|
|
@@ -98,6 +98,10 @@ class RunContext(BaseModel):
|
|
|
# 最终结果
|
|
|
final_output: str | None = None
|
|
|
|
|
|
+ # 评估缓存：避免重复评估相同文本
|
|
|
+ evaluation_cache: dict[str, tuple[float, str]] = Field(default_factory=dict)
|
|
|
+ # key: 文本, value: (score, reason)
|
|
|
+
|
|
|
|
|
|
# ============================================================================
|
|
|
# Agent 定义
|
|
|
@@ -441,6 +445,13 @@ word_selection_instructions = """
|
|
|
* combined_query: 组合后的新query（只包含seed和word的原始文本）
|
|
|
* reasoning: 选择理由（说明为什么选这个词）
|
|
|
- overall_reasoning: 整体选择思路（说明这5个词的选择逻辑）
|
|
|
+
|
|
|
+## JSON输出规范
|
|
|
+1. **格式要求**：必须输出标准的、完整的JSON格式
|
|
|
+2. **字符限制**：不要在JSON中使用任何不可见的特殊字符或控制字符
|
|
|
+3. **引号规范**：字符串中如需表达引用或强调，使用书名号《》或单书名号「」，不要使用英文引号或中文引号""
|
|
|
+4. **编码规范**：所有文本使用UTF-8编码，不要包含二进制或转义序列
|
|
|
+5. **完整性**：确保JSON的开始和结束括号完整匹配，所有字段都正确闭合
|
|
|
""".strip()
|
|
|
|
|
|
word_selector = Agent[None](
|
|
|
@@ -492,6 +503,13 @@ def calculate_final_score(motivation_score: float, category_score: float) -> flo
|
|
|
return base_score
|
|
|
|
|
|
|
|
|
+def clean_json_string(text: str) -> str:
|
|
|
+ """清理JSON中的非法控制字符（保留 \t \n \r）"""
|
|
|
+ import re
|
|
|
+ # 移除除了 \t(09) \n(0A) \r(0D) 之外的所有控制字符
|
|
|
+ return re.sub(r'[\x00-\x08\x0B\x0C\x0E-\x1F]', '', text)
|
|
|
+
|
|
|
+
|
|
|
def process_note_data(note: dict) -> Post:
|
|
|
"""处理搜索接口返回的帖子数据"""
|
|
|
note_card = note.get("note_card", {})
|
|
|
@@ -555,7 +573,7 @@ def process_note_data(note: dict) -> Post:
|
|
|
)
|
|
|
|
|
|
|
|
|
-async def evaluate_with_o(text: str, o: str) -> tuple[float, str]:
|
|
|
+async def evaluate_with_o(text: str, o: str, cache: dict[str, tuple[float, str]] | None = None) -> tuple[float, str]:
|
|
|
"""评估文本与原始问题o的相关度
|
|
|
|
|
|
采用两阶段评估 + 代码计算规则：
|
|
|
@@ -563,9 +581,20 @@ async def evaluate_with_o(text: str, o: str) -> tuple[float, str]:
|
|
|
2. 品类维度评估（权重30%）
|
|
|
3. 应用规则A/B/C调整得分
|
|
|
|
|
|
+ Args:
|
|
|
+ text: 待评估的文本
|
|
|
+ o: 原始问题
|
|
|
+ cache: 评估缓存（可选），用于避免重复评估
|
|
|
+
|
|
|
Returns:
|
|
|
tuple[float, str]: (最终相关度分数, 综合评估理由)
|
|
|
"""
|
|
|
+ # 检查缓存
|
|
|
+ if cache is not None and text in cache:
|
|
|
+ cached_score, cached_reason = cache[text]
|
|
|
+ print(f" ⚡ 缓存命中: {text} -> {cached_score:.2f}")
|
|
|
+ return cached_score, cached_reason
|
|
|
+
|
|
|
# 准备输入
|
|
|
eval_input = f"""
|
|
|
<原始问题>
|
|
|
@@ -630,6 +659,10 @@ async def evaluate_with_o(text: str, o: str) -> tuple[float, str]:
|
|
|
elif motivation_score <= 0.2:
|
|
|
combined_reason += "（应用规则B：动机低分限制机制）"
|
|
|
|
|
|
+ # 存入缓存
|
|
|
+ if cache is not None:
|
|
|
+ cache[text] = (final_score, combined_reason)
|
|
|
+
|
|
|
return final_score, combined_reason
|
|
|
|
|
|
except Exception as e:
|
|
|
@@ -684,7 +717,7 @@ async def initialize(o: str, context: RunContext) -> tuple[list[Seg], list[Word]
|
|
|
|
|
|
async def evaluate_seg(seg: Seg) -> Seg:
|
|
|
async with seg_semaphore:
|
|
|
- seg.score_with_o, seg.reason = await evaluate_with_o(seg.text, o)
|
|
|
+ seg.score_with_o, seg.reason = await evaluate_with_o(seg.text, o, context.evaluation_cache)
|
|
|
return seg
|
|
|
|
|
|
if seg_list:
|
|
|
@@ -813,7 +846,7 @@ async def run_round(
|
|
|
|
|
|
async def evaluate_sug(sug: Sug) -> Sug:
|
|
|
async with semaphore: # 限制并发数
|
|
|
- sug.score_with_o, sug.reason = await evaluate_with_o(sug.text, o)
|
|
|
+ sug.score_with_o, sug.reason = await evaluate_with_o(sug.text, o, context.evaluation_cache)
|
|
|
return sug
|
|
|
|
|
|
if all_sugs:
|
|
|
@@ -887,6 +920,7 @@ async def run_round(
|
|
|
# 4. 构建q_list_next
|
|
|
print(f"\n[步骤4] 构建q_list_next...")
|
|
|
q_list_next = []
|
|
|
+ existing_q_texts = set() # 用于去重
|
|
|
add_word_details = {} # 保存每个seed对应的组合词列表
|
|
|
all_seed_combinations = [] # 保存本轮所有seed的组合词（用于后续构建seed_list_next）
|
|
|
|
|
|
@@ -912,7 +946,7 @@ async def run_round(
|
|
|
|
|
|
print(f" 候选词数量: {len(candidate_words)}")
|
|
|
|
|
|
- # 调用Agent一次性选择并组合Top 5
|
|
|
+ # 调用Agent一次性选择并组合Top 5（添加重试机制）
|
|
|
candidate_words_text = ', '.join([w.text for w in candidate_words])
|
|
|
selection_input = f"""
|
|
|
<原始问题>
|
|
|
@@ -929,15 +963,34 @@ async def run_round(
|
|
|
|
|
|
请从候选词列表中选择最多5个最合适的词，分别与当前seed组合成新的query。
|
|
|
"""
|
|
|
- result = await Runner.run(word_selector, selection_input)
|
|
|
- selection_result: WordSelectionTop5 = result.final_output
|
|
|
+
|
|
|
+ # 重试机制
|
|
|
+ max_retries = 2
|
|
|
+ selection_result = None
|
|
|
+ for attempt in range(max_retries):
|
|
|
+ try:
|
|
|
+ result = await Runner.run(word_selector, selection_input)
|
|
|
+ selection_result = result.final_output
|
|
|
+ break # 成功则跳出
|
|
|
+ except Exception as e:
|
|
|
+ error_msg = str(e)
|
|
|
+ if attempt < max_retries - 1:
|
|
|
+ print(f" ⚠️ 选词失败 (尝试 {attempt+1}/{max_retries}): {error_msg[:100]}")
|
|
|
+ await asyncio.sleep(1)
|
|
|
+ else:
|
|
|
+ print(f" ❌ 选词失败，跳过该seed: {error_msg[:100]}")
|
|
|
+ break
|
|
|
+
|
|
|
+ if selection_result is None:
|
|
|
+ print(f" 跳过seed: {seed.text}")
|
|
|
+ continue
|
|
|
|
|
|
print(f" Agent选择了 {len(selection_result.combinations)} 个组合")
|
|
|
print(f" 整体选择思路: {selection_result.overall_reasoning}")
|
|
|
|
|
|
# 并发评估所有组合的相关度
|
|
|
async def evaluate_combination(comb: WordCombination) -> dict:
|
|
|
- score, reason = await evaluate_with_o(comb.combined_query, o)
|
|
|
+ score, reason = await evaluate_with_o(comb.combined_query, o, context.evaluation_cache)
|
|
|
return {
|
|
|
'word': comb.selected_word,
|
|
|
'query': comb.combined_query,
|
|
|
@@ -951,8 +1004,13 @@ async def run_round(
|
|
|
|
|
|
print(f" 评估完成，得到 {len(top_5)} 个组合")
|
|
|
|
|
|
- # 将Top 5全部加入q_list_next
|
|
|
+ # 将Top 5全部加入q_list_next（去重检查）
|
|
|
for comb in top_5:
|
|
|
+ # 去重检查
|
|
|
+ if comb['query'] in existing_q_texts:
|
|
|
+ print(f" ⊗ 跳过重复: {comb['query']}")
|
|
|
+ continue
|
|
|
+
|
|
|
print(f" ✓ {comb['query']} (分数: {comb['score']:.2f})")
|
|
|
|
|
|
new_q = Q(
|
|
|
@@ -962,6 +1020,7 @@ async def run_round(
|
|
|
from_source="add"
|
|
|
)
|
|
|
q_list_next.append(new_q)
|
|
|
+ existing_q_texts.add(comb['query']) # 记录到去重集合
|
|
|
|
|
|
# 记录已添加的词
|
|
|
seed.added_words.append(comb['word'])
|
|
|
@@ -980,10 +1039,15 @@ async def run_round(
|
|
|
# 保存到all_seed_combinations（用于构建seed_list_next）
|
|
|
all_seed_combinations.extend(top_5)
|
|
|
|
|
|
- # 4.2 对于sug_list_list中，每个sug大于来自的query分数，加到q_list_next
|
|
|
+ # 4.2 对于sug_list_list中，每个sug大于来自的query分数，加到q_list_next（去重检查）
|
|
|
print(f"\n 4.2 将高分sug加入q_list_next...")
|
|
|
for sug in all_sugs:
|
|
|
if sug.from_q and sug.score_with_o > sug.from_q.score_with_o:
|
|
|
+ # 去重检查
|
|
|
+ if sug.text in existing_q_texts:
|
|
|
+ print(f" ⊗ 跳过重复: {sug.text}")
|
|
|
+ continue
|
|
|
+
|
|
|
new_q = Q(
|
|
|
text=sug.text,
|
|
|
score_with_o=sug.score_with_o,
|
|
|
@@ -991,6 +1055,7 @@ async def run_round(
|
|
|
from_source="sug"
|
|
|
)
|
|
|
q_list_next.append(new_q)
|
|
|
+ existing_q_texts.add(sug.text) # 记录到去重集合
|
|
|
print(f" ✓ {sug.text} (分数: {sug.score_with_o:.2f} > {sug.from_q.score_with_o:.2f})")
|
|
|
|
|
|
# 5. 构建seed_list_next（关键修改：不保留上一轮的seed）
|
刘立冬