diff --git a/data_chain/apps/app.py b/data_chain/apps/app.py
index 6e69e20229680e9d2da9954c9ddaa278466788da..64fdf4efd0e2b3bc84035fb9f2615034611e2b50 100644
--- a/data_chain/apps/app.py
+++ b/data_chain/apps/app.py
@@ -52,7 +52,8 @@ from data_chain.rag import (
     vector_searcher,
     keyword_and_vector_searcher,
     doc2chunk_searcher,
-    doc2chunk_bfs_searcher
+    doc2chunk_bfs_searcher,
+    enhanced_by_llm_searcher
 )
 from data_chain.stores.database.database import DataBase, ActionEntity, DocumentTypeEntity
 from data_chain.manager.role_manager import RoleManager
diff --git a/data_chain/apps/base/task/worker/acc_testing_worker.py b/data_chain/apps/base/task/worker/acc_testing_worker.py
index 41fbfeb67430b6127c22e46f9f792022b590d127..8ae3ceb6efb3e8cfe2d27ac7b1133c43d7249cc3 100644
--- a/data_chain/apps/base/task/worker/acc_testing_worker.py
+++ b/data_chain/apps/base/task/worker/acc_testing_worker.py
@@ -143,6 +143,19 @@ class TestingWorker(BaseWorker):
             answer = qa_entity.answer
             chunk = qa_entity.chunk
             chunk_entities = await BaseSearcher.search(testing_entity.search_method, testing_entity.kb_id, question, top_k=testing_entity.top_k, doc_ids=None, banned_ids=[])
+            related_chunk_entities = []
+            banned_ids = [chunk_entity.id for chunk_entity in chunk_entities]
+            divide_tokens = llm.max_tokens // len(chunk_entities) if chunk_entities else llm.max_tokens
+            leave_tokens = 0
+            for chunk_entity in chunk_entities:
+                leave_tokens += divide_tokens - chunk_entity.tokens
+                sub_related_chunk_entities = await BaseSearcher.related_surround_chunk(chunk_entity, leave_tokens, banned_ids)
+                banned_ids += [sub_chunk_entity.id for sub_chunk_entity in sub_related_chunk_entities]
+                related_chunk_entities += sub_related_chunk_entities
+                for related_chunk_entity in sub_related_chunk_entities:
+                    leave_tokens -= related_chunk_entity.tokens
+                leave_tokens = max(leave_tokens, 0)
+            chunk_entities += related_chunk_entities
             doc_chunk_dict = {}
             for chunk_entity in chunk_entities:
                 if chunk_entity.doc_id not in doc_chunk_dict:
diff --git a/data_chain/apps/base/task/worker/parse_document_worker.py b/data_chain/apps/base/task/worker/parse_document_worker.py
index 32bc4e0bd24dd5f42bb40c54c4eba7a8d403c0fc..7b9a75ef4727fde404b1dcada2f6b554aeff48c9 100644
--- a/data_chain/apps/base/task/worker/parse_document_worker.py
+++ b/data_chain/apps/base/task/worker/parse_document_worker.py
@@ -289,6 +289,8 @@ class ParseDocumentWorker(BaseWorker):
                 if len(nodes) == 0 or (len(nodes) and (nodes[-1].type != ChunkType.TEXT or TokenTool.get_tokens(nodes[-1].content) + tokens > doc_entity.chunk_size)):
                     nodes.append(node)
                 else:
+                    if node.is_need_newline:
+                        nodes[-1].content += '\n'
                     nodes[-1].content += node.content
             else:
                 nodes.append(node)
diff --git a/data_chain/apps/service/chunk_service.py b/data_chain/apps/service/chunk_service.py
index 4e3d2e976b4f75589f072cc3d7e3dead12d04e4a..5c7fd21d52d0a2653b20df1083fd0b679a26fc39 100644
--- a/data_chain/apps/service/chunk_service.py
+++ b/data_chain/apps/service/chunk_service.py
@@ -101,10 +101,11 @@ class ChunkService:
             chunk_entities = await BaseSearcher.rerank(chunk_entities, req.query)
         chunk_entities = chunk_entities[:req.top_k]
         chunk_ids = [chunk_entity.id for chunk_entity in chunk_entities]
+        logging.error("[ChunkService] 搜索分片，查询结果数量: %s", len(chunk_entities))
         if req.is_related_surrounding:
             # 关联上下文
             tokens_limit = req.tokens_limit
-            tokens_limit_every_chunk = tokens_limit // len(chunk_entities)
+            tokens_limit_every_chunk = tokens_limit // len(chunk_entities) if len(chunk_entities) > 0 else tokens_limit
             leave_tokens = 0
             related_chunk_entities = []
             for chunk_entity in chunk_entities:
@@ -125,6 +126,7 @@ class ChunkService:
                 chunk_ids += [chunk_entity.id for chunk_entity in sub_related_chunk_entities]
                 related_chunk_entities += sub_related_chunk_entities
             chunk_entities += related_chunk_entities
+        logging.error(len(chunk_entities))
         search_chunk_msg = SearchChunkMsg(docChunks=[])
         if req.is_classify_by_doc:
             doc_chunks = await BaseSearcher.classify_by_doc_id(chunk_entities)
diff --git a/data_chain/parser/handler/pdf_parser.py b/data_chain/parser/handler/pdf_parser.py
index fcc93e0564beaa86bdb3fa5cf6c7c784969eb458..65248a01a3382f5c9a53350c114734d8fa302717 100644
--- a/data_chain/parser/handler/pdf_parser.py
+++ b/data_chain/parser/handler/pdf_parser.py
@@ -151,7 +151,44 @@ class PdfParser(BaseParser):
                 xref = image_info[0]
                 # 提取基础图片（如果存在）
                 base_image = pdf_doc.extract_image(xref)
-                position = page.get_image_rects(xref)[0]
+
+                # 检查提取的图片是否有效
+                if not base_image or "image" not in base_image:
+                    logging.warning("[PdfParser] 标准方法提取失败，尝试替代方法 xref=%s", xref)
+                    continue
+
+                # 检查位置信息
+                rects = page.get_image_rects(xref)
+                if not rects:
+                    logging.warning("[PdfParser] 找不到图片位置，尝试基于布局估算 xref=%s", xref)
+                    width, height = base_image.get("width", 0), base_image.get("height", 0)
+                    if width <= 0 or height <= 0:
+                        logging.warning("[PdfParser] 图片尺寸无效，跳过 xref=%s", xref)
+                        continue
+                    # 获取页面尺寸
+                    page_width, page_height = page.rect.width, page.rect.height
+
+                    # 方法1: 默认居中布局
+                    x0 = (page_width - width) / 2
+                    y0 = (page_height - height) / 2
+
+                    # 方法2: 考虑文本布局，假设图片在页面上半部分
+                    # 这里可以集成文本布局分析，例如获取页面上的文本块位置
+                    # 然后避免与文本重叠
+
+                    # 方法3: 基于图片大小的智能布局
+                    # 如果图片很大，可能是全页图片，位置应从(0,0)开始
+                    if width > page_width * 0.8 and height > page_height * 0.8:
+                        x0, y0 = 0, 0
+                    # 如果图片很小，可能是图标或装饰，可能在角落
+                    elif width < page_width * 0.2 and height < page_height * 0.2:
+                        # 放在右上角作为默认位置
+                        x0 = page_width - width - 10  # 留出边距
+                        y0 = 10  # 留出边距
+
+                    position = fitz.Rect(x0, y0, x0 + width, y0 + height)
+                else:
+                    position = rects[0]
                 # 获取图片的二进制数据
                 blob = base_image["image"]
 
@@ -262,6 +299,10 @@ class PdfParser(BaseParser):
                 sub_nodes_with_bbox, image_nodes_with_bbox)
 
             nodes_with_bbox.extend(sub_nodes_with_bbox)
+        for i in range(1, len(nodes_with_bbox)):
+            '''根据bbox判断是否要进行换行'''
+            if nodes_with_bbox[i].bbox.y0 > nodes_with_bbox[i-1].bbox.y1 + 1:
+                nodes_with_bbox[i].node.is_need_newline = True
 
         nodes = [node_with_bbox.node for node_with_bbox in nodes_with_bbox]
         PdfParser.image_related_node_in_link_nodes(nodes)  # 假设这个方法在别处定义
diff --git a/data_chain/parser/handler/xlsx_parser.py b/data_chain/parser/handler/xlsx_parser.py
index ba1facbbf1712c9d03cf3be447c74f31da6c2cd0..b261f71b09cccc0b226c11dc72bbce1f627f74bb 100644
--- a/data_chain/parser/handler/xlsx_parser.py
+++ b/data_chain/parser/handler/xlsx_parser.py
@@ -32,15 +32,37 @@ class XlsxParser(BaseParser):
 
     @staticmethod
     async def parser(file_path: str) -> ParseResult:
-        try:
-            if file_path.endswith(('.xlsx', '.xls')):
+        if file_path.endswith(('.xlsx', '.xls')):
+            try:
                 data = pd.read_excel(file_path, sheet_name=None, header=None)
-            elif file_path.endswith('.csv'):
+            except Exception as e:
+                err = f"[XlsxParser] 解析Excel文件失败，error: {e}"
+                logging.exception(err)
+                raise e
+        elif file_path.endswith('.csv'):
+            try:
                 data = pd.read_csv(file_path, header=None)
-        except Exception as e:
-            err = "读取xlsx文件失败"
-            logging.exception("[XlsxParser] %s", err)
-            raise e
+            except Exception as e:
+                err = f"[XlsxParser] 解析CSV文件失败，error: {e}"
+                logging.exception(err)
+                raise e
+        else:
+            data = None
+            try:
+                data = pd.read_excel(file_path, sheet_name=None, header=None)
+            except Exception as e:
+                err = f"[XlsxParser] 解析文件失败，error: {e}"
+                logging.exception(err)
+            try:
+                data = pd.read_csv(file_path, header=None)
+            except Exception as e:
+                err = f"[XlsxParser] 解析文件失败，error: {e}"
+                logging.exception(err)
+            if data is None:
+                err = f"[XlsxParser] 无法解析文件，file_path: {file_path}"
+                logging.exception(err)
+                raise Exception(err)
+
         nodes = []
         for sheet_name, df in data.items():
             table_array = await XlsxParser.extract_table_to_array(df)
diff --git a/data_chain/parser/parse_result.py b/data_chain/parser/parse_result.py
index b701c0ff528c1291c5ecb0cd17835467382aff53..24dee22c1241748eb9a3c8d1f96944567ebb81b3 100644
--- a/data_chain/parser/parse_result.py
+++ b/data_chain/parser/parse_result.py
@@ -18,6 +18,7 @@ class ParseNode(BaseModel):
     content: Any = Field(..., description="节点内容")
     type: ChunkType = Field(..., description="节点类型")
     link_nodes: list = Field(..., description="链接节点")
+    is_need_newline: bool = Field(default=False, description="是否需要换行")
 
 
 class ParseResult(BaseModel):
diff --git a/data_chain/parser/tools/ocr_tool.py b/data_chain/parser/tools/ocr_tool.py
index 11739632af28721abdc0521a951cb06f2a3b2633..bfc7aab6f3415e73335625a127f27d4cf520b775 100644
--- a/data_chain/parser/tools/ocr_tool.py
+++ b/data_chain/parser/tools/ocr_tool.py
@@ -1,5 +1,4 @@
-from PIL import Image
-import asyncio
+from PIL import Image, ImageEnhance
 import yaml
 from paddleocr import PaddleOCR
 import numpy as np
@@ -13,23 +12,47 @@ class OcrTool:
     det_model_dir = 'data_chain/parser/model/ocr/ch_PP-OCRv4_det_infer'
     rec_model_dir = 'data_chain/parser/model/ocr/ch_PP-OCRv4_rec_infer'
     cls_model_dir = 'data_chain/parser/model/ocr/ch_ppocr_mobile_v2.0_cls_infer'
+    # 优化 OCR 参数配置
     model = PaddleOCR(
         det_model_dir=det_model_dir,
         rec_model_dir=rec_model_dir,
         cls_model_dir=cls_model_dir,
-        use_angle_cls=True,  # 是否使用角度分类模型
-        use_space_char=True  # 是否使用空格字符
+        use_angle_cls=True,
+        use_space_char=True,
+        det_db_thresh=0.3,       # 降低文本检测阈值，提高敏感度
+        det_db_box_thresh=0.5,   # 调整文本框阈值
     )
 
     @staticmethod
     async def ocr_from_image(image: np.ndarray) -> list:
         try:
+
+            # 尝试OCR识别
             ocr_result = OcrTool.model.ocr(image)
-            if ocr_result is None or ocr_result[0] is None:
+
+            # 如果第一次尝试失败，尝试不同的参数配置
+            if ocr_result is None or len(ocr_result) == 0 or ocr_result[0] is None:
+                logging.warning("[OCRTool] 第一次OCR尝试失败，尝试降低阈值...")
+                # 创建临时OCR实例，使用更低的阈值
+                temp_ocr = PaddleOCR(
+                    det_model_dir=OcrTool.det_model_dir,
+                    rec_model_dir=OcrTool.rec_model_dir,
+                    cls_model_dir=OcrTool.cls_model_dir,
+                    use_angle_cls=True,
+                    use_space_char=True,
+                    det_db_thresh=0.2,       # 更低的检测阈值
+                    det_db_box_thresh=0.4,   # 更低的文本框阈值
+                )
+                ocr_result = temp_ocr.ocr(image)
+
+            # 记录OCR结果状态
+            if ocr_result is None or len(ocr_result) == 0 or ocr_result[0] is None:
+                logging.warning("[OCRTool] 图片无法识别文本")
                 return None
+
             return ocr_result
         except Exception as e:
-            err = f"[OCRTool] OCR识别失败 {e}"
+            err = f"[OCRTool] OCR识别失败: {e}"
             logging.exception(err)
             return None
 
diff --git a/data_chain/rag/base_searcher.py b/data_chain/rag/base_searcher.py
index 72e65a9d0063bff652e629ed86447cb6e8755164..ef96e0f61ca6d1f2b5a1e65937d3426638eb62bb 100644
--- a/data_chain/rag/base_searcher.py
+++ b/data_chain/rag/base_searcher.py
@@ -65,7 +65,7 @@ class BaseSearcher:
         :param query: 查询
         :return: 相关上下文
         """
-        chunk_entities = await ChunkManager.fetch_surrounding_chunk_by_doc_id_and_global_offset(chunk_entity.doc_id, chunk_entity.global_offset, 50, banned_ids)
+        chunk_entities = await ChunkManager.fetch_surrounding_chunk_by_doc_id_and_global_offset(chunk_entity.doc_id, chunk_entity.global_offset, 100, banned_ids)
         chunk_entity_dict = {}
         lower = chunk_entity.global_offset-1
         upper = chunk_entity.global_offset+1
diff --git a/data_chain/rag/doc2chunk_bfs_searcher.py b/data_chain/rag/doc2chunk_bfs_searcher.py
index d30f21b9225b13c8e8e257c8d8ae27f66d2fa926..c72e7bd1c1a05b2c9ea68e027dac522714871210 100644
--- a/data_chain/rag/doc2chunk_bfs_searcher.py
+++ b/data_chain/rag/doc2chunk_bfs_searcher.py
@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 from pydantic import BaseModel, Field
 import random
@@ -33,7 +34,15 @@ class Doc2ChunkBfsSearcher(BaseSearcher):
         try:
             root_chunk_entities_keyword = await ChunkManager.get_top_k_chunk_by_kb_id_keyword(kb_id, query, top_k//2, doc_ids, banned_ids, ChunkParseTopology.TREEROOT.value)
             banned_ids += [chunk_entity.id for chunk_entity in root_chunk_entities_keyword]
-            root_chunk_entities_vector = await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(root_chunk_entities_keyword), doc_ids, banned_ids, ChunkParseTopology.TREEROOT.value)
+            root_chunk_entities_vector = []
+            for _ in range(3):
+                try:
+                    root_chunk_entities_vector = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(root_chunk_entities_keyword), doc_ids, banned_ids, ChunkParseTopology.TREEROOT.value), timeout=3)
+                    break
+                except Exception as e:
+                    err = f"[KeywordVectorSearcher] 向量检索失败，error: {e}"
+                    logging.error(err)
+                    continue
             banned_ids += [chunk_entity.id for chunk_entity in root_chunk_entities_vector]
             chunk_entities = root_chunk_entities_keyword + root_chunk_entities_vector
             pre_ids = [chunk_entity.id for chunk_entity in chunk_entities]
@@ -42,13 +51,21 @@ class Doc2ChunkBfsSearcher(BaseSearcher):
             while rd < max_retry:
                 root_chunk_entities_keyword = await ChunkManager.get_top_k_chunk_by_kb_id_keyword(kb_id, query, top_k//2, doc_ids, banned_ids, None, pre_ids)
                 banned_ids += [chunk_entity.id for chunk_entity in root_chunk_entities_keyword]
-                root_chunk_entities_vector = await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(root_chunk_entities_keyword), doc_ids, banned_ids, None, pre_ids)
+                root_chunk_entities_vector = []
+                for _ in range(3):
+                    try:
+                        root_chunk_entities_vector = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(root_chunk_entities_keyword), doc_ids, banned_ids, None, pre_ids), timeout=3)
+                        break
+                    except Exception as e:
+                        err = f"[KeywordVectorSearcher] 向量检索失败，error: {e}"
+                        logging.error(err)
+                        continue
                 banned_ids += [chunk_entity.id for chunk_entity in root_chunk_entities_vector]
                 sub_chunk_entities = root_chunk_entities_keyword + root_chunk_entities_vector
                 if len(sub_chunk_entities) == 0:
                     break
                 chunk_entities += sub_chunk_entities
-                pre_ids += [chunk_entity.id for chunk_entity in sub_chunk_entities]
+                pre_ids = [chunk_entity.id for chunk_entity in sub_chunk_entities]
                 rd += 1
         except Exception as e:
             err = f"[KeywordVectorSearcher] 关键词向量检索失败，error: {e}"
diff --git a/data_chain/rag/doc2chunk_searcher.py b/data_chain/rag/doc2chunk_searcher.py
index 437670289ac03e2abbfdd55e495f4a04058b4758..fc438a0b11b3632bd14a6d3b48b86d2031f60985 100644
--- a/data_chain/rag/doc2chunk_searcher.py
+++ b/data_chain/rag/doc2chunk_searcher.py
@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 from pydantic import BaseModel, Field
 import random
@@ -32,11 +33,27 @@ class Doc2ChunkSearcher(BaseSearcher):
         try:
             doc_entities_keyword = await DocumentManager.get_top_k_document_by_kb_id_keyword(kb_id, query, top_k, doc_ids, banned_ids)
             use_doc_ids = [doc_entity.id for doc_entity in doc_entities_keyword]
-            doc_entities_vector = await DocumentManager.get_top_k_document_by_kb_id_vector(kb_id, vector, top_k-len(doc_entities_keyword), doc_ids, banned_ids)
+            doc_entities_vector = []
+            for _ in range(3):
+                try:
+                    doc_entities_vector = await asyncio.wait_for(DocumentManager.get_top_k_document_by_kb_id_vector(kb_id, vector, top_k-len(doc_entities_keyword), doc_ids, banned_ids), timeout=3)
+                    break
+                except Exception as e:
+                    err = f"[KeywordVectorSearcher] 向量检索失败，error: {e}"
+                    logging.error(err)
+                    continue
             use_doc_ids += [doc_entity.id for doc_entity in doc_entities_vector]
             chunk_entities_keyword = await ChunkManager.get_top_k_chunk_by_kb_id_keyword(kb_id, query, top_k//2, use_doc_ids, banned_ids)
             chunk_ids = [chunk_entity.id for chunk_entity in chunk_entities_keyword]
-            chunk_entities_vector += await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(chunk_entities_keyword), use_doc_ids, banned_ids+chunk_ids)
+            chunk_entities_vector = []
+            for _ in range(3):
+                try:
+                    chunk_entities_vector = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(chunk_entities_keyword), use_doc_ids, banned_ids+chunk_ids), timeout=3)
+                    break
+                except Exception as e:
+                    err = f"[KeywordVectorSearcher] 向量检索失败，error: {e}"
+                    logging.error(err)
+                    continue
             chunk_entities = chunk_entities_keyword + chunk_entities_vector
         except Exception as e:
             err = f"[KeywordVectorSearcher] 关键词向量检索失败，error: {e}"
diff --git a/data_chain/rag/enhanced_by_llm_searcher.py b/data_chain/rag/enhanced_by_llm_searcher.py
index aab7c04fca902b2558b943c196a8da4dfd26790d..9ac2dd4ad844d20267e29d770e29f094a8ff2e23 100644
--- a/data_chain/rag/enhanced_by_llm_searcher.py
+++ b/data_chain/rag/enhanced_by_llm_searcher.py
@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 import yaml
 from pydantic import BaseModel, Field
@@ -38,7 +39,7 @@ class EnhancedByLLMSearcher(BaseSearcher):
             prompt_template = prompt_dict['CHUNK_QUERY_MATCH_PROMPT']
             chunk_entities = []
             rd = 0
-            max_retry = 5
+            max_retry = 15
             llm = LLM(
                 openai_api_key=config['OPENAI_API_KEY'],
                 openai_api_base=config['OPENAI_API_BASE'],
@@ -47,18 +48,33 @@ class EnhancedByLLMSearcher(BaseSearcher):
             )
             while len(chunk_entities) < top_k and rd < max_retry:
                 rd += 1
-                sub_chunk_entities = await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k, doc_ids, banned_ids)
+                sub_chunk_entities_keyword = await ChunkManager.get_top_k_chunk_by_kb_id_keyword(kb_id, query, top_k, doc_ids, banned_ids)
+                chunk_ids = [chunk_entity.id for chunk_entity in sub_chunk_entities_keyword]
+                banned_ids += chunk_ids
+                sub_chunk_entities_vector = []
+                for _ in range(3):
+                    try:
+                        sub_chunk_entities_vector = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k, doc_ids, banned_ids), timeout=3)
+                        break
+                    except Exception as e:
+                        err = f"[EnhancedByLLMSearcher] 向量检索失败，error: {e}"
+                        logging.error(err)
+                        continue
+                chunk_ids = [chunk_entity.id for chunk_entity in sub_chunk_entities_vector]
+                banned_ids += chunk_ids
+                sub_chunk_entities = sub_chunk_entities_keyword + sub_chunk_entities_vector
                 for chunk_entity in sub_chunk_entities:
                     sys_call = prompt_template.format(
-                        chunk=chunk_entity.text,
-                        query=query,
+                        chunk=TokenTool.get_k_tokens_words_from_content(chunk_entity.text, llm.max_tokens),
+                        question=query,
                     )
                     user_call = "请输出YES或NO"
                     result = await llm.nostream([], sys_call, user_call)
                     result = result.lower()
                     if result == "yes":
                         chunk_entities.append(chunk_entity)
-                chunk_ids = [chunk_entity.id for chunk_entity in sub_chunk_entities]
+                        logging.info(
+                            f"[EnhancedByLLMSearcher] 匹配到分片: {chunk_entity.id}, 分片内容: {chunk_entity.text[:50]}...")
                 banned_ids += chunk_ids
             return chunk_entities[:top_k]
         except Exception as e:
diff --git a/data_chain/rag/keyword_and_vector_searcher.py b/data_chain/rag/keyword_and_vector_searcher.py
index 18cfdc6e7dfd161e4706ced7025d5f688644bc8b..d3753029c57c1ae2ae0a9f53632c0f543e445856 100644
--- a/data_chain/rag/keyword_and_vector_searcher.py
+++ b/data_chain/rag/keyword_and_vector_searcher.py
@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 from pydantic import BaseModel, Field
 import random
@@ -31,7 +32,15 @@ class KeywordVectorSearcher(BaseSearcher):
         try:
             chunk_entities_get_by_keyword = await ChunkManager.get_top_k_chunk_by_kb_id_keyword(kb_id, query, top_k//2, doc_ids, banned_ids)
             chunk_ids = [chunk_entity.id for chunk_entity in chunk_entities_get_by_keyword]
-            chunk_entities_get_by_vector = await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(chunk_entities_get_by_keyword), doc_ids, banned_ids+chunk_ids)
+            chunk_entities_get_by_vector = []
+            for _ in range(3):
+                try:
+                    chunk_entities_get_by_vector = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k-len(chunk_entities_get_by_keyword), doc_ids, banned_ids+chunk_ids), timeout=3)
+                    break
+                except Exception as e:
+                    err = f"[KeywordVectorSearcher] 向量检索失败，error: {e}"
+                    logging.error(err)
+                    continue
             chunk_entities = chunk_entities_get_by_keyword + chunk_entities_get_by_vector
         except Exception as e:
             err = f"[KeywordVectorSearcher] 关键词向量检索失败，error: {e}"
diff --git a/data_chain/rag/vector_searcher.py b/data_chain/rag/vector_searcher.py
index b55ab43892d3fc6dc8f64cd52487d91bb12e9a58..dad5e8676792927fa28f27a0ec9b8ac0cb08a079 100644
--- a/data_chain/rag/vector_searcher.py
+++ b/data_chain/rag/vector_searcher.py
@@ -1,3 +1,4 @@
+import asyncio
 import uuid
 from pydantic import BaseModel, Field
 import random
@@ -25,10 +26,13 @@ class VectorSearcher(BaseSearcher):
         :return: 检索结果
         """
         vector = await Embedding.vectorize_embedding(query)
-        try:
-            chunk_entities = await ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k, doc_ids, banned_ids)
-        except Exception as e:
-            err = f"[VectorSearcher] 向量检索失败，error: {e}"
-            logging.exception(err)
-            return []
+        chunk_entities = []
+        for _ in range(3):
+            try:
+                chunk_entities = await asyncio.wait_for(ChunkManager.get_top_k_chunk_by_kb_id_vector(kb_id, vector, top_k, doc_ids, banned_ids), timeout=3)
+                break
+            except Exception as e:
+                err = f"[VectorSearcher] 向量检索失败，error: {e}"
+                logging.exception(err)
+                continue
         return chunk_entities