Context, Knowledge, and Trusted Government Data

10 min

Prompt vs. Context

A Prompt is "what is being asked this time"; Context is "the background information AI should draw on when answering." Conflating the two is often where government AI applications go wrong: a caseworker assumes a clearly written question is enough, forgetting that AI has no idea about internal procedures or case history unless given them.

• Context Engineering sorts information into four categories: Include (put directly into the prompt), Retrieve (pulled dynamically), Remember (kept across turns), and Exclude (explicitly withheld).
• Include: suited to small, always-needed, rarely-changing information, e.g. the complaint text itself.
• Retrieve: suited to large, not-always-needed, updating information, e.g. a full regulation database.
• Remember: suited to information that must carry over across multiple interactions, e.g. the case number and current status.
• Exclude: information explicitly withheld, this does not mean "forgotten," it means "never provided in the first place."

Key takeaway: Prompt decides "what is being asked"; Context decides "whether AI can get it right"; most government AI quality problems trace back to poorly designed Context, not to an unclear prompt.

14 min

RAG, source trust, and the data lifecycle

RAG (Retrieval-Augmented Generation) is the most common way to implement Knowledge Retrieval: before answering, AI dynamically pulls relevant content from a knowledge base and generates its answer based on what it retrieved. Government work often needs regulations, internal procedures, and past cases, and the quality and management of those sources directly determines how reliable a RAG answer can be.

• Data lifecycle: data moves through published, effective, revised, and retired states, and the knowledge base must stay in sync or it will cite a retired rule.
• Outdated-source risk: if a regulation is revised but the knowledge base is not updated in time, AI may generate an answer based on the old rule, and the caseworker may not notice.
• Conflicting sources: when sources contradict each other (e.g. an old and new version both exist), a human should decide which to use, AI should never decide this on its own or silently ignore the conflict.

Key takeaway: How well RAG answers depends on the knowledge base's metadata, version management, and source trust; outdated or conflicting sources must always be resolved by a human, never left to AI's own judgment.

12 min

Government data classification: six tiers

Key takeaway: The six classification tiers are not terms to memorize, they are the question you should ask every single time before handing data to AI: "which tier does this belong to?"

12 min

Government scenario: context design for a complaint reply

• This scenario shows how Include / Retrieve / Remember / Exclude can all appear in the same use case, complementing rather than excluding each other.
• The Exclude list is often more important than the Include list: missing an item that should be excluded is usually far riskier than missing one that should be included.
• Citation and provenance: an AI answer should be able to point to which regulation or handling principle it drew on, so the caseworker can check the draft item by item.

Key takeaway: Good context design is not about giving as much information as possible, it is about precisely deciding what to give and what must never be given; how rigorous the Exclude list is often determines whether the whole use case is safe.

12 min

Practice: What Should the AI Know?

Activity 3 continues the use case selected in Module 1, classifying every candidate piece of information as Include / Retrieve / Remember / Exclude, and filling in the Government Context & Data Canvas.

1. 1List candidate information
   List every information source this use case might draw on.
2. 2Classify as Include / Retrieve / Remember / Exclude
   Judge each source item by item.
3. 3Mark the source and sensitivity
   For each Include/Retrieve item, note the owning unit and data classification.
4. 4Write the exclusion reason
   For each Exclude item, write one sentence on why it is withheld from AI.

• Government Context & Data Canvas minimum fields: Use Case, Required Context, Trusted Sources, Source Owner, Retrieval Method, Version / Freshness Rule, Metadata, Personal Data, Sensitive Data, Restricted Data, Exclude Rules, Citation Requirement, Conflict Resolution Rule, Retention / Memory Rule, Human Verification.

• Go to the practice page: Activity 3 What Should the AI Know?

Key takeaway: This Canvas turns Module 3's judgment method into concrete rules that Module 4 (tool permissions) and Module 5 (governance) can directly reuse.