shared vector space so nearest-neighbour finds matches whichever modality did the asking. Getting it right means separating retrieval from reasoning, and keeping the distance metric matched to the model.
Read article
Exam Room · Advanced GenAI
Tuning How a Model Samples: Temperature, Top-P, and Top-K
Temperature, top-p, and top-k all change how a model picks its next token, and reaching for the wrong one turns a crisp classifier into a coin flip or a brainstorm into a monotone. Each parameter shapes the distribution in a different place, they interact in ways that surprise people who turn all three dials at once, and the right setting is a property of the task rather than a house style. The job is knowing which knob does what and leaving the rest alone.
Read article
Exam Room · Advanced GenAI
Budgeting Tokens for a Long-Document Workload
A workload that reads long documents keeps hitting the context window and the bill keeps climbing, and the instinct is to reach for the model with the biggest window and stuff whole files into it. That instinct is expensive, slower, and can quietly lower answer quality as the relevant facts get buried. The real job is reasoning about tokens: what fits, what it costs on the way in and on the way out, and how to shrink the input without losing the part that matters.
Read article
Exam Room · Advanced GenAI
SageMaker JumpStart or Bedrock for the Same Model
A team that wants Llama 3.3 70B in production can take it through SageMaker JumpStart, pushing a button to deploy onto an endpoint they own, or through Bedrock's model catalog with per-token pricing and no infrastructure. Same model, sometimes the same base weights, two quite different operational shapes. The correct choice depends on how much of the serving layer you actually want to touch.
Read article
Exam Room · Advanced GenAI
Making an LLM Output Reproducible
Two people run the same prompt through the same model and get two different answers, and now a compliance reviewer wants to know which one the system actually produced. Temperature zero helps, pinning the model version helps more, and only a cache in front of the model gives you byte-identical repeats. The real question is how much determinism the use case genuinely needs before you pay for it.
Read article
Exam Room · Advanced GenAI
Why Your RAG Returns the Wrong Chunk
A retrieval system answers with confidence and cites the wrong passage, and the reasons are frustratingly varied: a distance metric that disagrees with the embedding model, chunks too big or too small, a query phrased nothing like the documents, an SKU that semantic search cannot match, a missing tenant filter, no reranking, or an index that never re-synced. This walks each symptom back to its cause and lands it on a fix.
Read article
Exam Room · Advanced GenAI
Metadata Filtering for Multi-Tenant Retrieval
One retrieval index serves every tenant, so the question that decides everything is whether a query can ever surface a document its asker was never meant to see. Metadata filters keyed on verified identity are the boundary; a prompt that politely asks the model to stay in its lane is not. This walks through attaching tenant and access metadata at ingestion, filtering at query time, and why pre-filtering beats filtering the results afterwards.
Read article
Exam Room · Advanced GenAI
Parent-Document Retrieval: Small Chunks, Big Context
The chunk size that retrieves precisely and the chunk size that gives the model enough context pull in opposite directions: small chunks embed cleanly and match the query but hand the model a fragment, while large chunks carry the surrounding context but blur the embedding and match worse. The fix is to stop treating them as one number and decouple the unit you search from the unit you return, which is what parent-document retrieval, hierarchical chunking, and sentence-window retrieval each do.
Read article
Exam Room · Advanced GenAI
Lab: Put a Guardrail in Front of a Bedrock Model
A hands-on lab. You are handed a working Bedrock model call and a fully-configured guardrail, and your job is the two lines that connect them. Deploy real infrastructure with CloudFormation, block a financial-advice question, watch a phone number get redacted, then tear it all down. The reading tells you what a guardrail is; this makes you wire one in.
Read article
Exam Room · Advanced GenAI
Agentic RAG: When Retrieval Needs to Reason
A support assistant that answered one kind of question from one knowledge base now gets questions that need two lookups, a reformulated search, and a second retrieval once the first one comes back thin. Plain RAG runs a fixed pipeline: embed the query, fetch the top matches once, generate. Agentic RAG hands the retrieval decisions to the model, which can choose whether to search at all, which source to hit, how to rewrite the query, and whether to go back for more. The first is cheap and predictable; the second is the answer when one pass genuinely cannot get there, and a needless tax when it can.
Read article
Exam Room · Advanced GenAI
Designing Safe Tool Schemas for an Agent Action Group
An agent action group is the set of tools you hand a foundation model and let it call on its own judgement, and every tool is a capability you are granting. The arguments arrive already shaped by a model that can be steered by whatever text reached the prompt, so a tool that runs a free-form command lets one coaxed call reach far further than you meant. Design each tool narrow, strongly typed, and validated where it runs, and the worst a manipulated call can do stays small. The craft is scoping every tool to the least it needs.
Read article
Exam Room · Advanced GenAI
Surviving a Model Deprecation on Bedrock
A foundation model you pinned two quarters ago will not live forever; Bedrock versions its models, marks the old ones legacy, and eventually retires them. Pin a version and behaviour stays stable, but the migration bill comes due on a schedule you do not control. Chase the newest version and you trade that bill for silent behaviour drift. The way through is a rehearsed migration: pin, watch the notices, test the successor against a saved eval set, and cut over behind a flag with a rollback ready.
Read article
Exam Room · Advanced GenAI
Measuring Hallucination in a RAG System
A retrieval-augmented assistant that answers confidently and wrong is failing in one of two very different places, and the fix depends on which. A hallucination can come from retrieval, where the context never held the answer, or from generation, where the model ran past what the context said, and a single quality number hides both. Measuring it well means separating faithfulness to the retrieved passages from whether the answer is even the right one, then attributing each miss to the stage that caused it.
Read article
Exam Room · Advanced GenAI
Choosing an Embedding Model for a Multilingual Corpus
A support knowledge base written in English, French, and Japanese needs semantic search, and the obvious move of running an English embedding model over everything quietly fails: a French question never finds the English answer that solves it. Multilingual embedding models put every language into one shared vector space so a query in one language retrieves documents in another, but only some models do it, and the choice turns on language coverage, whether cross-lingual matching is needed at all, embedding dimension, and how much text fits in one call.
Read article
Exam Room · Advanced GenAI
LLM-as-a-Judge: Designing a Rubric You Can Trust
Handing evaluation to a second model is fast and it scales, but a judge with a vague rubric or an unchecked bias just launders its own preferences into a number you then trust. Pointwise scoring against explicit criteria and pairwise comparison answer different questions, and both leak position, verbosity, and self-preference bias unless you design against it. The work is writing a concrete rubric, controlling the known biases, and calibrating the judge against human labels before you let it grade at scale.
Read article
Exam Room · Advanced GenAI
Monitoring a Production Bedrock App
A generative-AI feature on Amazon Bedrock breaks in three directions: the bill climbs, the responses slow down, and the answers quietly get worse. The platform hands you metrics and logs for the first two almost for free, but quality is the one signal Bedrock cannot measure for you, so you have to build it. Knowing which of the three you are looking at, and where the signal lives, is the whole job.
Read article
Exam Room · Advanced GenAI
Lab: Get Structured JSON Out With Tool Use
A hands-on lab. Asking a model for JSON in the prompt and hoping is how parsers break at 3am. This lab has you do it the reliable way: declare a tool schema, let the model fill it in, and read the answer back as parsed arguments rather than a string. You deploy a Lambda, turn a free-text support message into a structured record, and find out how much of the shape the schema really buys you.
Read article
Exam Room · Advanced GenAI
Building Deterministic Pipelines With Bedrock Flows
When a generative workflow has a sequence you already know, letting a model pick the order each time is the wrong trade: nondeterministic, harder to test, and an extra model call at every decision point. Amazon Bedrock Flows lets you draw the sequence as a fixed graph of nodes wired together with data links, so a designer owns the control flow and the model does the work inside a step. This walks the space between a Flow, a single agent, and Step Functions, and lands on when the drawn graph is the right home.
Read article
Exam Room · Advanced GenAI
Encrypting a Bedrock App End to End With KMS
Everything a Bedrock app stores is encrypted at rest by default, so the question is never whether the data is encrypted; it is who holds the key that decrypts it. A customer-managed KMS key turns encryption from a checkbox into a control you own: a lock you set the policy on, an audit trail of every use, and a switch you can throw to cut access without deleting a single byte. This walks the persistent artefacts one by one and asks, for each, whose key it is.
Read article
Exam Room · Advanced GenAI
Event-Driven GenAI: Processing Documents Asynchronously
Summarising a 200-page report or enriching a batch of records is a job that takes minutes, so wiring it behind a synchronous HTTP request guarantees timeouts, retries that re-run expensive work, and a bill that scales with impatience. The fix is to make the upload an event and let the work happen in the background. S3 notifications, SQS buffering, Lambda workers, Step Functions pipelines, and Bedrock batch inference each fit a different point on the volume-and-latency curve, and picking the wrong one shows up as either idle capacity or dropped documents.
Read article
Exam Room · Advanced GenAI
Caching LLM Responses Without Stale Answers
Thirty percent of the support assistant's queries are paraphrases of each other, 'how do I cancel?' 'can I cancel?' 'where's the cancel button?', and every one pays full model price. Caching LLM responses isn't as simple as hashing a prompt: exact-match, semantic, and prefix caching answer different questions, and getting the boundary wrong serves yesterday's answer to today's question.
Read article
Exam Room · Advanced GenAI
A/B Testing Prompts and Models in Production
A new prompt scores better on the test set, a newer model is cheaper per token, and both look like obvious upgrades until they meet real traffic and quietly regress a slice of it. Offline scores tell you a variant is plausible; only live traffic tells you it is better. The job is deciding how much risk each change deserves, shadowing or splitting accordingly, measuring quality and latency and cost together, and rolling the winner out in a way you can undo in seconds.
Read article
Exam Room · Advanced GenAI
Right-Sizing Provisioned Throughput for a Custom Model
A fine-tuned model on Amazon Bedrock cannot be called on demand; the only way to serve it is Provisioned Throughput, bought in model units you reserve and pay for whether traffic fills them or not. Size it from peak tokens per minute with headroom, then choose a commitment term, because a six-month lock is cheap per unit and unforgiving if demand moves. Over-provision and you burn money on idle units; under-provision and you throttle real requests. The sizing is the whole game.
Read article
Exam Room · Advanced GenAI
Defending Against Indirect Prompt Injection in RAG
A retrieval-augmented assistant answers from a knowledge base you built over your own documents, so the retrieved context feels trusted. It is not. The moment a document can be edited by a partner, crawled from the web, or fed in from user-generated content, an attacker can plant instructions that ride into the model alongside the genuine text, and the model cannot tell your data apart from a command hidden inside it.
Read article
Exam Room · Advanced GenAI
Cost Attribution and Tagging for GenAI Workloads
A single Bedrock bill arrives as one on-demand line, and the finance team wants to know which team, feature, and customer spent it. Application inference profiles, cost allocation tags, model invocation logging, Cost Explorer, and Budgets each pin the spend at a different grain, from whole account down to a single request. The job is choosing the grain that answers the chargeback question, then wiring the alert that closes the loop before the next bill lands.
Read article
Exam Room · Advanced GenAI
Lab: Give a Bedrock Chatbot a Memory
A hands-on lab. A model call is stateless: ask a chatbot your name, ask it back, and it has already forgotten. Memory is not a feature you switch on, it is a transcript you replay, stored somewhere between requests. This lab hands you a Lambda and a DynamoDB table and has you build the load-and-save loop that turns single calls into a conversation, then reason about what replaying all those turns costs.
Read article
Exam Room · Advanced GenAI
Building a Voice Assistant: Transcribe, Bedrock, and Polly
A voice assistant is really four pieces bolted together: speech becomes text, a model reasons over the text, the reply becomes speech, and something manages the back-and-forth. The trap is picking each piece in isolation and discovering at integration time that the latency has stacked up past what a caller will tolerate, or that a name has landed in a model prompt that a compliance rule says should never have been kept. Getting it right means budgeting latency across the whole chain and putting safety on the text stage, where the words actually live.
Read article
Exam Room · Advanced GenAI
Handling Throttling and Rate Limits Gracefully
A Bedrock feature that ran clean in testing starts returning ThrottlingException the week traffic doubles, and the temptation is to bolt on a retry loop and move on. Retries smooth a transient spike, but they add no capacity, so a genuinely undersized workload just retries its way into a slower failure. The real work is telling a transient limit from a structural one, then reaching for backoff, a quota increase, cross-region inference, provisioned throughput, or a queue depending on which one the traffic actually needs.
Read article
Exam Room · Advanced GenAI
Cutting Ingestion Cost by Caching and Batching Embeddings
A retrieval system re-embeds its whole corpus on every update, and the bill scales with a knowledge base that barely changes. Embedding is a per-token cost paid on every chunk you send, so re-embedding text that has not changed is money spent to arrive at the vector you already had. Incremental sync, content-hash caching, de-duplication, request batching, and a smaller embedding dimension each cut a different slice of that waste, and each carries a bookkeeping cost worth naming before you reach for it.
Read article
Exam Room · Advanced GenAI
Choosing Between Kiro, Amazon Quick, and Bedrock
AWS sells finished AI assistants and it sells the platform to build your own, and most bad decisions in this corner come from confusing the two. Kiro is the assistant developers adopt; Amazon Quick is the assistant staff use over enterprise data; Bedrock is the platform a customer-facing feature gets built on. Sort by who the output is for and the choice mostly makes itself.
Read article
Exam Room · Advanced GenAI
Keeping a Knowledge Base Fresh Without Re-Embedding Everything
A retrieval assistant is only as good as the documents behind it, and those documents keep changing while the embedding bill for re-processing them keeps climbing. A support team on Amazon Bedrock has a Knowledge Base over forty thousand policy and product files, and their nightly full re-sync costs more than the queries it serves. The fix is knowing what actually needs re-embedding, when to trigger the work, and when retrieval is the wrong tool for the fact entirely.
Read article
Exam Room · Advanced GenAI
How to Wire Function Calling Through Bedrock
An assistant that knows the answers but can't act on them is half a tool. Function calling, letting the model invoke tools we define, with arguments it chooses, turns understanding into action. Bedrock's Converse API has native tool-use support; so does Anthropic's Messages API via Bedrock; so do Bedrock Agents as a higher-level wrapper. Each exposes function calling through a different surface, and picking wrong makes the simple case hard.
Read article
Exam Room · Advanced GenAI
When to Orchestrate With Step Functions Instead of an Agent
A multi-step GenAI job can be run three ways: a Bedrock agent that lets the model decide the sequence, a Bedrock Flow that draws the sequence as a fixed graph, or a Step Functions state machine that runs the sequence as a durable workflow with the model as one step among many. The deciding line is whether the control flow is known ahead of time and how much durability, retry, and cross-service reach the job needs. This walks the space and lands on when the state machine is the right home.
Read article
Exam Room · Advanced GenAI
Building a Golden Dataset for LLM Evaluation
Every prompt tweak, model swap, and RAG change gets judged against something, and if that something is a handful of screenshots in a chat thread, the judgement is noise. A golden dataset is the fixed yardstick: representative inputs paired with accepted answers, covering the easy middle, the known-hard cases, and the questions the system should refuse. Building one that stays trustworthy and reusable is the work that makes every later evaluation mean something.
Read article
Exam Room · Advanced GenAI
Lab: Build RAG From Scratch
A hands-on lab. A Knowledge Base hides retrieval behind an API; this lab makes you build the thing it hides. Five documents, a model, and one function you write: embed the question, compare it to the documents, ground the answer in the closest ones. No vector store, so the mechanics of RAG are in plain sight, and the concept stops being magic.
Read article
Exam Room · Advanced GenAI
Cheat Sheet: Model Selection and Inference
A condensed revision sheet for choosing a model on Amazon Bedrock and picking how it runs: providers, the Nova family, selection axes, and every inference mode from on-demand to Provisioned Throughput to SageMaker hosting.
Read article
Exam Room · Advanced GenAI
Red-Teaming a Bedrock Application
A team ships a generative-AI assistant on Bedrock and waits to see what breaks. Red-teaming flips that around: you attack your own application on purpose, before an outsider does, probing for jailbreaks, injected instructions, leaked data, and tools that can be talked into doing damage. The work that lasts is not the single afternoon of creative attacks; it is turning every finding into a test that runs forever.
Read article
Exam Room · Advanced GenAI
Content Moderation With Rekognition, Comprehend, and Guardrails
A single generative app now takes uploaded photos, voice notes, and free text, and it also emits model output that has to be safe to show. One moderation service cannot cover all of that, because the right tool is decided by the media and by the stage of the pipeline. Rekognition reads images and video, Comprehend reads text, Transcribe bridges audio into text, and Bedrock Guardrails sits on the model prompt and completion. The job is routing each piece of content to the service built for it rather than forcing one tool to do everything.
Read article
Exam Room · Advanced GenAI
Cost Guardrails: Budgets, Quotas, and Model Choice
A generative-AI feature can double its own bill between two invoices, and the first anyone hears of it is the invoice. Per-token pricing, model tier, replayed context, call volume, and Provisioned Throughput commitments are the dials that set the number, and every one of them can be bounded in advance. The job is designing the spend limits in up front, budgets and quotas and model choice together, so an overrun trips an alert on day two rather than a surprise on the statement.
Read article
Exam Room · Advanced GenAI
Routing Requests Between a Cheap and a Capable Model
Most traffic to an LLM feature is easy: a label, a short extraction, a one-line answer that a small cheap model handles as well as the flagship. A minority needs the big model to reason through several steps. Paying flagship prices on every request treats the whole workload as if it were the hard tail, so the money and the latency go where they are not needed. Routing splits the stream, and the risk is a misroute that sends a hard request to the weak model and gets a confident wrong answer.
Read article
Exam Room · Advanced GenAI
Summarising Long Conversations to Fit the Context Window
A support copilot on Bedrock works fine for the first dozen turns, then the transcript it replays on every call grows until the bill climbs and the oldest messages fall off the edge of the context window. Replaying every turn is the simple thing, and it is the thing that breaks first. Sliding windows, running summaries, and extracted facts each trade history for room in a different way, and the split between a recent transcript and a durable store is the line between short-term and long-term memory.
Read article
Exam Room · Advanced GenAI
Choosing an Embedding Dimension and Its Storage Cost
Embedding dimension looks like a knob where higher is simply better, but every extra dimension multiplies across every vector in the index, and storage plus search cost climb in lockstep. Amazon Titan Text Embeddings v2 lets you ask for 1024, 512, or 256 dimensions, so the choice is a real trade of retrieval quality against footprint and latency. The job is to size the corpus, set a quality bar, and measure what each dimension actually buys on your own data.
Read article
Exam Room · Advanced GenAI
Lab: Wire a Tool the Model Can Call
A hands-on lab. In an earlier lab the model filled in a schema and stopped; here it goes further, deciding on its own to call a tool, waiting for the result, and answering from it. That decide-call-read-continue cycle is exactly what a managed agent runs under the hood. You build it by hand, so an action group stops being a black box.
Read article
Exam Room · Advanced GenAI
Cheat Sheet: Prompt Engineering
A condensed revision sheet for prompt engineering on Bedrock: techniques, decision rules, traps, and one-line facts for the AIP-C01 Generative AI Developer track.
Read article
Exam Room · Advanced GenAI
Tracing an Agent's Decisions in Production
An agent told a subscriber they were owed the wrong refund, and the final answer says nothing about why. To debug it you have to reconstruct the run: which tool it called, with what arguments, what came back, and how it reasoned from one step to the next. Bedrock Agents can emit a trace of exactly that, model invocation logging captures the prompts and completions, and distributed tracing ties the Lambda tools into one picture. Getting all three in place is the difference between guessing and knowing.
Read article
Exam Room · Advanced GenAI
Multi-Region Resilience for a GenAI Service
A Bedrock service is fine until the day its region has a bad hour, or until on-demand throttling caps it at the worst possible moment. Cross-region inference profiles, a warm second region with the same guardrails and knowledge base, replicated RAG data, and an honest read on which models the failover region actually supports all pull in the same direction. The decision comes down to the availability target, where the data is allowed to be processed, and what a second region costs to keep warm.
Read article
Exam Room · Advanced GenAI
Delivering Responses: Sync, Async, or Streaming
Every generative feature has to decide how the answer reaches the user: one synchronous reply, a stream of tokens as they are generated, or an asynchronous job that finishes in the background. The wrong choice shows up as a spinner that stalls for thirty seconds, a gateway timeout on a long completion, or a batch of ten thousand documents jammed through a real-time endpoint one at a time. The deciding factors are how interactive the moment is, how long the output runs, and whether the caller can afford to wait.
Read article
Exam Room · Advanced GenAI
Preparing a Dataset for Fine-Tuning
Fine-tuning a foundation model lives or dies on the data you feed it, and most teams reach for volume when they should be reaching for consistency. A few hundred clean, representative, identically formatted prompt-completion pairs will out-teach a scraped heap of thousands. The work is shaping that set: the JSONL schema the model expects, a train and validation split with no leakage, PII and duplicates stripped, and a held-out set to judge the result. Get the data right and the job is almost a formality.
Read article
Exam Room · Advanced GenAI
Choosing Between Kendra and a Bedrock Knowledge Base
A retrieval-augmented system lives or dies on its retrieval layer, and on AWS that used to be a choice between two very different tools: a Bedrock Knowledge Base that hands you the vector pipeline, or Amazon Kendra that hands you managed connectors, mature relevance, and document-level access control. Kendra went into maintenance mode on 30 June 2026 and closed to new customers on 30 July, so for a new build the choice is already made. The useful question is which of the things Kendra was doing the replacement kept, and the answer is more of them than the migration guide suggests.
Read articleExtracting Structured Data From Documents at Scale
A pile of invoices, forms, and scanned contracts has to become clean database rows, and the tool choice decides everything downstream. Amazon Textract reads layout and text deterministically; a foundation model on Bedrock reasons over meaning but needs validating; Bedrock Data Automation wraps the whole pipeline; and the reliable pattern for most real documents combines them. The job is matching the tool to whether the work is reading pixels or understanding meaning.
Generative AI Developer · AIP-C01
Coming soon