ml-knowledge-platform/knowledge_platform/feedback/adaptive_prompt.py
2026-02-16 04:50:51 -08:00

268 lines
8.3 KiB
Python

"""Adaptive system prompt builder.
Dynamically adjusts system prompts based on user interaction patterns
to reduce repeated mistakes and improve relevance.
"""
from __future__ import annotations
from pathlib import Path
from typing import Any
from .user_stats import UserStats, UserStatsTracker
class AdaptivePromptBuilder:
"""Builds adaptive system prompts based on user statistics."""
def __init__(
self,
user_id: str,
storage_dir: Path | None = None,
stats_days: int = 30,
) -> None:
"""Initialize adaptive prompt builder.
Args:
user_id: User identifier for personalization
storage_dir: Feedback storage directory (default: ~/.cache/crystal/feedback)
stats_days: Number of days to analyze for stats
"""
self.user_id = user_id
self.stats_days = stats_days
if storage_dir is None:
storage_dir = Path.home() / ".cache" / "crystal" / "feedback"
self.tracker = UserStatsTracker(storage_dir)
self._stats: UserStats | None = None
@property
def stats(self) -> UserStats:
"""Get or compute user stats (cached)."""
if self._stats is None:
self._stats = self.tracker.get_user_stats(
self.user_id, days=self.stats_days
)
return self._stats
def build(
self,
base_prompt: str,
context: dict[str, Any] | None = None,
) -> str:
"""Build adaptive prompt based on user stats.
Args:
base_prompt: Original system prompt
context: Optional context dict with:
- recent_low_confidence: bool
- current_topic: str
- session_corrections: int
Returns:
Enhanced system prompt with user-specific adaptations
"""
context = context or {}
sections = [base_prompt]
# Add frequent correction reminders
correction_section = self._build_correction_reminders()
if correction_section:
sections.append(correction_section)
# Add low-confidence topic disclaimers
confidence_section = self._build_confidence_disclaimers(context)
if confidence_section:
sections.append(confidence_section)
# Add domain focus hints
domain_section = self._build_domain_focus()
if domain_section:
sections.append(domain_section)
# Add error prevention hints
error_section = self._build_error_prevention()
if error_section:
sections.append(error_section)
return "\n\n".join(sections)
def _build_correction_reminders(self) -> str | None:
"""Build reminder section for frequent corrections."""
if not self.stats.corrections:
return None
patterns = self.stats.corrections.frequent_patterns
if not patterns:
return None
# Top 3 most frequent corrections
top_patterns = patterns[:3]
lines = ["## Frequent Corrections", "", "You often correct these patterns:"]
for original, replacement, count in top_patterns:
lines.append(f"- **Avoid:** \"{original}\" → **Use:** \"{replacement}\" (corrected {count}x)")
lines.append("")
lines.append("Please be extra careful with these terms to reduce repeated corrections.")
return "\n".join(lines)
def _build_confidence_disclaimers(
self, context: dict[str, Any]
) -> str | None:
"""Build disclaimer for low-confidence topics."""
if not self.stats.topics:
return None
low_conf_topics = self.stats.topics.low_confidence_topics
if not low_conf_topics:
return None
# Check if current context involves a low-confidence topic
current_topic = context.get("current_topic", "").lower()
is_low_confidence = any(
topic in current_topic for topic, _ in low_conf_topics
)
if not is_low_confidence and not context.get("recent_low_confidence"):
return None
lines = [
"## Confidence Notice",
"",
"Based on recent interactions, confidence may be lower for these topics:",
]
for topic, conf in low_conf_topics[:3]:
conf_pct = int(conf * 100)
lines.append(f"- {topic} (avg confidence: {conf_pct}%)")
lines.append("")
lines.append(
"Consider requesting additional context or verification for these areas."
)
return "\n".join(lines)
def _build_domain_focus(self) -> str | None:
"""Build domain-specific focus hints."""
if not self.stats.topics:
return None
primary_topics = self.stats.topics.primary_topics
if not primary_topics:
return None
# Check if user has strong focus on specific domains
total_interactions = sum(count for _, count in primary_topics)
if not total_interactions:
return None
# Calculate focus percentage for top topic
top_topic, top_count = primary_topics[0]
focus_pct = (top_count / total_interactions) * 100
# Only add hint if >60% focused on one topic
if focus_pct < 60:
return None
lines = [
"## Domain Focus",
"",
f"Note: This user primarily works with **{top_topic}** topics ({int(focus_pct)}% of interactions).",
"Prioritize information relevant to this domain when appropriate.",
]
return "\n".join(lines)
def _build_error_prevention(self) -> str | None:
"""Build error prevention hints based on common mistakes."""
if not self.stats.corrections:
return None
error_types = self.stats.corrections.common_error_types
if not error_types:
return None
# Only show if there are significant error patterns
total_errors = sum(count for _, count in error_types)
if total_errors < 5:
return None
lines = [
"## Error Prevention",
"",
"Common error types in corrections:",
]
for error_type, count in error_types[:3]:
pct = int((count / total_errors) * 100)
lines.append(f"- {error_type}: {pct}% of corrections")
lines.append("")
lines.append("Double-check content for these error types before responding.")
return "\n".join(lines)
def refresh_stats(self) -> None:
"""Force refresh of user stats from storage."""
self.tracker.clear_cache(self.user_id)
self._stats = None
def get_stats_summary(self) -> dict[str, Any]:
"""Get summary of user stats for debugging/monitoring.
Returns:
Dict with user statistics summary
"""
stats = self.stats
summary = {
"user_id": stats.user_id,
"period": {
"start": stats.period_start,
"end": stats.period_end,
},
"total_interactions": stats.total_interactions,
}
if stats.corrections:
summary["corrections"] = {
"total": stats.corrections.total_corrections,
"avg_confidence": round(stats.corrections.avg_confidence, 2),
"frequent_patterns_count": len(stats.corrections.frequent_patterns),
"top_error_types": [
type_name for type_name, _ in stats.corrections.common_error_types[:3]
],
}
if stats.topics:
summary["topics"] = {
"primary": [topic for topic, _ in stats.topics.primary_topics[:3]],
"low_confidence_count": len(stats.topics.low_confidence_topics),
}
return summary
def build_adaptive_prompt(
base_prompt: str,
user_id: str,
context: dict[str, Any] | None = None,
storage_dir: Path | None = None,
) -> str:
"""Convenience function to build adaptive prompt in one call.
Args:
base_prompt: Original system prompt
user_id: User identifier
context: Optional context dict
storage_dir: Feedback storage directory
Returns:
Enhanced system prompt
"""
builder = AdaptivePromptBuilder(user_id, storage_dir)
return builder.build(base_prompt, context)