fss-mini-rag-github/claude_rag/llm_safeguards.py
BobAi 5f42751e9a 🛡️ Add comprehensive LLM safeguards and dual-mode demo scripts
🛡️ SMART MODEL SAFEGUARDS:
- Implement runaway prevention with pattern detection (repetition, thinking loops, rambling)
- Add context length management with optimal parameters per model size
- Quality validation prevents problematic responses before reaching users
- Helpful explanations when issues occur with recovery suggestions
- Model-specific parameter optimization (qwen3:0.6b vs 1.7b vs 3b+)
- Timeout protection and graceful degradation

 OPTIMAL PERFORMANCE SETTINGS:
- Context window: 32k tokens for good balance
- Repeat penalty: 1.15 for 0.6b, 1.1 for 1.7b, 1.05 for larger models
- Presence penalty: 1.5 for quantized models to prevent repetition
- Smart output limits: 1500 tokens for 0.6b, 2000+ for larger models
- Top-p/top-k tuning based on research best practices

🎬 DUAL-MODE DEMO SCRIPTS:
- create_synthesis_demo.py: Shows fast search with AI synthesis workflow
- create_exploration_demo.py: Interactive thinking mode with conversation memory
- Realistic typing simulation and response timing for quality GIFs
- Clear demonstration of when to use each mode

Perfect for creating compelling demo videos showing both RAG experiences!
2025-08-12 19:07:48 +10:00

320 lines
12 KiB
Python

#!/usr/bin/env python3
"""
LLM Safeguards for Small Model Management
Provides runaway prevention, context management, and intelligent detection
of problematic model behaviors to ensure reliable user experience.
"""
import re
import time
import logging
from typing import Optional, Dict, List, Tuple
from dataclasses import dataclass
logger = logging.getLogger(__name__)
@dataclass
class SafeguardConfig:
"""Configuration for LLM safeguards."""
max_output_tokens: int = 2000 # Prevent excessive generation
max_repetition_ratio: float = 0.3 # Max ratio of repeated content
max_response_time: int = 60 # Max seconds for response
min_useful_length: int = 20 # Minimum useful response length
context_window: int = 32768 # Ollama context window
enable_thinking_detection: bool = True # Detect thinking patterns
class ModelRunawayDetector:
"""Detects and prevents model runaway behaviors."""
def __init__(self, config: SafeguardConfig = None):
self.config = config or SafeguardConfig()
self.response_patterns = self._compile_patterns()
def _compile_patterns(self) -> Dict[str, re.Pattern]:
"""Compile regex patterns for runaway detection."""
return {
# Excessive repetition patterns
'word_repetition': re.compile(r'\b(\w+)\b(?:\s+\1\b){3,}', re.IGNORECASE),
'phrase_repetition': re.compile(r'(.{10,50}?)\1{2,}', re.DOTALL),
# Thinking loop patterns (small models get stuck)
'thinking_loop': re.compile(r'(let me think|i think|thinking|consider|actually|wait|hmm|well)\s*[.,:]*\s*\1', re.IGNORECASE),
# Rambling patterns
'excessive_filler': re.compile(r'\b(um|uh|well|you know|like|basically|actually|so|then|and|but|however)\b(?:\s+[^.!?]*){5,}', re.IGNORECASE),
# JSON corruption patterns
'broken_json': re.compile(r'\{[^}]*\{[^}]*\{'), # Nested broken JSON
'json_repetition': re.compile(r'("[\w_]+"\s*:\s*"[^"]*",?\s*){4,}'), # Repeated JSON fields
}
def check_response_quality(self, response: str, query: str, start_time: float) -> Tuple[bool, Optional[str], Optional[str]]:
"""
Check response quality and detect runaway behaviors.
Returns:
(is_valid, issue_type, user_explanation)
"""
if not response or len(response.strip()) < self.config.min_useful_length:
return False, "too_short", self._explain_too_short()
# Check response time
elapsed = time.time() - start_time
if elapsed > self.config.max_response_time:
return False, "timeout", self._explain_timeout()
# Check for repetition issues
repetition_issue = self._check_repetition(response)
if repetition_issue:
return False, repetition_issue, self._explain_repetition(repetition_issue)
# Check for thinking loops
if self.config.enable_thinking_detection:
thinking_issue = self._check_thinking_loops(response)
if thinking_issue:
return False, thinking_issue, self._explain_thinking_loop()
# Check for rambling
rambling_issue = self._check_rambling(response)
if rambling_issue:
return False, rambling_issue, self._explain_rambling()
# Check JSON corruption (for structured responses)
if '{' in response and '}' in response:
json_issue = self._check_json_corruption(response)
if json_issue:
return False, json_issue, self._explain_json_corruption()
return True, None, None
def _check_repetition(self, response: str) -> Optional[str]:
"""Check for excessive repetition."""
# Word repetition
if self.response_patterns['word_repetition'].search(response):
return "word_repetition"
# Phrase repetition
if self.response_patterns['phrase_repetition'].search(response):
return "phrase_repetition"
# Calculate repetition ratio
words = response.split()
if len(words) > 10:
unique_words = set(words)
repetition_ratio = 1 - (len(unique_words) / len(words))
if repetition_ratio > self.config.max_repetition_ratio:
return "high_repetition_ratio"
return None
def _check_thinking_loops(self, response: str) -> Optional[str]:
"""Check for thinking loops (common in small models)."""
if self.response_patterns['thinking_loop'].search(response):
return "thinking_loop"
# Check for excessive meta-commentary
thinking_words = ['think', 'considering', 'actually', 'wait', 'hmm', 'let me']
thinking_count = sum(response.lower().count(word) for word in thinking_words)
if thinking_count > 5 and len(response.split()) < 200:
return "excessive_thinking"
return None
def _check_rambling(self, response: str) -> Optional[str]:
"""Check for rambling or excessive filler."""
if self.response_patterns['excessive_filler'].search(response):
return "excessive_filler"
# Check for extremely long sentences (sign of rambling)
sentences = re.split(r'[.!?]+', response)
long_sentences = [s for s in sentences if len(s.split()) > 50]
if len(long_sentences) > 2:
return "excessive_rambling"
return None
def _check_json_corruption(self, response: str) -> Optional[str]:
"""Check for JSON corruption in structured responses."""
if self.response_patterns['broken_json'].search(response):
return "broken_json"
if self.response_patterns['json_repetition'].search(response):
return "json_repetition"
return None
def _explain_too_short(self) -> str:
return """🤔 The AI response was too short to be helpful.
**Why this happens:**
• The model might be confused by the query
• Context might be insufficient
• Model might be overloaded
**What to try:**
• Rephrase your question more specifically
• Try a broader search term first
• Use exploration mode for complex questions: `rag-mini explore`"""
def _explain_timeout(self) -> str:
return """⏱️ The AI took too long to respond (over 60 seconds).
**Why this happens:**
• Small models sometimes get "stuck" thinking
• Complex queries can overwhelm smaller models
• System might be under load
**What to try:**
• Try a simpler, more direct question
• Use synthesis mode for faster responses: `--synthesize`
• Consider using a larger model if available"""
def _explain_repetition(self, issue_type: str) -> str:
return f"""🔄 The AI got stuck in repetition loops ({issue_type}).
**Why this happens:**
• Small models sometimes repeat when uncertain
• Query might be too complex for the model size
• Context window might be exceeded
**What to try:**
• Try a more specific question
• Break complex questions into smaller parts
• Use exploration mode which handles context better: `rag-mini explore`
• Consider: A larger model (qwen3:1.7b or qwen3:3b) would help"""
def _explain_thinking_loop(self) -> str:
return """🧠 The AI got caught in a "thinking loop" - overthinking the response.
**Why this happens:**
• Small models sometimes over-analyze simple questions
• Thinking mode can cause loops in smaller models
• Query complexity exceeds model capabilities
**What to try:**
• Ask more direct, specific questions
• Use synthesis mode (no thinking) for faster results
• Try: "What does this code do?" instead of "Explain how this works"
• Larger models (qwen3:1.7b+) handle thinking better"""
def _explain_rambling(self) -> str:
return """💭 The AI started rambling instead of giving focused answers.
**Why this happens:**
• Small models sometimes lose focus on complex topics
• Query might be too broad or vague
• Model trying to cover too much at once
**What to try:**
• Ask more specific questions
• Break broad questions into focused parts
• Example: "How is data validated?" instead of "Explain the whole system"
• Exploration mode helps maintain focus across questions"""
def _explain_json_corruption(self) -> str:
return """🔧 The AI response format got corrupted.
**Why this happens:**
• Small models sometimes struggle with structured output
• Context limits can cause format errors
• Complex analysis might overwhelm formatting
**What to try:**
• Try the question again (often resolves itself)
• Use simpler questions for better formatting
• Synthesis mode sometimes gives cleaner output
• This is less common with larger models"""
def get_recovery_suggestions(self, issue_type: str, query: str) -> List[str]:
"""Get specific recovery suggestions based on the issue."""
suggestions = []
if issue_type in ['thinking_loop', 'excessive_thinking']:
suggestions.extend([
f"Try synthesis mode: `rag-mini search . \"{query}\" --synthesize`",
"Ask more direct questions without 'why' or 'how'",
"Break complex questions into smaller parts"
])
elif issue_type in ['word_repetition', 'phrase_repetition', 'high_repetition_ratio']:
suggestions.extend([
"Try rephrasing your question completely",
"Use more specific technical terms",
f"Try exploration mode: `rag-mini explore .`"
])
elif issue_type == 'timeout':
suggestions.extend([
"Try a simpler version of your question",
"Use synthesis mode for faster responses",
"Check if Ollama is under heavy load"
])
# Universal suggestions
suggestions.extend([
"Consider using a larger model if available (qwen3:1.7b or qwen3:3b)",
"Check model status: `ollama list`"
])
return suggestions
def get_optimal_ollama_parameters(model_name: str) -> Dict[str, any]:
"""Get optimal parameters for different Ollama models."""
base_params = {
"num_ctx": 32768, # Good context window for most uses
"num_predict": 2000, # Reasonable response length
"temperature": 0.3, # Balanced creativity/consistency
}
# Model-specific optimizations
if "qwen3:0.6b" in model_name.lower():
return {
**base_params,
"repeat_penalty": 1.15, # Prevent repetition in small model
"presence_penalty": 1.5, # Suppress repetitive outputs
"top_p": 0.8, # Focused sampling
"top_k": 20, # Limit choices
"num_predict": 1500, # Shorter responses for reliability
}
elif "qwen3:1.7b" in model_name.lower():
return {
**base_params,
"repeat_penalty": 1.1, # Less aggressive for larger model
"presence_penalty": 1.0, # Balanced
"top_p": 0.9, # More creative
"top_k": 40, # More choices
}
elif any(size in model_name.lower() for size in ["3b", "7b", "8b"]):
return {
**base_params,
"repeat_penalty": 1.05, # Minimal for larger models
"presence_penalty": 0.5, # Light touch
"top_p": 0.95, # High creativity
"top_k": 50, # Many choices
"num_predict": 3000, # Longer responses OK
}
return base_params
# Quick test
def test_safeguards():
"""Test the safeguard system."""
detector = ModelRunawayDetector()
# Test repetition detection
bad_response = "The user authentication system works by checking user credentials. The user authentication system works by checking user credentials. The user authentication system works by checking user credentials."
is_valid, issue, explanation = detector.check_response_quality(bad_response, "auth", time.time())
print(f"Repetition test: Valid={is_valid}, Issue={issue}")
if explanation:
print(explanation)
if __name__ == "__main__":
test_safeguards()