diff --git a/scripts/make_rag_database.py b/scripts/make_rag_database.py
index e69de29..61bffc7 100644
--- a/scripts/make_rag_database.py
+++ b/scripts/make_rag_database.py
@@ -0,0 +1,70 @@
+import pandas as pd 
+import os.path as osp
+import os
+
+from lang_agent.rag.emb import QwenEmbeddings
+
+from langchain.text_splitter import CharacterTextSplitter
+from langchain_community.vectorstores import FAISS
+from langchain.chains import RetrievalQA
+from langchain.llms import OpenAI
+from langchain.schema import Document
+
+def main(save_path = "assets/xiaozhan_emb"):
+    cat_f = "assets/xiaozhan_data/catering_end_category.csv"
+    desc_f = "assets/xiaozhan_data/catering_end_dish.csv"
+
+    df_cat = pd.read_csv(cat_f)
+    df_desc = pd.read_csv(desc_f)
+
+    id_desc_dic = {}
+    for _, (id, name, desc) in df_cat[["id", "name", "description"]].iterrows():
+        id_desc_dic[id] = f"{name}-{desc}"
+
+    df_desc["cat_desc"] = df_desc["category_id"].map(id_desc_dic)
+    
+    data = []
+    for _, (name, desc, px, cat_desc) in df_desc[["name", "description", "price", "cat_desc"]].iterrows():
+        sen = f"茶名称：{name}|茶描述：{desc}|价格{px}|饮品类:{cat_desc}"
+        data.append(sen)
+    
+
+    # texts = [Document(e) for e in data]
+    texts = data
+    embeddings = QwenEmbeddings(
+        api_key=os.environ.get("ALI_API_KEY")
+    )  # Needs OPENAI_API_KEY
+
+    if not osp.exists(save_path):
+        # --- STEP 2: Create vector store ---
+        # vectorstore = FAISS.from_documents(texts, embeddings)
+        out_emb = embeddings.batch_embed_documents(texts)
+        vectorstore = FAISS.from_embeddings(zip(texts, out_emb), embeddings)
+
+        # --- STEP 3: SAVE the FAISS index to local files ---
+        vectorstore.save_local(save_path)
+        print(f"✅ Saved FAISS index to: {save_path}")
+
+    # --- STEP 4: LOAD later from disk in a separate session ---
+    # (You can imagine this being a new Python script.)
+    loaded_vectorstore = FAISS.load_local(
+        folder_path=save_path,
+        embeddings=embeddings,
+        allow_dangerous_deserialization=True  # Required for LangChain >= 0.1.1
+    )
+    print("✅ Loaded FAISS index successfully!")
+
+    # --- STEP 5: Use the retriever/QA chain on the loaded store ---
+    retriever = loaded_vectorstore.as_retriever(search_kwargs={
+        "k":3
+    })
+
+    res = retriever.invoke("灯与尘")
+
+    for doc in res:
+        print(doc)
+        print("==============================================")
+
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file