# Refusal to Answer

## The Problem

LLM refuses to answer legitimate queries, incorrectly flagging them as unsafe or outside its capabilities despite having relevant context.

### Symptoms

* ❌ "I cannot help with that" for normal queries
* ❌ False safety triggers
* ❌ Refuses despite having answer in context
* ❌ Overly cautious responses
* ❌ Generic deflections

### Real-World Example

```
Query: "How do I delete my account?"

Retrieved context: Clear deletion procedure in docs

AI response: "I cannot provide guidance on account deletion.
Please contact customer support for account management."

Problem: Safety training makes model cautious about "deletion"
Ignores legitimate documented procedure
```

***

## Deep Technical Analysis

### Safety Training Conflicts

RLHF safety vs helpfulness:

**Overactive Safety:**

```
Model trained to refuse:
→ Harmful instructions
→ Personal data requests
→ Illegal activities

But triggers on legitimate queries:
→ "How to hack my own API?" (security testing)
→ "Delete user data" (GDPR compliance)
→ "Bypass rate limit" (documented method)

Safety overrides RAG context
```

**Capability Uncertainty:**

```
Model uncertain if it can answer:
→ Specialized technical question
→ "I'm not qualified to..."
→ Even when context has answer

Training to admit limits becomes excessive caution
```

***

## How to Solve

**Add explicit permission in system prompt ("You have complete knowledge in provided context") + use RAG-specific fine-tuned models less prone to refusal + implement output retry with rephrased prompt if refused + whitelist domains known to be safe.** See [Refusal Handling](/rag-scenarios-and-solutions/llm/refusal-to-answer.md).


---

# Agent Instructions: Querying This Documentation

If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://help.twig.so/rag-scenarios-and-solutions/llm/refusal-to-answer.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.