Update LLM response generator and system prompts

- Add "/no_think" directive to English and Chinese system prompts.
- Handle `<think>` and `</think>` tags in response chunk processing.
- Update LLM model file to `Qwen3-8B-Q6_K.gguf`.
- Adjust model parameters: increase context length, modify temperature, top_p, and top_k, and include `presence_penalty`.

assets/models/llm/qwen/{Qwen2.5-14B-Instruct.Q4_0.gguf → Qwen3-8B-Q6_K.gguf}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1c9d82509108e9f758ff9ad1034d2c23088c611c76eaec308ad246caf8decac5
-size 8517725952
+oid sha256:cb042ccd76795a8830d6be6bd4165245847cc68e41797b13bd61aed4c2cfbce6
+size 6725899040

src/voice_dialogue/services/text/generator.py

@@ -12,15 +12,21 @@ from voice_dialogue.models.voice_task import VoiceTask, QuestionDisplayMessage
 from voice_dialogue.services.text.processor import preprocess_sentence_text, \
     create_langchain_chat_llamacpp_instance, create_langchain_pipeline, warmup_langchain_pipeline
 
-CHINESE_SYSTEM_PROMPT = ("你是善于模拟真实的思考过程的AI助手。"
-                         "回答时，必须首先生成一个不超过5个字的简短句子，"
-                         "比如：\"让我想一下\"、\"在我看来\"、\"稍等我理一理\"、\"不错的问题\"、\"稍等片刻\"等，然后再进行正式回答，"
-                         "保持中文口语化表达，禁用emoji和系统相关描述，确保衔接词与内容存在合理逻辑关联。")
-
-ENGLISH_SYSTEM_PROMPT = ("You are an AI assistant skilled at simulating authentic thinking processes. "
-                         "When responding, you must first generate a brief phrase of no more than 5 words, "
-                         "such as: 'Let me think', 'I see', 'Let me process this', 'Good question', 'One moment', etc., then proceed with your formal response. "
-                         "Maintain natural conversational English expression, avoid emojis and system-related descriptions, and ensure logical coherence between transitional phrases and content.")
+CHINESE_SYSTEM_PROMPT = (
+    "你是善于模拟真实的思考过程的AI助手。"
+    "回答时，必须首先生成一个不超过5个字的简短句子，"
+    "比如：\"让我想一下\"、\"在我看来\"、\"稍等我理一理\"、\"不错的问题\"、\"稍等片刻\"等，然后再进行正式回答，"
+    "保持中文口语化表达，禁用emoji和系统相关描述，确保衔接词与内容存在合理逻辑关联。"
+    "/no_think"
+)
+
+ENGLISH_SYSTEM_PROMPT = (
+    "You are an AI assistant skilled at simulating authentic thinking processes. "
+    "When responding, you must first generate a brief phrase of no more than 5 words, "
+    "such as: 'Let me think', 'I see', 'Let me process this', 'Good question', 'One moment', etc., then proceed with your formal response. "
+    "Maintain natural conversational English expression, avoid emojis and system-related descriptions, and ensure logical coherence between transitional phrases and content."
+    "/no_think"
+)
 
 
 class LLMResponseGenerator(BaseThread):
@@ -144,6 +150,9 @@ class LLMResponseGenerator(BaseThread):
             for chunk in pipeline.stream(input={'input': user_question}, config=config):
                 if not chunk.content.strip():
                     continue
+                elif chunk.content.strip() in {'<think>', '</think>'}:
+                    continue
+
                 chunk_content = f'{chunk.content}'
 
                 sentence_end_mark, remain_content = self._process_chunk_content(chunk_content)
@@ -181,17 +190,18 @@ class LLMResponseGenerator(BaseThread):
         self._send_sentence_to_queue(voice_task, sentence, answer_index)
 
     def run(self):
-        model_path = paths.LLM_MODELS_PATH / 'qwen' / 'Qwen2.5-14B-Instruct.Q4_0.gguf'
+        model_path = paths.LLM_MODELS_PATH / 'qwen' / 'Qwen3-8B-Q6_K.gguf'
         model_params = {
             'streaming': True,
-            'n_gpu_layers': -1,
-            'n_batch': 512,
-            'n_ctx': 2048,
-            'f16_kv': True,
-            'temperature': 0.8,
-            # 'n_predict': -1,
-            'top_k': 50,
-            'top_p': 1.0,
+            'n_ctx': 32768,
+            'temperature': 0.7,
+            'top_p': 0.9,
+            'top_k': 20,
+            'model_kwargs': {
+                'mini_p': 0,
+                'presence_penalty': 1.5
+            },
+            'verbose': False,
         }
         self.model_instance = create_langchain_chat_llamacpp_instance(
             local_model_path=model_path, model_params=model_params