#!/usr/bin/env python3
"""
Bootstrap script to increase RL confidence by adding positive feedback data.
This script simulates successful tool executions with positive ratings to
quickly train the RL system and boost confidence scores.
"""
import os
from datetime import datetime, timedelta
from sqlalchemy import create_engine
from sqlalchemy.orm import sessionmaker
from fccs_agent.config import config
from fccs_agent.services.feedback_service import FeedbackService
from fccs_agent.services.rl_service import init_rl_service, get_rl_service
# Common successful tools and their typical contexts
SUCCESSFUL_TOOL_PATTERNS = [
{
"tool_name": "get_application_info",
"rating": 5,
"execution_time_ms": 150,
"count": 10
},
{
"tool_name": "get_dimensions",
"rating": 5,
"execution_time_ms": 200,
"count": 8
},
{
"tool_name": "get_members",
"rating": 4,
"execution_time_ms": 300,
"count": 8
},
{
"tool_name": "smart_retrieve",
"rating": 5,
"execution_time_ms": 400,
"count": 10
},
{
"tool_name": "get_journals",
"rating": 4,
"execution_time_ms": 350,
"count": 6
},
{
"tool_name": "run_business_rule",
"rating": 5,
"execution_time_ms": 1200,
"count": 5
},
{
"tool_name": "export_data_slice",
"rating": 4,
"execution_time_ms": 800,
"count": 6
},
{
"tool_name": "list_jobs",
"rating": 5,
"execution_time_ms": 180,
"count": 7
},
]
def bootstrap_rl_confidence():
"""Add synthetic positive feedback to boost RL confidence."""
print("š Boosting RL Confidence...\n")
# Initialize services
engine = create_engine(config.database_url)
Session = sessionmaker(bind=engine)
feedback_service = FeedbackService(Session)
rl_service = init_rl_service(
feedback_service=feedback_service,
db_url=config.database_url,
exploration_rate=config.rl_exploration_rate,
learning_rate=config.rl_learning_rate,
discount_factor=config.rl_discount_factor,
min_samples=config.rl_min_samples
)
total_added = 0
base_time = datetime.utcnow() - timedelta(days=7) # Start from 7 days ago
for pattern in SUCCESSFUL_TOOL_PATTERNS:
tool_name = pattern["tool_name"]
rating = pattern["rating"]
exec_time = pattern["execution_time_ms"]
count = pattern["count"]
print(f"š Adding {count} successful executions for '{tool_name}'...")
for i in range(count):
# Vary the execution time slightly
varied_time = exec_time + (i * 10) - 50
# Create synthetic tool execution
session_id = f"bootstrap_session_{total_added + i}"
# Add feedback
feedback_service.record_tool_execution(
session_id=session_id,
tool_name=tool_name,
parameters={},
result={"success": True, "data": "synthetic_data"},
success=True,
execution_time_ms=varied_time,
user_rating=rating,
timestamp=base_time + timedelta(hours=total_added + i)
)
# Calculate reward and update policy
executions = feedback_service.get_session_history(session_id)
if executions:
execution = executions[0]
reward = rl_service.calculate_reward(execution)
# Update RL policy
context_hash = "default" # Use default context for bootstrapping
rl_service.update_policy(
session_id=session_id,
tool_name=tool_name,
context_hash=context_hash,
reward=reward,
next_context_hash=None,
is_terminal=True
)
total_added += count
print(f"\nā
Successfully added {total_added} synthetic tool executions!")
print(f"šÆ RL system has been bootstrapped with positive feedback.\n")
# Display updated statistics
print("š Updated RL Statistics:")
stats = rl_service.get_learning_stats()
if stats:
print(f" ā¢ Policy Updates: {stats.get('update_count', 0)}")
print(f" ā¢ Average Reward: {stats.get('avg_reward', 0):.2f}")
print(f" ā¢ Total Tools Tracked: {stats.get('total_tools', 0)}")
# Show top tool confidences
print("\nš Top Tool Confidences:")
policy_dict = rl_service._get_policy_dict()
# Group by tool name
tool_confidences = {}
for key, action_value in policy_dict.items():
if ":" in key:
tool_name, context = key.split(":", 1)
confidence = rl_service.get_tool_confidence(tool_name, context)
if tool_name not in tool_confidences or confidence > tool_confidences[tool_name]:
tool_confidences[tool_name] = confidence
# Sort and display
sorted_tools = sorted(tool_confidences.items(), key=lambda x: x[1], reverse=True)
for tool_name, confidence in sorted_tools[:10]:
print(f" ā¢ {tool_name:30s} {confidence*100:5.1f}%")
if tool_confidences:
avg_confidence = sum(tool_confidences.values()) / len(tool_confidences) * 100
print(f"\nš Average RL Confidence: {avg_confidence:.1f}%")
print("\nš RL confidence boost complete!")
print("š” Restart your dashboard to see updated confidence scores.\n")
if __name__ == "__main__":
try:
bootstrap_rl_confidence()
except Exception as e:
print(f"ā Error: {e}")
import traceback
traceback.print_exc()
