query translate

tangwang
1 parent af03fdef
Showing 4 changed files with 31 additions and 9 deletions Show diff stats
config/config.yaml
config/config_loader.py
docs/TODO.txt
query/query_parser.py
@@ -241,10 +241,10 @@ services:
         infer_batch_size: 64
         sort_by_doc_length: true
         length_sort_mode: "char"  # char | token
-        instruction: "Given a shopping query, rank product titles by relevance"
+        instruction: "rank products by given query"
       qwen3_transformers:
         model_name: "Qwen/Qwen3-Reranker-0.6B"
-        instruction: "Given a shopping query, rank product titles by relevance"
+        instruction: "rank products by given query"
         max_length: 8192
         batch_size: 64
         use_fp16: true
@@ -67,6 +67,14 @@ class QueryConfig:
     original_query_fallback_boost_when_translation_missing: float = 0.2
     tie_breaker_base_query: float = 0.9
+    # Query-time translation model selection (configurable)
+    # - zh_to_en_model: model for zh -> en
+    # - en_to_zh_model: model for en -> zh
+    # - default_translation_model: fallback model for all other language pairs
+    zh_to_en_model: str = "opus-mt-zh-en"
+    en_to_zh_model: str = "opus-mt-en-zh"
+    default_translation_model: str = "nllb-200-distilled-600m"
+
 @dataclass
 class SPUConfig:
@@ -262,6 +270,11 @@ class ConfigLoader:
                 text_strategy_cfg.get("original_query_fallback_boost_when_translation_missing", 0.2)
             ),
             tie_breaker_base_query=float(text_strategy_cfg.get("tie_breaker_base_query", 0.9)),
+            zh_to_en_model=str(query_config_data.get("zh_to_en_model") or "opus-mt-zh-en"),
+            en_to_zh_model=str(query_config_data.get("en_to_zh_model") or "opus-mt-en-zh"),
+            default_translation_model=str(
+                query_config_data.get("default_translation_model") or "nllb-200-distilled-600m"
+            ),
         )
         # Parse Function Score configuration
@@ -114,6 +114,10 @@ https://modelscope.cn/models/dengcao/Qwen3-Reranker-4B-GGUF/summary
 reranker 补充：nvidia/llama-nemotron-rerank-1b-v2
+https://huggingface.co/nvidia/llama-nemotron-rerank-1b-v2
+后端推理也建议使用vLLM
+注意搜索相关资料，挖掘我的特斯拉 T4 GPU 的性能，充分挖掘性能
+你有充足的自由度进行实验
 encoder架构。
 比较新。
 性能更好。
@@ -152,15 +152,20 @@ class QueryParser:
         return self._translator
     @staticmethod
-    def _pick_query_translation_model(source_lang: str, target_lang: str) -> str:
-        """Pick the translation capability for query-time translation."""
+    def _pick_query_translation_model(source_lang: str, target_lang: str, config: SearchConfig) -> str:
+        """Pick the translation capability for query-time translation (configurable)."""
         src = str(source_lang or "").strip().lower()
         tgt = str(target_lang or "").strip().lower()
+
+        # Use dedicated models for zh<->en if configured
         if src == "zh" and tgt == "en":
-            return "opus-mt-zh-en"
+            return config.query_config.zh_to_en_model
         if src == "en" and tgt == "zh":
-            return "opus-mt-en-zh"
-        return "deepl"
+            return config.query_config.en_to_zh_model
+
+        # For any other language pairs, fall back to the configurable default model.
+        # By default this is `nllb-200-distilled-600m` (multi-lingual local model).
+        return config.query_config.default_translation_model
     def _simple_tokenize(self, text: str) -> List[str]:
         """
@@ -363,7 +368,7 @@ class QueryParser:
                             thread_name_prefix="query-translation-wait",
                         )
                         for lang in target_langs:
-                            model_name = self._pick_query_translation_model(detected_lang, lang)
+                            model_name = self._pick_query_translation_model(detected_lang, lang, self.config)
                             log_debug(
                                 f"Submitting query translation | source={detected_lang} target={lang} model={model_name}"
                             )
@@ -377,7 +382,7 @@ class QueryParser:
                             )
                     else:
                         for lang in target_langs:
-                            model_name = self._pick_query_translation_model(detected_lang, lang)
+                            model_name = self._pick_query_translation_model(detected_lang, lang, self.config)
                             log_debug(
                                 f"Submitting query translation | source={detected_lang} target={lang} model={model_name}"
                             )