Major changes: - Remove fake OPRO evaluation (no more fake 0.5 scores) - Add simple refinement based on user selection - New endpoint: POST /opro/refine (selected + rejected instructions) - Update prompt generation to focus on comprehensive coverage instead of style variety - All generated instructions now start with role definition (你是一个...) - Update README to reflect new approach and API endpoints Technical details: - Added refine_based_on_selection() in prompt_utils.py - Added refine_instruction_candidates() in user_prompt_optimizer.py - Added OPRORefineReq model and /opro/refine endpoint in api.py - Updated frontend handleContinueOptimize() to use new refinement flow - Changed prompt requirements from 'different styles' to 'comprehensive coverage' - Added role definition requirement as first item in all prompt templates
189 lines
6.6 KiB
Python
189 lines
6.6 KiB
Python
from typing import List, Tuple
|
||
|
||
# ============================================================================
|
||
# OLD FUNCTIONS (Query Rewriting - NOT true OPRO, kept for compatibility)
|
||
# ============================================================================
|
||
|
||
def refine_instruction(query: str) -> str:
|
||
"""
|
||
LEGACY: Generates query rewrites (NOT true OPRO).
|
||
This is query expansion, not system instruction optimization.
|
||
"""
|
||
return f"""
|
||
你是一个“问题澄清与重写助手”。
|
||
请根据用户的原始问题:
|
||
【{query}】
|
||
生成不少于20条多角度、可直接执行的问题改写,每行一条。
|
||
"""
|
||
|
||
def refine_instruction_with_history(query: str, rejected_list: list) -> str:
|
||
"""
|
||
LEGACY: Generates query rewrites with rejection history (NOT true OPRO).
|
||
"""
|
||
rejected_text = "\n".join(f"- {r}" for r in rejected_list) if rejected_list else ""
|
||
return f"""
|
||
你是一个“问题澄清与重写助手”。
|
||
原始问题:
|
||
{query}
|
||
|
||
以下改写已被否定:
|
||
{rejected_text}
|
||
|
||
请从新的角度重新生成至少20条不同的改写问题,每条单独一行。
|
||
"""
|
||
|
||
|
||
# ============================================================================
|
||
# TRUE OPRO FUNCTIONS (System Instruction Optimization)
|
||
# ============================================================================
|
||
|
||
def generate_initial_system_instruction_candidates(task_description: str, pool_size: int = None) -> str:
|
||
"""
|
||
TRUE OPRO: Generates initial candidate System Instructions for a new OPRO run.
|
||
|
||
Args:
|
||
task_description: Description of the task the LLM should perform
|
||
pool_size: Number of candidates to generate (defaults to config.GENERATION_POOL_SIZE)
|
||
|
||
Returns:
|
||
Meta-prompt that instructs the optimizer LLM to generate system instruction candidates
|
||
"""
|
||
import config
|
||
pool_size = pool_size or config.GENERATION_POOL_SIZE
|
||
|
||
return f"""
|
||
你是一个"系统指令生成助手"。
|
||
目标任务描述:
|
||
【{task_description}】
|
||
|
||
请根据以上任务,生成 {pool_size} 条高质量、全面的"System Instruction"候选指令。
|
||
|
||
要求:
|
||
1. 每条指令必须以角色定义开头(例如:"你是一个..."、"你是..."等)
|
||
2. 每条指令必须全面覆盖任务的所有要求和细节
|
||
3. 指令应清晰、具体、可执行,能够有效指导LLM完成任务
|
||
4. 确保指令包含必要的行为规范、输出格式、注意事项等
|
||
5. 每条指令单独成行,不包含编号或额外说明
|
||
6. 所有生成的指令必须使用简体中文
|
||
|
||
生成 {pool_size} 条指令:
|
||
"""
|
||
|
||
|
||
def generate_optimized_system_instruction(
|
||
task_description: str,
|
||
trajectory: List[Tuple[str, float]],
|
||
pool_size: int = None
|
||
) -> str:
|
||
"""
|
||
TRUE OPRO: Analyzes performance trajectory and generates optimized System Instructions.
|
||
|
||
This is the core OPRO function that uses an LLM as an optimizer to improve
|
||
system instructions based on historical performance scores.
|
||
|
||
Args:
|
||
task_description: Description of the task the LLM should perform
|
||
trajectory: List of (instruction, score) tuples, sorted by score (highest first)
|
||
pool_size: Number of candidates to generate (defaults to config.GENERATION_POOL_SIZE)
|
||
|
||
Returns:
|
||
Meta-prompt that instructs the optimizer LLM to generate better system instructions
|
||
"""
|
||
import config
|
||
pool_size = pool_size or config.GENERATION_POOL_SIZE
|
||
|
||
if not trajectory:
|
||
# If no trajectory, fall back to initial generation
|
||
return generate_initial_system_instruction_candidates(task_description, pool_size)
|
||
|
||
# Format the trajectory for the Optimizer LLM
|
||
formatted_history = "\n".join(
|
||
f"--- Instruction Score: {score:.4f}\n{instruction}"
|
||
for instruction, score in trajectory
|
||
)
|
||
|
||
# Determine the current highest score to set the optimization goal
|
||
highest_score = max(score for _, score in trajectory)
|
||
|
||
# Construct the Meta-Prompt (The OPRO Instruction)
|
||
return f"""
|
||
你是一个"System Prompt 优化器"。
|
||
你的任务是改进一个LLM的系统指令,以最大化其在以下任务中的性能:
|
||
【{task_description}】
|
||
|
||
---
|
||
**历史性能轨迹 (Instructions and Scores):**
|
||
{formatted_history}
|
||
---
|
||
**当前最高得分: {highest_score:.4f}**
|
||
|
||
请分析得分最高的指令的特点和得分最低指令的缺陷。
|
||
然后,生成 {pool_size} 条新的、有潜力超越 {highest_score:.4f} 分的System Instruction。
|
||
|
||
要求:
|
||
1. 每条指令必须以角色定义开头(例如:"你是一个..."、"你是..."等)
|
||
2. 每条指令必须全面覆盖任务的所有要求和细节
|
||
3. 结合高分指令的优点,避免低分指令的缺陷
|
||
4. 指令应清晰、具体、可执行,能够有效指导LLM完成任务
|
||
5. 每条指令单独成行,不包含编号或额外说明
|
||
6. 所有生成的指令必须使用简体中文
|
||
|
||
生成 {pool_size} 条优化后的指令:
|
||
"""
|
||
|
||
|
||
def refine_based_on_selection(
|
||
task_description: str,
|
||
selected_instruction: str,
|
||
rejected_instructions: List[str],
|
||
pool_size: int = None
|
||
) -> str:
|
||
"""
|
||
Simple refinement: Generate variations based on selected instruction while avoiding rejected ones.
|
||
|
||
This is NOT OPRO - it's straightforward iterative refinement based on user preference.
|
||
No scoring, no trajectory, just: "I like this one, give me more like it (but not like those)."
|
||
|
||
Args:
|
||
task_description: Description of the task
|
||
selected_instruction: The instruction the user selected
|
||
rejected_instructions: The instructions the user didn't select
|
||
pool_size: Number of new candidates to generate
|
||
|
||
Returns:
|
||
Prompt for generating refined candidates
|
||
"""
|
||
import config
|
||
pool_size = pool_size or config.GENERATION_POOL_SIZE
|
||
|
||
rejected_text = ""
|
||
if rejected_instructions:
|
||
rejected_formatted = "\n".join(f"- {inst}" for inst in rejected_instructions)
|
||
rejected_text = f"""
|
||
**用户未选择的指令(避免这些方向):**
|
||
{rejected_formatted}
|
||
"""
|
||
|
||
return f"""
|
||
你是一个"System Prompt 改进助手"。
|
||
目标任务描述:
|
||
【{task_description}】
|
||
|
||
**用户选择的指令(基于此改进):**
|
||
{selected_instruction}
|
||
{rejected_text}
|
||
|
||
请基于用户选择的指令,生成 {pool_size} 条改进版本。
|
||
|
||
要求:
|
||
1. 每条指令必须以角色定义开头(例如:"你是一个..."、"你是..."等)
|
||
2. 保留用户选择指令的核心优点
|
||
3. 每条指令必须全面覆盖任务的所有要求和细节
|
||
4. 指令应清晰、具体、可执行,能够有效指导LLM完成任务
|
||
5. 避免与未选择指令相似的方向
|
||
6. 每条指令单独成行,不包含编号或额外说明
|
||
7. 所有生成的指令必须使用简体中文
|
||
|
||
生成 {pool_size} 条改进后的指令:
|
||
"""
|