rerank

tangwang
1 parent 7746376c
Showing 9 changed files with 170 additions and 88 deletions Show diff stats
api/models.py
api/routes/search.py
config/config.yaml
config/config_loader.py
docs/搜索API对接指南.md
query/test_translation.py
reranker/bge_reranker.py
search/rerank_client.py
search/searcher.py
@@ -151,9 +151,23 @@ class SearchRequest(BaseModel):
     min_score: Optional[float] = Field(None, ge=0, description="最小相关性分数阈值")
     highlight: bool = Field(False, description="是否高亮搜索关键词（暂不实现）")
     debug: bool = Field(False, description="是否返回调试信息")
-    ai_search: bool = Field(
+    enable_rerank: bool = Field(
         False,
-        description="是否开启 AI 搜索（调用本地重排服务对 ES 结果进行二次排序）"
+        description="是否开启重排（调用外部重排服务对 ES 结果进行二次排序）"
+    )
+    rerank_query_template: Optional[str] = Field(
+        None,
+        description=(
+            "重排 query 模板（可选）。支持 {query} 占位符。"
+            "不传则使用服务端配置的 rerank_query_template。"
+        ),
+    )
+    rerank_doc_template: Optional[str] = Field(
+        None,
+        description=(
+            "重排 doc 模板（可选）。支持 {title} {brief} {vendor} {description} {category_path} 占位符。"
+            "不传则使用服务端配置的 rerank_doc_template。"
+        ),
     )
  
     # SKU筛选参数
@@ -84,7 +84,9 @@ async def search(request: SearchRequest, http_request: Request):
             f"min_score: {request.min_score} | "
             f"language: {request.language} | "
             f"debug: {request.debug} | "
-            f"ai_search: {request.ai_search} | "
+            f"enable_rerank: {request.enable_rerank} | "
+            f"rerank_query_template: {request.rerank_query_template} | "
+            f"rerank_doc_template: {request.rerank_doc_template} | "
             f"sku_filter_dimension: {request.sku_filter_dimension} | "
             f"filters: {request.filters} | "
             f"range_filters: {request.range_filters} | "
@@ -112,7 +114,9 @@ async def search(request: SearchRequest, http_request: Request):
             debug=request.debug,
             language=request.language,
             sku_filter_dimension=request.sku_filter_dimension,
-            ai_search=request.ai_search,
+            enable_rerank=request.enable_rerank,
+            rerank_query_template=request.rerank_query_template,
+            rerank_doc_template=request.rerank_doc_template,
         )
  
         # Include performance summary in response
@@ -133,14 +133,19 @@ function_score:
   boost_mode: "multiply"
   functions: []
  
-# 重排配置（唯一实现：外部 BGE 重排服务，由请求参数 ai_search 控制是否执行）
-# ai_search 且 from+size<=rerank_window 时：从 ES 取前 rerank_window 条、重排后再按 from/size 分页
+# 重排配置（唯一实现：外部 BGE 重排服务，由请求参数 enable_rerank 控制是否执行）
+# enable_rerank 且 from+size<=rerank_window 时：从 ES 取前 rerank_window 条、重排后再按 from/size 分页
 rerank:
   rerank_window: 1000
   # service_url: "http://127.0.0.1:6007/rerank"  # 可选，不填则用默认端口 6007
   timeout_sec: 15.0   # 文档多时重排耗时长，可按需调大
   weight_es: 0.4
   weight_ai: 0.6
+  # 模板：用于将搜索请求/文档字段组装成重排服务输入
+  # - rerank_query_template：支持 {query}
+  # - rerank_doc_template：支持 {title} {brief} {vendor} {description} {category_path}
+  rerank_query_template: "{query}"
+  rerank_doc_template: "{title}"
  
 # SPU配置（已启用，使用嵌套skus）
 spu_config:
@@ -88,14 +88,19 @@ class RankingConfig:
  
 @dataclass
 class RerankConfig:
-    """重排配置（唯一实现：调用外部 BGE 重排服务，由请求参数 ai_search 控制是否执行）"""
-    # 重排窗口：ai_search 且 from+size<=rerank_window 时，从 ES 取前 rerank_window 条重排后再分页
+    """重排配置（唯一实现：调用外部 BGE 重排服务，由请求参数 enable_rerank 控制是否执行）"""
+    # 重排窗口：enable_rerank 且 from+size<=rerank_window 时，从 ES 取前 rerank_window 条重排后再分页
     rerank_window: int = 1000
     # 可选：重排服务 URL，为空时使用 reranker 模块默认端口 6007
     service_url: Optional[str] = None
     timeout_sec: float = 15.0
     weight_es: float = 0.4
     weight_ai: float = 0.6
+    # 模板：用于将搜索请求/文档字段组装成重排服务输入
+    # - rerank_query_template：支持 {query}
+    # - rerank_doc_template：支持 {title} {brief} {vendor} {description} {category_path}
+    rerank_query_template: str = "{query}"
+    rerank_doc_template: str = "{title}"
  
  
 @dataclass
@@ -267,7 +272,7 @@ class ConfigLoader:
             functions=fs_data.get("functions") or []
         )
  
-        # Parse Rerank configuration（唯一实现：外部重排服务，由 ai_search 控制）
+        # Parse Rerank configuration（唯一实现：外部重排服务，由 enable_rerank 控制）
         rerank_data = config_data.get("rerank", {})
         rerank = RerankConfig(
             rerank_window=int(rerank_data.get("rerank_window", 1000)),
@@ -275,6 +280,8 @@ class ConfigLoader:
             timeout_sec=float(rerank_data.get("timeout_sec", 15.0)),
             weight_es=float(rerank_data.get("weight_es", 0.4)),
             weight_ai=float(rerank_data.get("weight_ai", 0.6)),
+            rerank_query_template=str(rerank_data.get("rerank_query_template") or "{query}"),
+            rerank_doc_template=str(rerank_data.get("rerank_doc_template") or "{title}"),
         )
  
         # Parse SPU config
@@ -410,6 +417,8 @@ class ConfigLoader:
                 "timeout_sec": config.rerank.timeout_sec,
                 "weight_es": config.rerank.weight_es,
                 "weight_ai": config.rerank.weight_ai,
+                "rerank_query_template": config.rerank.rerank_query_template,
+                "rerank_doc_template": config.rerank.rerank_doc_template,
             },
             "spu_config": {
                 "enabled": config.spu_config.enabled,
@@ -167,7 +167,9 @@ curl -X POST &quot;http://120.76.41.98:6002/search/&quot; \
   "min_score": 0.0,
   "sku_filter_dimension": ["string"],
   "debug": false,
-  "ai_search": false,
+  "enable_rerank": false,
+  "rerank_query_template": "{query}",
+  "rerank_doc_template": "{title}",
   "user_id": "string",
   "session_id": "string"
 }
@@ -189,7 +191,9 @@ curl -X POST &quot;http://120.76.41.98:6002/search/&quot; \
 | `min_score` | float | N | null | 最小相关性分数阈值 |
 | `sku_filter_dimension` | array[string] | N | null | 子SKU筛选维度列表（见[SKU筛选维度](#35-sku筛选维度)） |
 | `debug` | boolean | N | false | 是否返回调试信息 |
-| `ai_search` | boolean | N | false | 是否开启 AI 搜索（调用本地重排服务对 ES 结果进行二次排序） |
+| `enable_rerank` | boolean | N | false | 是否开启重排（调用外部重排服务对 ES 结果进行二次排序）。开启后若 `from+size<=rerank_window` 才会触发重排 |
+| `rerank_query_template` | string | N | null | 重排 query 模板（可选）。支持 `{query}` 占位符；不传则使用服务端配置 |
+| `rerank_doc_template` | string | N | null | 重排 doc 模板（可选）。支持 `{title} {brief} {vendor} {description} {category_path}`；不传则使用服务端配置 |
 | `user_id` | string | N | null | 用户ID（用于个性化，预留） |
 | `session_id` | string | N | null | 会话ID（用于分析，预留） |
  
 #!/usr/bin/env python3
 """
-翻译功能测试脚本。
+Translation function test script.
  
-测试内容：
-1. 翻译提示词配置加载
-2. 同步翻译（索引场景）
-3. 异步翻译（查询场景）
-4. 不同提示词的使用
-5. 缓存功能
-6. DeepL Context参数使用
+Test content:
+1. Translation prompt configuration loading
+2. Synchronous translation (indexing scenario)
+3. Asynchronous translation (query scenario)
+4. Usage of different prompts
+5. Cache functionality
+6. DeepL Context parameter usage
 """
  
 import sys
@@ -31,37 +31,37 @@ logger = logging.getLogger(__name__)
  
  
 def test_config_loading():
-    """测试配置加载"""
+    """Test configuration loading"""
     print("\n" + "="*60)
-    print("测试1: 配置加载")
+    print("Test 1: Configuration loading")
     print("="*60)
  
     try:
         config_loader = ConfigLoader()
         config = config_loader.load_config()
  
-        print(f"✓ 配置加载成功")
-        print(f"  翻译服务: {config.query_config.translation_service}")
-        print(f"  翻译提示词配置:")
+        print(f"✓ Configuration loaded successfully")
+        print(f"  Translation service: {config.query_config.translation_service}")
+        print(f"  Translation prompt configuration:")
         for key, value in config.query_config.translation_prompts.items():
             print(f"    {key}: {value[:60]}..." if len(value) > 60 else f"    {key}: {value}")
  
         return config
     except Exception as e:
-        print(f"✗ 配置加载失败: {e}")
+        print(f"✗ Configuration loading failed: {e}")
         import traceback
         traceback.print_exc()
         return None
  
  
 def test_translator_sync(config):
-    """测试同步翻译（索引场景）"""
+    """Test synchronous translation (indexing scenario)"""
     print("\n" + "="*60)
-    print("测试2: 同步翻译（索引场景）")
+    print("Test 2: Synchronous translation (indexing scenario)")
     print("="*60)
  
     if not config:
-        print("✗ 跳过：配置未加载")
+        print("✗ Skipped: Configuration not loaded")
         return None
  
     try:
@@ -90,10 +90,10 @@ def test_translator_sync(config):
                 else:
                     prompt = config.query_config.translation_prompts.get('default_en')
  
-            print(f"\n翻译测试:")
-            print(f"  原文 ({source_lang}): {text}")
-            print(f"  目标语言: {target_lang}")
-            print(f"  提示词: {prompt[:50] if prompt else 'None'}...")
+            print(f"\nTranslation test:")
+            print(f"  Original text ({source_lang}): {text}")
+            print(f"  Target language: {target_lang}")
+            print(f"  Prompt: {prompt[:50] if prompt else 'None'}...")
  
             result = translator.translate(
                 text,
@@ -103,28 +103,28 @@ def test_translator_sync(config):
             )
  
             if result:
-                print(f"  结果: {result}")
-                print(f"  ✓ 翻译成功")
+                print(f"  Result: {result}")
+                print(f"  ✓ Translation successful")
             else:
-                print(f"  ⚠ 翻译返回None（可能是mock模式或无API key）")
+                print(f"  ⚠ Translation returned None (possibly mock mode or no API key)")
  
         return translator
  
     except Exception as e:
-        print(f"✗ 同步翻译测试失败: {e}")
+        print(f"✗ Synchronous translation test failed: {e}")
         import traceback
         traceback.print_exc()
         return None
  
  
 def test_translator_async(config, translator):
-    """测试异步翻译（查询场景）"""
+    """Test asynchronous translation (query scenario)"""
     print("\n" + "="*60)
-    print("测试3: 异步翻译（查询场景）")
+    print("Test 3: Asynchronous translation (query scenario)")
     print("="*60)
  
     if not config or not translator:
-        print("✗ 跳过：配置或翻译器未初始化")
+        print("✗ Skipped: Configuration or translator not initialized")
         return
  
     try:
@@ -134,9 +134,9 @@ def test_translator_async(config, translator):
  
         query_prompt = config.query_config.translation_prompts.get('query_zh')
  
-        print(f"查询文本: {query_text}")
-        print(f"目标语言: {target_langs}")
-        print(f"提示词: {query_prompt}")
+        print(f"Query text: {query_text}")
+        print(f"Target languages: {target_langs}")
+        print(f"Prompt: {query_prompt}")
  
         # 异步模式（立即返回，后台翻译）
         results = translator.translate_multi(
@@ -148,15 +148,15 @@ def test_translator_async(config, translator):
             prompt=query_prompt
         )
  
-        print(f"\n异步翻译结果:")
+        print(f"\nAsynchronous translation results:")
         for lang, translation in results.items():
             if translation:
-                print(f"  {lang}: {translation} (缓存命中)")
+                print(f"  {lang}: {translation} (cache hit)")
             else:
-                print(f"  {lang}: None (后台翻译中...)")
+                print(f"  {lang}: None (translating in background...)")
  
         # 同步模式（等待完成）
-        print(f"\n同步翻译（等待完成）:")
+        print(f"\nSynchronous translation (waiting for completion):")
         results_sync = translator.translate_multi(
             query_text,
             target_langs,
@@ -170,7 +170,7 @@ def test_translator_async(config, translator):
             print(f"  {lang}: {translation}")
  
     except Exception as e:
-        print(f"✗ 异步翻译测试失败: {e}")
+        print(f"✗ Asynchronous translation test failed: {e}")
         import traceback
         traceback.print_exc()
  
@@ -178,7 +178,7 @@ def test_translator_async(config, translator):
 def test_cache():
     """测试缓存功能"""
     print("\n" + "="*60)
-    print("测试4: 缓存功能")
+    print("Test 4: Cache functionality")
     print("="*60)
  
     try:
@@ -195,29 +195,29 @@ def test_cache():
         source_lang = "zh"
         prompt = config.query_config.translation_prompts.get('default_zh')
  
-        print(f"第一次翻译（应该调用API或返回mock）:")
+        print(f"First translation (should call API or return mock):")
         result1 = translator.translate(test_text, target_lang, source_lang, prompt=prompt)
-        print(f"  结果: {result1}")
-        
-        print(f"\n第二次翻译（应该使用缓存）:")
+        print(f"  Result: {result1}")
+
+        print(f"\nSecond translation (should use cache):")
         result2 = translator.translate(test_text, target_lang, source_lang, prompt=prompt)
-        print(f"  结果: {result2}")
-        
+        print(f"  Result: {result2}")
+
         if result1 == result2:
-            print(f"  ✓ 缓存功能正常")
+            print(f"  ✓ Cache functionality working properly")
         else:
-            print(f"  ⚠ 缓存可能有问题")
+            print(f"  ⚠ Cache might have issues")
  
     except Exception as e:
-        print(f"✗ 缓存测试失败: {e}")
+        print(f"✗ Cache test failed: {e}")
         import traceback
         traceback.print_exc()
  
  
 def test_context_parameter():
-    """测试DeepL Context参数使用"""
+    """Test DeepL Context parameter usage"""
     print("\n" + "="*60)
-    print("测试5: DeepL Context参数")
+    print("Test 5: DeepL Context parameter")
     print("="*60)
  
     try:
@@ -233,8 +233,8 @@ def test_context_parameter():
         text = "手机"
         prompt = config.query_config.translation_prompts.get('query_zh')
  
-        print(f"测试文本: {text}")
-        print(f"提示词（作为context）: {prompt}")
+        print(f"Test text: {text}")
+        print(f"Prompt (as context): {prompt}")
  
         # 带context的翻译
         result_with_context = translator.translate(
@@ -243,7 +243,7 @@ def test_context_parameter():
             source_lang='zh',
             prompt=prompt
         )
-        print(f"\n带context翻译结果: {result_with_context}")
+        print(f"\nTranslation result with context: {result_with_context}")
  
         # 不带context的翻译
         result_without_context = translator.translate(
@@ -252,21 +252,21 @@ def test_context_parameter():
             source_lang='zh',
             prompt=None
         )
-        print(f"不带context翻译结果: {result_without_context}")
+        print(f"Translation result without context: {result_without_context}")
  
-        print(f"\n✓ Context参数测试完成")
-        print(f"  注意：根据DeepL API，context参数影响翻译但不参与翻译本身")
+        print(f"\n✓ Context parameter test completed")
+        print(f"  Note: According to DeepL API, context parameter affects translation but does not participate in translation itself")
  
     except Exception as e:
-        print(f"✗ Context参数测试失败: {e}")
+        print(f"✗ Context parameter test failed: {e}")
         import traceback
         traceback.print_exc()
  
  
 def main():
-    """主测试函数"""
+    """Main test function"""
     print("="*60)
-    print("翻译功能测试")
+    print("Translation function test")
     print("="*60)
  
     # 测试1: 配置加载
@@ -285,7 +285,7 @@ def main():
     test_context_parameter()
  
     print("\n" + "="*60)
-    print("测试完成")
+    print("Test completed")
     print("="*60)
  
  
@@ -112,6 +112,17 @@ class BGEReranker:
         total_docs = len(docs)
         output_scores: List[float] = [0.0] * total_docs
  
+        # Log request summary (query + first 3 docs preview)
+        preview_docs: List[str] = []
+        for d in docs[:3]:
+            preview_docs.append("" if d is None else str(d))
+        logger.info(
+            "[BGE_RERANKER] Request | query=%r | docs=%d | docs_preview=%s",
+            query,
+            total_docs,
+            preview_docs,
+        )
+
         indexed_docs: List[Tuple[int, str]] = []
         for i, doc in enumerate(docs):
             if doc is None:
@@ -158,6 +169,15 @@ class BGEReranker:
         for (orig_idx, _text), unique_idx in zip(indexed_docs, position_to_unique):
             output_scores[orig_idx] = float(unique_scores[unique_idx])
  
+        # Log per-doc scores (aligned to original docs order)
+        try:
+            lines = []
+            for i, d in enumerate(docs[:100]):
+                lines.append(f"{output_scores[i]},{'' if d is None else str(d)}")
+            logger.info("[BGE_RERANKER] query:%s Scores (score,doc):\n%s", query, "\n".join(lines))
+        except Exception:
+            pass
+
         elapsed_ms = (time.time() - start_ts) * 1000.0
         dedup_ratio = 0.0
         if indexed_docs:
@@ -22,12 +22,13 @@ DEFAULT_TIMEOUT_SEC = 15.0
 def build_docs_from_hits(
     es_hits: List[Dict[str, Any]],
     language: str = "zh",
+    doc_template: str = "{title}",
 ) -> List[str]:
     """
     从 ES 命中结果构造重排服务所需的文档文本列表（与 hits 一一对应）。
  
-    文本由 title、brief、description、vendor、category_path 等多语言字段拼接，
-    按 language 优先选取对应语言；若无内容则用 spu_id 兜底。
+    使用 doc_template 将文档字段组装为重排服务输入。
+    支持占位符：{title} {brief} {vendor} {description} {category_path}
  
     Args:
         es_hits: ES 返回的 hits 列表，每项含 _source
@@ -47,16 +48,29 @@ def build_docs_from_hits(
             return str(obj.get(lang) or obj.get("zh") or obj.get("en") or "").strip()
         return str(obj).strip()
  
+    class _SafeDict(dict):
+        def __missing__(self, key: str) -> str:
+            return ""
+
     docs: List[str] = []
+    only_title = "{title}" == doc_template
+    need_brief = "{brief}" in doc_template
+    need_vendor = "{vendor}" in doc_template
+    need_description = "{description}" in doc_template
+    need_category_path = "{category_path}" in doc_template
     for hit in es_hits:
         src = hit.get("_source") or {}
-        parts: List[str] = []
-        for key in ("title", "brief", "description", "vendor", "category_path"):
-            parts.append(pick_lang_text(src.get(key)))
-        text = " ".join(p for p in parts if p).strip()
-        if not text:
-            text = str(src.get("spu_id", ""))
-        docs.append(text)
+        if only_title:
+            docs.append(pick_lang_text(src.get("title")))
+        else:
+            values = _SafeDict(
+                title=pick_lang_text(src.get("title")),
+                brief=pick_lang_text(src.get("brief")) if need_brief else "",
+                vendor=pick_lang_text(src.get("vendor")) if need_vendor else "",
+                description=pick_lang_text(src.get("description")) if need_description else "",
+                category_path=pick_lang_text(src.get("category_path")) if need_category_path else "",
+            )
+            docs.append(str(doc_template).format_map(values))
     return docs
  
  
@@ -188,6 +202,8 @@ def run_rerank(
     timeout_sec: float = DEFAULT_TIMEOUT_SEC,
     weight_es: float = DEFAULT_WEIGHT_ES,
     weight_ai: float = DEFAULT_WEIGHT_AI,
+    rerank_query_template: str = "{query}",
+    rerank_doc_template: str = "{title}",
 ) -> Tuple[Dict[str, Any], Optional[Dict[str, Any]], List[Dict[str, Any]]]:
     """
     完整重排流程：从 es_response 取 hits -> 构造 docs -> 调服务 -> 融合分数并重排 -> 更新 max_score。
@@ -222,8 +238,10 @@ def run_rerank(
     if not hits:
         return es_response, None, []
  
-    docs = build_docs_from_hits(hits, language=language)
-    scores, meta = call_rerank_service(query, docs, url, timeout_sec=timeout_sec)
+    # Apply query template (supports {query})
+    query_text = str(rerank_query_template).format_map({"query": query})
+    docs = build_docs_from_hits(hits, language=language, doc_template=rerank_doc_template)
+    scores, meta = call_rerank_service(query_text, docs, url, timeout_sec=timeout_sec)
  
     if scores is None or len(scores) != len(hits):
         return es_response, None, []
@@ -135,7 +135,9 @@ class Searcher:
         debug: bool = False,
         language: str = "en",
         sku_filter_dimension: Optional[List[str]] = None,
-        ai_search: bool = False,
+        enable_rerank: bool = False,
+        rerank_query_template: Optional[str] = None,
+        rerank_doc_template: Optional[str] = None,
     ) -> SearchResult:
         """
         Execute search query (外部友好格式).
@@ -167,11 +169,11 @@ class Searcher:
         index_langs = tenant_cfg.get("index_languages") or []
         enable_translation = len(index_langs) > 0
         enable_embedding = self.config.query_config.enable_text_embedding
-        # 重排仅由请求参数 ai_search 控制，唯一实现为调用外部 BGE 重排服务
-        enable_rerank = bool(ai_search)
+        # 重排仅由请求参数 enable_rerank 控制，唯一实现为调用外部 BGE 重排服务
+        do_rerank = bool(enable_rerank)
         rerank_window = self.config.rerank.rerank_window or 1000
         # 若开启重排且请求范围在窗口内：从 ES 取前 rerank_window 条、重排后再按 from/size 分页；否则不重排，按原 from/size 查 ES
-        in_rerank_window = enable_rerank and (from_ + size) <= rerank_window
+        in_rerank_window = do_rerank and (from_ + size) <= rerank_window
         es_fetch_from = 0 if in_rerank_window else from_
         es_fetch_size = rerank_window if in_rerank_window else size
  
@@ -180,7 +182,7 @@ class Searcher:
  
         context.logger.info(
             f"开始搜索请求 | 查询: '{query}' | 参数: size={size}, from_={from_}, "
-            f"enable_rerank={enable_rerank}, in_rerank_window={in_rerank_window}, es_fetch=({es_fetch_from},{es_fetch_size}) | "
+            f"enable_rerank={do_rerank}, in_rerank_window={in_rerank_window}, es_fetch=({es_fetch_from},{es_fetch_size}) | "
             f"enable_translation={enable_translation}, enable_embedding={enable_embedding}, min_score={min_score}",
             extra={'reqid': context.reqid, 'uid': context.uid}
         )
@@ -192,12 +194,14 @@ class Searcher:
             'es_fetch_from': es_fetch_from,
             'es_fetch_size': es_fetch_size,
             'in_rerank_window': in_rerank_window,
+            'rerank_query_template': rerank_query_template,
+            'rerank_doc_template': rerank_doc_template,
             'filters': filters,
             'range_filters': range_filters,
             'facets': facets,
             'enable_translation': enable_translation,
             'enable_embedding': enable_embedding,
-            'enable_rerank': enable_rerank,
+            'enable_rerank': do_rerank,
             'min_score': min_score,
             'sort_by': sort_by,
             'sort_order': sort_order
@@ -206,7 +210,7 @@ class Searcher:
         context.metadata['feature_flags'] = {
             'translation_enabled': enable_translation,
             'embedding_enabled': enable_embedding,
-            'rerank_enabled': enable_rerank
+            'rerank_enabled': do_rerank
         }
  
         # Step 1: Parse query
@@ -374,13 +378,15 @@ class Searcher:
             context.end_stage(RequestContextStage.ELASTICSEARCH_SEARCH)
  
         # Optional Step 4.5: AI reranking（仅当请求范围在重排窗口内时执行）
-        if enable_rerank and in_rerank_window:
+        if do_rerank and in_rerank_window:
             context.start_stage(RequestContextStage.RERANKING)
             try:
                 from .rerank_client import run_rerank
  
                 rerank_query = parsed_query.original_query if parsed_query else query
                 rc = self.config.rerank
+                effective_query_template = rerank_query_template or rc.rerank_query_template
+                effective_doc_template = rerank_doc_template or rc.rerank_doc_template
                 es_response, rerank_meta, fused_debug = run_rerank(
                     query=rerank_query,
                     es_response=es_response,
@@ -389,6 +395,8 @@ class Searcher:
                     timeout_sec=rc.timeout_sec,
                     weight_es=rc.weight_es,
                     weight_ai=rc.weight_ai,
+                    rerank_query_template=effective_query_template,
+                    rerank_doc_template=effective_doc_template,
                 )
  
                 if rerank_meta is not None: