feat: integrate RAG rag_jrxml submodule and fix Anthropic API key

Add rag submodule for semantic JRXML chunk retrieval, refactor
retrieve node to use RAGSearcher, and fix missing api_key in
Anthropic SDK client initialization.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/embeddings.py

@@ -1,4 +1,9 @@
-"""嵌入模型工厂：支持本地 sentence-transformers 和云端 API。"""
+"""嵌入模型工厂：支持本地 Sentence-Transformers 和云端 API。
+
+调用方式:
+    get_embeddings()  → LangChain 兼容的 embeddings 对象
+    get_st_embeddings() → 原始 SentenceTransformer 实例
+"""
 
 import os
 from dotenv import load_dotenv
@@ -7,6 +12,7 @@ load_dotenv()
 
 
 def get_embeddings():
+    """返回 LangChain 兼容的 embeddings 对象（用于 langchain_chroma 等）。"""
     backend = os.getenv("EMBED_BACKEND", "local")
     if backend == "cloud":
         from langchain_openai import OpenAIEmbeddings
@@ -22,5 +28,21 @@ def get_embeddings():
         except ImportError:
             from langchain_community.embeddings import HuggingFaceEmbeddings
 
-        model = os.getenv("LOCAL_EMBED_MODEL", "Qwen/Qwen3-Embedding-0.6B")
+        model = os.getenv("RAG_EMBED_MODEL", os.getenv("LOCAL_EMBED_MODEL", "Qwen/Qwen3-Embedding-0.6B"))
         return HuggingFaceEmbeddings(model_name=model)
+
+
+def get_st_model():
+    """返回原始 SentenceTransformer 实例（与 rag_jrxml 子模块使用方式一致）。"""
+    import torch
+    from sentence_transformers import SentenceTransformer
+
+    model_name = os.getenv("RAG_EMBED_MODEL", os.getenv("LOCAL_EMBED_MODEL", "Qwen/Qwen3-Embedding-0.6B"))
+    use_gpu = os.getenv("RAG_USE_GPU", "true").lower() in ("true", "1")
+    use_fp16 = os.getenv("RAG_USE_FP16", "true").lower() in ("true", "1")
+
+    device = "cuda" if (use_gpu and torch.cuda.is_available()) else "cpu"
+    model = SentenceTransformer(model_name, device=device)
+    if device == "cuda" and use_fp16:
+        model = model.half()
+    return model