An output can change because the request, available context, retrieved evidence, tool access, system instructions, or workflow changed. These layers improve capability only within their own quality and permission limits.
Plan for this page
Understand the idea, test it actively, then explain it in your own words.
You will leave with
distinguish prompt wording from the full context assembled by a product explain retrieval and tool use without treating either as automatic truth choose when to add information, retrieve a source, use a deterministic tool, or require human action
Time
30 min
Before you begin
AI from Zero is useful preparation, but it is not a locked prerequisite. Each unit states the concepts it uses.
Do this now
Read the outcomes and useful terms before the first section.
distinguish prompt wording from the full context assembled by a product
explain retrieval and tool use without treating either as automatic truth
choose when to add information, retrieve a source, use a deterministic tool, or require human action
Terms you will use
These definitions prepare you for the reading; you do not need to memorize them.Prompt
Input or instruction supplied to a model, whether written by a user, developer, system, template, or another tool.
Retrieval
Finding external records or documents and adding selected material to the current task or context.
Tool use
Calling a separate capability—such as search, calculation, code, a database, or an action API—under defined permissions.
01
Prompting is task communication, not a spell
Clear goals, relevant context, constraints, examples, output format, and uncertainty instructions can make a task easier to perform and review. They cannot create missing facts, permissions, model capability, or a guarantee.
The visible user message may be only one part of context. Products can add system instructions, conversation history, retrieved snippets, memory, tool results, metadata, and safety rules. Ask what the system actually receives before attributing behavior to wording alone.
02
Retrieval changes access, not source quality
Retrieval-augmented systems search a collection and supply selected material to generation or another process. Results depend on the collection, indexing, query, ranking, filters, access controls, document freshness, and how the model uses the excerpts.
A retrieved document can be authoritative, irrelevant, outdated, malicious, misunderstood, or insufficient. Citation presence therefore shows a trace to inspect, not proof that the claim is supported.
03
Tools expand consequences and control needs
A calculator may improve arithmetic, code execution may test a procedure, and an API may read a calendar or send a message. The model still needs to choose the right tool, form correct arguments, interpret the result, and respect permissions.
Read-only lookup, drafting, proposing, and executing are different authority levels. Place confirmation, scoped credentials, logs, rate limits, validation, and recovery according to what the tool can change and who may be affected.
Interactive model
Choose the missing layer
Selected layerClarify
Improve the goal, audience, constraints, examples, format, and definition of a good result.
Read the complete text alternative
ClarifyImprove the goal, audience, constraints, examples, format, and definition of a good result.
Supply contextProvide relevant authorized information that fits the context and privacy boundary.
Retrieve or calculateUse current documents, structured records, search, code, or deterministic computation when the task needs them.
Control actionLimit permissions, validate arguments, request approval, log effects, and provide recovery before execution.
Misconception to correct
“If an answer includes a citation, retrieval has verified it.”
Retrieval selects material. Verification still asks whether the source is authoritative, current, in scope, accurately represented, and sufficient for the exact claim.
Active checks
Decide, then compare the reasoning
No grade, score, or streak. Progress records completed learning objects, not your worth or ability.01
An assistant must answer today's train-departure time. Which change matters most?
02
A meeting assistant can read calendars and draft invitations. What should happen before it sends one?
Explain back and transfer
Diagnose the missing layer
Choose an AI task that currently performs poorly. Decide whether it needs clearer instructions, authorized context, retrieval, a deterministic tool, a permission boundary, human judgment, or no AI at all.
Answer both checks before completing.
Source context
What supports this unit and where it stops
Evidence status: established
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)U.S. National Institute of Standards and Technology
Risk and control context for generative systems; retrieval or tool access does not remove the need to evaluate sources, actions, and system behavior.
inAi