Use decoupled way for backend client

services/llm.py

@@ -7,10 +7,10 @@ for real-time voice conversation.
 import os
 import asyncio
 import uuid
-from typing import AsyncIterator, Optional, List, Dict, Any
+from typing import AsyncIterator, Optional, List, Dict, Any, Callable, Awaitable
 from loguru import logger
 
-from app.backend_client import search_knowledge_context
+from app.backend_adapters import build_backend_adapter_from_settings
 from services.base import BaseLLMService, LLMMessage, LLMStreamEvent, ServiceState
 
 # Try to import openai
@@ -37,6 +37,7 @@ class OpenAILLMService(BaseLLMService):
         base_url: Optional[str] = None,
         system_prompt: Optional[str] = None,
         knowledge_config: Optional[Dict[str, Any]] = None,
+        knowledge_searcher: Optional[Callable[..., Awaitable[List[Dict[str, Any]]]]] = None,
     ):
         """
         Initialize OpenAI LLM service.
@@ -60,6 +61,11 @@ class OpenAILLMService(BaseLLMService):
         self.client: Optional[AsyncOpenAI] = None
         self._cancel_event = asyncio.Event()
         self._knowledge_config: Dict[str, Any] = knowledge_config or {}
+        if knowledge_searcher is None:
+            adapter = build_backend_adapter_from_settings()
+            self._knowledge_searcher = adapter.search_knowledge_context
+        else:
+            self._knowledge_searcher = knowledge_searcher
         self._tool_schemas: List[Dict[str, Any]] = []
 
     _RAG_DEFAULT_RESULTS = 5
@@ -224,7 +230,7 @@ class OpenAILLMService(BaseLLMService):
         n_results = self._coerce_int(cfg.get("nResults"), self._RAG_DEFAULT_RESULTS)
         n_results = max(1, min(n_results, self._RAG_MAX_RESULTS))
 
-        results = await search_knowledge_context(
+        results = await self._knowledge_searcher(
             kb_id=kb_id,
             query=latest_user,
             n_results=n_results,