diff --git a/services/rag_service.py b/services/rag_service.py
index c126944..8c4873e 100644
--- a/services/rag_service.py
+++ b/services/rag_service.py
@@ -9,6 +9,7 @@ from transformers import AutoTokenizer, AutoModelForSequenceClassification
import os
import time
import torch
+import json
class AsyncRAGService:
@@ -38,14 +39,17 @@ class AsyncRAGService:
),
)
- self.tokenizer = AutoTokenizer.from_pretrained(
- "/Volumes/LRW/Model/Qwen3-Embedding-0.6B", trust_remote_code=True
- )
- self.rerank_model = AutoModelForSequenceClassification.from_pretrained(
- "/Volumes/LRW/Model/Qwen3-Embedding-0.6B",
- trust_remote_code=True,
- device_map="auto", # 或 "cuda"
- )
+ self.tokenizer = AutoTokenizer.from_pretrained("/Volumes/LRW/Model/Qwen3-Embedding-0.6B", padding_side="left")
+ # 强制设置 padding token
+ if self.tokenizer.pad_token is None:
+ self.tokenizer.pad_token = self.tokenizer.eos_token
+ self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+ self.rerank_model = AutoModelForSequenceClassification.from_pretrained("/Volumes/LRW/Model/Qwen3-Embedding-0.6B").eval()
+
+ # 确保模型配置与 tokenizer 一致
+ if hasattr(self.rerank_model.config, "pad_token_id"):
+ self.rerank_model.config.pad_token_id = self.tokenizer.pad_token_id
self.logger.info("RAG服务初始化完成")
@@ -214,37 +218,106 @@ class AsyncRAGService:
return await asyncio.to_thread(_format_sources)
async def _rerank_results(
- self, question: str, search_results: List[Dict[str, Any]]
+ self, question: str, search_results: List[Dict[str, Any]], skip_rerank: bool = True
) -> List[Dict[str, Any]]:
"""使用 Qwen3-Reranker 对搜索结果批量重排序"""
- # 准备 batch 输入:格式必须是 Query: xxx\nDocument: yyy
- batch_texts = [
- f"Query: {question}\nDocument: {r['content'][:1000]}" # 可以根据显存调整截断长度
+ if skip_rerank:
+ self.logger.info("跳过重排序")
+ return search_results
+
+ if not search_results:
+ return []
+
+ # 模型相关常量(可初始化时提前保存)
+ instruction = (
+ "Given a web search query, retrieve relevant passages that answer the query"
+ )
+ prefix = '<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>\n<|im_start|>user\n'
+ suffix = "<|im_end|>\n<|im_start|>assistant\n\n\n\n\n"
+
+ prefix_tokens = self.tokenizer.encode(prefix, add_special_tokens=False)
+ suffix_tokens = self.tokenizer.encode(suffix, add_special_tokens=False)
+
+ # 构造符合格式的输入
+ def format_pair(query, doc):
+ return f": {instruction}\n: {query}\n: {doc}"
+
+ pairs = [
+ format_pair(question, r["content"][:1000]) # 文本截断,避免超长
for r in search_results
]
- # 使用 tokenizer 构建 batch 输入
+ # 分词 + 拼接前后缀 + padding
inputs = self.tokenizer(
- batch_texts,
+ pairs,
+ padding="max_length",
+ truncation="longest_first",
+ return_attention_mask=True, # 确保返回 attention_mask
+ max_length=8192 - len(prefix_tokens) - len(suffix_tokens),
return_tensors="pt",
- padding=True,
- truncation=True,
- max_length=1024, # Qwen3 的最大上下文长度,建议限制
- ).to(self.rerank_model.device)
+ )
- # 推理打分(关闭梯度计算)
+ # 手动添加前后缀
+ batch_size = inputs["input_ids"].shape[0]
+ max_len = 8192
+
+ # 创建新的输入张量
+ new_input_ids = torch.full((batch_size, max_len), self.tokenizer.pad_token_id, dtype=torch.long)
+ new_attention_mask = torch.zeros((batch_size, max_len), dtype=torch.long)
+
+ for i in range(batch_size):
+ # 获取原始序列(去除padding)
+ original_ids = inputs["input_ids"][i]
+ original_mask = inputs["attention_mask"][i]
+ actual_length = original_mask.sum().item()
+
+ # 构建新序列:prefix + original + suffix
+ new_sequence = (
+ prefix_tokens + original_ids[:actual_length].tolist() + suffix_tokens
+ )
+ new_length = len(new_sequence)
+
+ if new_length <= max_len:
+ new_input_ids[i, :new_length] = torch.tensor(new_sequence)
+ new_attention_mask[i, :new_length] = 1
+
+ inputs = {
+ "input_ids": new_input_ids.to(self.rerank_model.device),
+ "attention_mask": new_attention_mask.to(self.rerank_model.device),
+ }
+
+ # 获取 yes / no 的 token id(初始化时保存也可)
+ token_true_id = self.tokenizer.convert_tokens_to_ids("yes")
+ token_false_id = self.tokenizer.convert_tokens_to_ids("no")
+
+ # 推理评分
with torch.no_grad():
outputs = self.rerank_model(**inputs)
- logits = outputs.logits.squeeze(-1)
+ logits = outputs.logits
- # 如果是二分类模型,通常需要做 sigmoid 激活
- scores = torch.sigmoid(logits).tolist()
+ # 检查 logits 的维度
+ if logits.dim() == 3:
+ # 如果是3维,取最后一个token的logits
+ logits = logits[:, -1, :]
+ elif logits.dim() == 2:
+ # 如果是2维,直接使用
+ pass
+ else:
+ raise ValueError(f"Unexpected logits dimension: {logits.dim()}")
- # 写入到每个 search_result 中
+ # 提取 yes/no token 的 logits
+ true_logits = logits[:, token_true_id]
+ false_logits = logits[:, token_false_id]
+
+ stacked = torch.stack([false_logits, true_logits], dim=1)
+ probs = torch.nn.functional.softmax(stacked, dim=1)
+ scores = probs[:, 1].tolist() # 取 "yes" 的概率值
+
+ # 写入每条结果
for r, score in zip(search_results, scores):
- r["rerank_score"] = max(0.0, min(score, 1.0)) # 保证分数在 0-1 范围
-
+ r["rerank_score"] = round(float(score), 4)
+ self.logger.info(f"重排序完成,得分范围: {min(scores)} - {max(scores)} \n\n {json.dumps(search_results, indent=4)}")
return search_results
async def _build_context_async(self, search_results: List[Dict[str, Any]]) -> str: