Skip to content
NTPU AI Learn

AI 102 · Module 3

Context Selection: What Is Worth Letting the AI Know?

Which information should the AI really see?

Start Section 1
  • Duration60 min
  • LevelFoundation
  • PrerequisiteAI 101 recommended, plus Modules 1 and 2 of this course
0 / 7 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Use RAFS (Relevance, Authority, Freshness, Sensitivity) to judge whether a piece of information belongs in the context
  • Explain what a Source of Truth is, and decide which source prevails when sources conflict
  • Recognize the risk of Context Dumping and trim context using the need-to-know principle
  • Make a Use, Retrieve, Verify or Exclude judgment for a set of sources, and explain the reasoning

The Problem

7 min

Findable Does Not Mean Fit to Give the AI

Once you know the types of context, the next question is selection. In practice, data is rarely scarce; judgment is. Which of it is trusted, which is out of date, and which should not be provided?

Work scenario (illustrative)

A colleague wants the AI to explain how annual leave is calculated for new staff. On hand: the official HR regulations, a two-year-old version of them, a manager's explanation in a group chat, an article from an online forum, and some HR records containing employee names.

All of these are "related". But what happens if they all go to the AI?

  • The old and current regulations contradict each other, and the AI may pick either one
  • The forum article looks detailed, yet nobody answers for it
  • HR records with names have nothing to do with the question, yet they were sent along

Warning

Handing the AI everything you can find is not caution; it hands the judging over to the AI. Choosing which information may be used is a human responsibility.

Key takeaway: Plenty of data is not the problem; missing selection is. A person does the selecting.

The RAFS Quality Model

9 min

RAFS: Four Criteria

Faced with a piece of information, four questions give a quick check. Their initials spell the RAFS quality model.

The RAFS context quality model
CriterionQuestion to askTypical consequence of failing
R: Relevance (相關性)Will this really affect the result of the current task?Context Dumping: more noise, key points drowned out
A: Authority (權威性)How trustworthy is the source? Who answers for it?Relying on an unreliable claim, or choosing wrongly among conflicting sources
F: Freshness (時效性)Is it current? Which date and which version?Information that was correct but is now out of date
S: Sensitivity (敏感性)Is it appropriate to provide to the AI?Personal, confidential or unauthorized data used improperly

Read the four together. A document can be highly relevant and authoritative yet out of date, or fresh and relevant yet confidential and not to be shared.

Key takeaway: RAFS: relevance, authority, freshness, sensitivity. Check all four together.

8 min

Relevance and Freshness

Context Dumping

  • Whole handbooks and whole folders pasted in
  • Higher cost, and key information drowned out
  • Results may be unstable
  • Hard to tell which passage the AI actually relied on

Focused context

  • Only passages related to this task
  • Every item can be justified with "why it is needed"
  • Results are easier to check and reproduce
  • When more is needed, it is obtained separately

Freshness problems are subtler: data can be entirely correct and simply belong to the past. A 2024 price list, for example, is no basis for a 2026 purchase. Build this habit:

Information + Timestamp + Version
  1. Information: the content itself
  2. Timestamp: which day the data is from
  3. Version: which edition it is, and whether it is the current one

Info

Data without a date or version is not automatically unusable, but the user should know they are taking on extra uncertainty.
Check yourself

To answer one employee's annual leave question, someone attaches almost 200 company documents, saying "this is the safest". By RAFS, what is the most reasonable comment?

To answer one employee's annual leave question, someone attaches almost 200 company documents, saying "this is the safest". By RAFS, what is the most reasonable comment?

Key takeaway: Put information in context for a reason, and know its date and version.

Authority and Source of Truth

10 min

Authority and Source of Truth

Sources differ in trustworthiness. When several disagree, you need a rule for which one prevails. The source that settles the matter is called the Source of Truth (權威資料來源).

Authority ladder (illustrative, highest to lowest)
  1. Official policy
  2. Manager-approved version
  3. Internal wiki
  4. Old email
  5. Chat message
  6. Personal recollection

Info

This ladder is illustrative, not a universal rule. The order can differ by organization and task; in some fields, for instance, law outranks internal policy. The point is to decide in advance rather than pick by feel in the moment.
  1. Look at authority first
    Formal sources that someone answers for and that went through a publishing process outrank informal claims.
  2. Then version and date
    Among versions of the same source, the current one usually prevails.
  3. If a conflict remains, surface it
    Do not let the AI quietly choose one. List the conflict and hand it to someone with authority to decide.

High authority does not mean always right: official documents can contain errors or be replaced by a newer notice. So read authority together with freshness.

Check yourself

The current official travel policy (version 4) caps lodging at NT$2,500 per night. A manager once said in a chat group, "the cap is NT$3,000 this time". You ask the AI to prepare a trip budget. What is the most reasonable handling?

The current official travel policy (version 4) caps lodging at NT$2,500 per night. A manager once said in a chat group, "the cap is NT$3,000 this time". You ask the AI to prepare a trip budget. What is the most reasonable handling?

Key takeaway: Decide the Source of Truth in advance. When sources conflict, surface it and let someone with authority rule.

Sensitivity

8 min

Sensitivity: Need-to-know

Relevant does not mean permitted. Some information should stay out of an AI's context even if it would help the task. Common types include:

  • Personal data
  • Confidential data and trade secrets
  • Credentials such as accounts, passwords and API keys
  • Information not yet public
  • Documents you are not authorized to use

Warning

Need-to-know context: provide only what must be known to do this task. Before using any AI tool, check what your organization's rules say about data use.

Pasting the original complaint

Asking the AI to draft a complaint reply and pasting the full complaint, including the customer's name, phone, address and order number.

Adjusted by need-to-know

De-identify
Remove the name, phone and address, and refer to "Customer A".
Keep only what is needed
Keep the reason for the complaint and how it was handled; that is what a reply requires.
A person reviews and restores
Where personal details are needed, the case handler fills them in outside the AI output.
Check yourself

You want an AI to help draft a reply to a customer complaint. Which approach best fits need-to-know?

You want an AI to help draft a reply to a customer complaint. Which approach best fits need-to-know?

Key takeaway: Relevant is not the same as permitted. Provide only what the task must know.

Practice

12 min

Practice: 12 Sources, Use, Retrieve, Verify or Exclude?

Continuing the leave scenario: a colleague asks an AI to help answer "what are the leave rules for new staff?". Below are 12 sources on hand (all illustrative). Choose how to handle each one and explain with RAFS. The same source can call for a different answer on a different task.

12 sources (decide for yourself)
SourceThink about
1. Official HR policyIs it the current version? It is large; should all of it be given?
2. HR policy from two years agoDoes it match the current one? Could it be useful for comparison?
3. A manager's explanation on SlackIs there a formal basis? Is it a temporary verbal instruction?
4. An online forum postWho answers for the content? Does it apply to our organization?
5. A government websiteIs it the source of the law? Is the page the latest announcement?
6. A draft PDFIt is not final; can it serve as a basis?
7. Personal notesWhose are they? When were they written?
8. CRM dataDoes it contain customer personal data? Which fields does this task need?
9. A customer emailMay contain personal data and unverified claims
10. A news reportIs it a second-hand source? Can it be traced to the original announcement?
11. A database query resultWhen was it run and with what conditions? How often is the data updated?
12. An AI-generated summaryWhich original text is it based on? Has it been checked?
  • Use: use directly, because it is relevant, trusted, current and appropriate to provide
  • Retrieve: the material is large or changes often; fetch the relevant part when needed
  • Verify: check the source, version or content first, and use it only once confirmed
  • Exclude: should not be provided, or should not serve as a basis
Check yourself

Among the sources is an "AI-generated summary" of how annual leave is calculated. You plan to hand it to another AI as a basis. By RAFS, what is the most reasonable judgment?

Among the sources is an "AI-generated summary" of how annual leave is calculated. You plan to hand it to another AI as a basis. By RAFS, what is the most reasonable judgment?

Key takeaway: The same source can be handled differently on different tasks. Giving reasons with RAFS matters more than memorizing answers.

Recap

6 min

Recap: Selection Is a Human Responsibility

Done

Good context is the right information, trusted, sufficient and safe, provided at the right time. Choosing it is a person's responsibility.
  • RAFS: Relevance, Authority, Freshness, Sensitivity
  • Avoid Context Dumping: every item needs a stated reason
  • Decide the Source of Truth in advance; when sources conflict, flag it and hand it to someone with authority
  • Carry a date and version with data; out of date is not the same as wrong, but it cannot be used as a basis directly
  • Need-to-know: relevant is not the same as permitted

The following modules turn these judgments into concrete strategy: what to include directly (Include), what to retrieve when needed (Retrieve), what is worth remembering (Remember), and what to exclude (Exclude).

Key takeaway: Select information with RAFS, and remember that source trust, freshness and whether it may be provided remain a person's responsibility.

Save as PDF

What next?