RAG Answer Policy Skill

Purpose

Use this skill to design a response policy for a RAG assistant or knowledge-base bot.

The goal is to prevent unsupported answers, source confusion, hallucinated details, and overconfident responses when context is incomplete.

When to use

Use this skill when building prompts for:

• customer-support RAG bots;
• internal knowledge-base assistants;
• course assistants;
• documentation assistants;
• legal/financial/medical document assistants;
• website chatbots trained on company content;
• assistants that answer using retrieved documents.

Required input

Collect or infer:

• assistant purpose;
• knowledge sources;
• whether external knowledge is allowed;
• citation requirements;
• answer format;
• escalation rules;
• no-answer behavior;
• audience expertise.

Procedure

1. Define allowed sources

Specify what the assistant may use:

• retrieved documents;
• uploaded files;
• website content;
• database records;
• user-provided context;
• model knowledge;
• web search.

Define source priority order.

2. Define forbidden inference

List what the assistant must not invent:

• facts not present in sources;
• prices;
• policies;
• legal claims;
• availability;
• contact information;
• medical or financial recommendations;
• internal company details;
• user preferences.

3. Define citation behavior

Specify when citations are required.

For source-grounded systems, prefer:

• cite key factual claims;
• cite the exact document or section when available;
• do not cite sources that do not support the claim;
• say when no supporting source is available.

4. Define insufficient context behavior

When sources do not contain enough information, the assistant should not guess.

Choose one or more behaviors:

• say that the available sources do not contain the answer;
• ask a clarifying question;
• suggest where the user may find the answer;
• escalate to a human;
• provide a partial answer with explicit uncertainty.

5. Define conflicting source behavior

When sources conflict, the assistant should:

• identify the conflict;
• prefer the highest-priority or most recent source if rules allow;
• state uncertainty;
• avoid merging conflicting facts into one confident answer;
• escalate if needed.

6. Define answer style

Specify:

• language;
• tone;
• depth;
• formatting;
• whether to include summaries;
• whether to include next steps.

7. Define escalation rules

Escalate when:

• the answer affects high-stakes decisions;
• the source is missing or conflicting;
• the user requests something outside assistant scope;
• the user asks for human support;
• the assistant cannot complete the request safely.

Output format

Return a ready-to-use RAG answer policy:

# RAG Answer Policy

## Purpose

This assistant answers user questions using approved knowledge sources.

## Source Priority

Use sources in this order:

1.
2.
3.

## Allowed Sources

The assistant may use:

- ...

## Forbidden Inference

The assistant must not invent:

- ...

## Citation Rules

- Cite factual claims when sources are available.
- Do not cite unsupported sources.
- If the source does not contain the answer, say so.

## Insufficient Context

If the available sources do not contain enough information, the assistant should:

1.
2.
3.

## Conflicting Sources

If sources conflict, the assistant should:

1.
2.
3.

## Answer Style

- Language:
- Tone:
- Detail level:
- Formatting:

## Escalation Rules

Escalate to a human when:

- ...

## Final Check

Before answering, verify:

- [ ] The answer is grounded in allowed sources.
- [ ] Unsupported claims are not presented as facts.
- [ ] Missing context is handled explicitly.
- [ ] Source conflicts are disclosed.
- [ ] The answer follows the required format.

Quality checklist

A good RAG policy should define:

• allowed sources;
• source priority;
• citation behavior;
• insufficient context behavior;
• conflict resolution;
• no-answer policy;
• escalation rules.

Boundaries

This skill creates an answer policy. It does not implement retrieval, ranking, chunking, embeddings, or citation extraction.