codebook / potato /solo_mode /prompt_optimizer.py
davidjurgens's picture
Deploy: Potato — Codebook Annotation
aceb1b2 verified
Raw
History Blame Contribute Delete
19 kB
"""
Prompt Optimizer for Solo Mode
This module implements DSPy-style automatic prompt optimization.
It uses labeled examples to iteratively improve prompts for better
accuracy while maintaining brevity.
"""
import json
import logging
import re
import threading
import time
from dataclasses import dataclass, field
from datetime import datetime
from typing import Any, Dict, List, Optional, Tuple
from queue import Queue
logger = logging.getLogger(__name__)
OPTIMIZATION_PROMPT_TEMPLATE = """You are an expert at improving annotation prompts.
Given the current prompt and some examples where the LLM made mistakes,
suggest improvements to the prompt that would help get the correct labels.
## Current Prompt
{current_prompt}
## Correct Examples
These are examples where the LLM got the right answer:
{correct_examples}
## Incorrect Examples
These are examples where the LLM got the wrong answer (with corrections):
{incorrect_examples}
## Optimization Goals
1. Improve accuracy on the incorrect examples
2. Keep the prompt concise (shorter is better)
3. Make instructions clearer and more specific
4. Add clarifying examples if helpful
## Requirements
- Focus on patterns in the errors
- Don't make the prompt too long
- Keep successful patterns from the current prompt
- Be specific about edge cases
## Output Format
Respond with JSON:
{{
"improved_prompt": "<the improved prompt text>",
"changes_made": ["<change 1>", "<change 2>", ...],
"rationale": "<why these changes should help>"
}}
"""
@dataclass
class OptimizationResult:
"""Result of a prompt optimization run."""
original_prompt: str
optimized_prompt: str
changes_made: List[str]
rationale: str
accuracy_before: float
accuracy_after: Optional[float] = None
timestamp: datetime = field(default_factory=datetime.now)
model_used: str = ""
num_examples_used: int = 0
def to_dict(self) -> Dict[str, Any]:
"""Serialize to dictionary."""
return {
'original_prompt': self.original_prompt,
'optimized_prompt': self.optimized_prompt,
'changes_made': self.changes_made,
'rationale': self.rationale,
'accuracy_before': self.accuracy_before,
'accuracy_after': self.accuracy_after,
'timestamp': self.timestamp.isoformat(),
'model_used': self.model_used,
'num_examples_used': self.num_examples_used,
}
@dataclass
class OptimizationConfig:
"""Configuration for prompt optimization."""
enabled: bool = True
find_smallest_model: bool = True
target_accuracy: float = 0.85
min_examples_for_optimization: int = 10
optimization_interval_seconds: int = 300 # 5 minutes
max_prompt_length: int = 2000
accuracy_weight: float = 0.7
length_weight: float = 0.2
consistency_weight: float = 0.1
class PromptOptimizer:
"""
DSPy-style automatic prompt optimization.
Optimizes prompts based on:
1. Accuracy on labeled examples
2. Prompt length (shorter is better)
3. Prediction consistency
Can run in background or be triggered on-demand.
"""
def __init__(
self,
config: Dict[str, Any],
solo_config: Any,
prompt_getter: callable,
prompt_setter: callable,
examples_getter: callable,
):
"""
Initialize the prompt optimizer.
Args:
config: Full application configuration
solo_config: SoloModeConfig instance
prompt_getter: Callable that returns current prompt text
prompt_setter: Callable to update the prompt
examples_getter: Callable that returns labeled examples
"""
self.config = config
self.solo_config = solo_config
self.prompt_getter = prompt_getter
self.prompt_setter = prompt_setter
self.examples_getter = examples_getter
# Load optimization config. solo_config.prompt_optimization may be a
# PromptOptimizationConfig dataclass (typical) or a plain dict; read
# fields from either form.
opt_config = getattr(solo_config, 'prompt_optimization', None)
if opt_config:
def _opt(key, default):
if isinstance(opt_config, dict):
return opt_config.get(key, default)
return getattr(opt_config, key, default)
self.opt_config = OptimizationConfig(
enabled=_opt('enabled', True),
find_smallest_model=_opt('find_smallest_model', True),
target_accuracy=_opt('target_accuracy', 0.85),
)
else:
self.opt_config = OptimizationConfig()
self._lock = threading.RLock()
# Optimization history
self.optimization_history: List[OptimizationResult] = []
# Background optimization
self._background_thread: Optional[threading.Thread] = None
self._stop_event = threading.Event()
self._optimization_queue: Queue = Queue()
# AI endpoint (lazy init)
self._endpoint = None
# Cached labeled examples
self._cached_examples: Dict[str, Dict[str, Any]] = {}
def _get_endpoint(self) -> Optional[Any]:
"""Get or create the optimization endpoint."""
if self._endpoint is not None:
return self._endpoint
if not self.solo_config.revision_models:
logger.warning("No revision models configured for optimization")
return None
try:
from potato.ai.ai_endpoint import AIEndpointFactory
for model_config in self.solo_config.revision_models:
try:
endpoint_config = model_config.to_endpoint_config(temperature_override=0.3)
endpoint = AIEndpointFactory.create_endpoint(endpoint_config)
if endpoint:
self._endpoint = endpoint
return endpoint
except Exception as e:
logger.debug(f"Failed to create optimization endpoint: {e}")
continue
except Exception as e:
logger.error(f"Error creating optimization endpoint: {e}")
return None
def optimize(self, force: bool = False) -> Optional[OptimizationResult]:
"""
Run prompt optimization.
Args:
force: Run even if not enough examples
Returns:
OptimizationResult if optimization was performed
"""
with self._lock:
# Get labeled examples
examples = self.examples_getter()
if not examples and not force:
logger.info("No labeled examples available for optimization")
return None
if len(examples) < self.opt_config.min_examples_for_optimization and not force:
logger.info(
f"Not enough examples for optimization "
f"({len(examples)} < {self.opt_config.min_examples_for_optimization})"
)
return None
# Get current prompt
current_prompt = self.prompt_getter()
if not current_prompt:
logger.warning("No current prompt to optimize")
return None
# Get endpoint
endpoint = self._get_endpoint()
if endpoint is None:
logger.warning("No endpoint available for optimization")
return None
# Split examples into correct and incorrect
correct, incorrect = self._split_examples(examples)
if not incorrect:
logger.info("No incorrect predictions to optimize for")
return None
# Calculate current accuracy
accuracy_before = len(correct) / len(examples) if examples else 0.0
# Check if already above target
if accuracy_before >= self.opt_config.target_accuracy and not force:
logger.info(
f"Accuracy ({accuracy_before:.2%}) already above target "
f"({self.opt_config.target_accuracy:.2%})"
)
return None
# Generate optimized prompt
result = self._generate_optimized_prompt(
current_prompt,
correct[:5], # Limit examples
incorrect[:10],
endpoint,
accuracy_before,
)
if result:
self.optimization_history.append(result)
# Update prompt if optimization was successful
if result.optimized_prompt and result.optimized_prompt != current_prompt:
self.prompt_setter(
result.optimized_prompt,
source='llm_optimization',
source_description='; '.join(result.changes_made)
)
logger.info(f"Prompt optimized: {len(result.changes_made)} changes made")
return result
def _split_examples(
self,
examples: List[Dict[str, Any]]
) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
"""Split examples into correct and incorrect predictions."""
correct = []
incorrect = []
for ex in examples:
if ex.get('agrees', True):
correct.append(ex)
else:
incorrect.append(ex)
return correct, incorrect
def _generate_optimized_prompt(
self,
current_prompt: str,
correct_examples: List[Dict[str, Any]],
incorrect_examples: List[Dict[str, Any]],
endpoint: Any,
accuracy_before: float,
) -> Optional[OptimizationResult]:
"""Generate an optimized prompt using the LLM."""
try:
# Format examples
correct_text = self._format_examples(correct_examples, show_correction=False)
incorrect_text = self._format_examples(incorrect_examples, show_correction=True)
optimization_prompt = OPTIMIZATION_PROMPT_TEMPLATE.format(
current_prompt=current_prompt,
correct_examples=correct_text or "None available",
incorrect_examples=incorrect_text or "None available",
)
from pydantic import BaseModel
class OptimizationResponse(BaseModel):
improved_prompt: str = ""
changes_made: List[str] = []
rationale: str = ""
response = endpoint.query(optimization_prompt, OptimizationResponse)
# Parse response
if isinstance(response, str):
response_data = self._parse_json_response(response)
elif hasattr(response, 'model_dump'):
response_data = response.model_dump()
else:
response_data = response
improved_prompt = response_data.get('improved_prompt', '')
changes_made = response_data.get('changes_made', [])
rationale = response_data.get('rationale', '')
# Validate improved prompt
if not improved_prompt:
logger.warning("Optimization returned empty prompt")
return None
if len(improved_prompt) > self.opt_config.max_prompt_length:
logger.warning(
f"Optimized prompt too long ({len(improved_prompt)} > {self.opt_config.max_prompt_length})"
)
# Truncate if necessary
improved_prompt = improved_prompt[:self.opt_config.max_prompt_length]
return OptimizationResult(
original_prompt=current_prompt,
optimized_prompt=improved_prompt,
changes_made=changes_made,
rationale=rationale,
accuracy_before=accuracy_before,
model_used=getattr(endpoint, 'model', ''),
num_examples_used=len(correct_examples) + len(incorrect_examples),
)
except Exception as e:
logger.error(f"Error generating optimized prompt: {e}")
return None
def _format_examples(
self,
examples: List[Dict[str, Any]],
show_correction: bool = False
) -> str:
"""Format examples for the optimization prompt."""
if not examples:
return ""
formatted = []
for i, ex in enumerate(examples[:10], 1):
text = ex.get('text', '')[:200] # Truncate long text
predicted = ex.get('predicted_label', '')
if show_correction:
actual = ex.get('actual_label', ex.get('human_label', ''))
formatted.append(
f"{i}. Text: \"{text}\"\n"
f" LLM predicted: {predicted}\n"
f" Correct label: {actual}"
)
else:
formatted.append(
f"{i}. Text: \"{text}\"\n"
f" Label: {predicted}"
)
return '\n\n'.join(formatted)
def _parse_json_response(self, response: str) -> Dict[str, Any]:
"""Parse JSON from response."""
content = response.strip()
if '```json' in content:
match = re.search(r'```json\s*([\s\S]*?)\s*```', content)
if match:
content = match.group(1).strip()
elif '```' in content:
match = re.search(r'```\s*([\s\S]*?)\s*```', content)
if match:
content = match.group(1).strip()
try:
return json.loads(content)
except json.JSONDecodeError:
return {}
# === Background Optimization ===
def start_background_optimization(self) -> bool:
"""Start background optimization thread."""
with self._lock:
if self._background_thread is not None and self._background_thread.is_alive():
logger.warning("Background optimization already running")
return False
if not self.opt_config.enabled:
logger.info("Prompt optimization is disabled")
return False
self._stop_event.clear()
self._background_thread = threading.Thread(
target=self._background_optimization_loop,
name="PromptOptimizationThread",
daemon=True
)
self._background_thread.start()
logger.info("Started background prompt optimization")
return True
def stop_background_optimization(self) -> None:
"""Stop background optimization thread."""
if self._background_thread is None:
return
self._stop_event.set()
self._background_thread.join(timeout=5.0)
self._background_thread = None
logger.info("Stopped background prompt optimization")
def is_running(self) -> bool:
"""Check if background optimization is running."""
return (
self._background_thread is not None and
self._background_thread.is_alive()
)
def _background_optimization_loop(self) -> None:
"""Main loop for background optimization."""
interval = self.opt_config.optimization_interval_seconds
logger.info(f"Background optimization started (interval={interval}s)")
while not self._stop_event.is_set():
try:
# Wait for interval
if self._stop_event.wait(timeout=interval):
break # Stop event was set
# Run optimization
result = self.optimize()
if result:
logger.info(
f"Background optimization completed: "
f"accuracy {result.accuracy_before:.2%} -> {result.accuracy_after or 'pending'}"
)
except Exception as e:
logger.error(f"Error in background optimization: {e}")
# === Model Selection ===
def find_smallest_accurate_model(
self,
models: List[Any],
test_examples: List[Dict[str, Any]],
prompt: str,
) -> Optional[str]:
"""
Find the smallest model that achieves target accuracy.
Args:
models: List of model configs (ordered small to large)
test_examples: Examples to test accuracy on
prompt: The prompt to use
Returns:
Model name if found, None otherwise
"""
if not self.opt_config.find_smallest_model:
return None
target = self.opt_config.target_accuracy
for model_config in models:
try:
accuracy = self._test_model_accuracy(
model_config, test_examples, prompt
)
if accuracy >= target:
logger.info(
f"Model {model_config.model} achieves {accuracy:.2%} accuracy"
)
return model_config.model
except Exception as e:
logger.debug(f"Error testing model {model_config.model}: {e}")
continue
logger.warning("No model achieved target accuracy")
return None
def _test_model_accuracy(
self,
model_config: Any,
examples: List[Dict[str, Any]],
prompt: str,
) -> float:
"""Test a model's accuracy on examples."""
# This is a placeholder - full implementation would
# run predictions with the model and calculate accuracy
return 0.0
# === Status and History ===
def get_optimization_history(self) -> List[OptimizationResult]:
"""Get optimization history."""
with self._lock:
return self.optimization_history.copy()
def get_last_optimization(self) -> Optional[OptimizationResult]:
"""Get the most recent optimization result."""
with self._lock:
if self.optimization_history:
return self.optimization_history[-1]
return None
def get_status(self) -> Dict[str, Any]:
"""Get optimizer status."""
with self._lock:
last = self.get_last_optimization()
return {
'enabled': self.opt_config.enabled,
'is_running': self.is_running(),
'optimization_count': len(self.optimization_history),
'last_optimization': last.to_dict() if last else None,
'target_accuracy': self.opt_config.target_accuracy,
'interval_seconds': self.opt_config.optimization_interval_seconds,
}
def clear_history(self) -> None:
"""Clear optimization history."""
with self._lock:
self.optimization_history.clear()