修复CLI使用的是props字段，而你的脚本使用的是columns字段的问题确保正确 python 能正确加载

src/visualization/comparison.py

@@ -13,24 +13,30 @@ from typing import Optional, List, Dict, Any
 import numpy as np
 
 def launch_interactive_viewer(df: pd.DataFrame, text_column: str, port: int = 5055, host: str = "0.0.0.0"):
-    """使用Python API启动交互式服务器"""
+    """使用Python API启动交互式服务器
+    
+    Args:
+        df: 原始的pandas DataFrame，包含投影坐标和邻居信息（作为Python对象，而非JSON字符串）
+        text_column: 用作文本内容的列名
+        port: 启动服务器的端口
+        host: 服务器绑定的主机地址
+    """
     try:
         from embedding_atlas.server import make_server
         from embedding_atlas.data_source import DataSource
         from embedding_atlas.utils import Hasher
         import pathlib
         
-        # 创建metadata - 添加database配置
+        # 创建metadata - 使用与CLI一致的props字段结构
         metadata = {
-            "columns": {
+            "props": {
-                "id": "_row_index",
+                "data": {
-                "text": text_column,
+                    "id": "_row_index",
-                "embedding": {"x": "projection_x", "y": "projection_y"},
+                    "text": text_column,
-                "neighbors": "__neighbors"  # 使用双下划线
+                    "projection": {"x": "projection_x", "y": "projection_y"},
-            },
+                    "neighbors": "__neighbors"
-            "database": {
+                },
-                "type": "wasm",
+                "initialState": {"version": "0.0.0"}
-                "load": True  # 关键: 告诉前端加载数据
             }
         }
         
@@ -40,7 +46,8 @@ def launch_interactive_viewer(df: pd.DataFrame, text_column: str, port: int = 50
         hasher.update(metadata)
         identifier = hasher.hexdigest()
         
-        # 创建DataSource
+        # 创建DataSource - 直接使用传入的原始DataFrame
+        # 确保df['__neighbors'] 是 Python dict + numpy arrays 的结构，而不是 JSON 字符串
         dataset = DataSource(identifier, df, metadata)
         
         # 获取静态文件路径
@@ -49,7 +56,7 @@ def launch_interactive_viewer(df: pd.DataFrame, text_column: str, port: int = 50
             (pathlib.Path(embedding_atlas.__file__).parent / "static").resolve()
         )
         
-        # 创建并启动服务器
+        # 创建并启动服务器 - 使用默认的duckdb_uri参数(即"wasm")
         app = make_server(dataset, static_path=static_path)
         
         import uvicorn
@@ -115,38 +122,55 @@ def create_embedding_service(
         print(f"Error computing projections: {e}")
         return
     
-    # 4. 创建metadata (修复: 添加database配置和正确的neighbors列名)
+    # 4. 创建metadata (使用与CLI一致的props字段结构)
     metadata = {
-        "columns": {
+        "props": {
-            "id": "_row_index",
+            "data": {
-            "text": text_column,
+                "id": "_row_index",
-            "embedding": {
+                "text": text_column,
-                "x": "projection_x",
+                "projection": {
-                "y": "projection_y"
+                    "x": "projection_x",
+                    "y": "projection_y"
+                },
+                "neighbors": "__neighbors"
             },
-            "neighbors": "__neighbors"  # 修复: 使用双下划线
+            "initialState": {"version": "0.0.0"}
-        },
-        "database": {
-            "type": "wasm",
-            "load": True  # 修复: 添加database配置
         }
     }
     
-    # 5. 生成数据集标识符
+    # 5. 保存原始DataFrame用于DataSource
+    df_for_datasource = combined_df.copy()
+    
+    # 6. 转换neighbors列为JSON字符串(仅用于保存parquet)
+    if '__neighbors' in combined_df.columns:
+        import json
+        combined_df['__neighbors'] = combined_df['__neighbors'].apply(
+            lambda x: json.dumps({
+                'ids': x['ids'].tolist() if hasattr(x['ids'], 'tolist') else list(x['ids']),
+                'distances': x['distances'].tolist() if hasattr(x['distances'], 'tolist') else list(x['distances'])
+            }) if x is not None else None
+        )
+    
+    # 7. 保存parquet文件
+    parquet_path = "combined_dataset.parquet"
+    combined_df.to_parquet(parquet_path, index=False)
+    print(f"Dataset saved to {parquet_path}")
+    
+    # 8. 生成数据集标识符
     hasher = Hasher()
     hasher.update(["custom_dataset"])
     hasher.update(metadata)
     identifier = hasher.hexdigest()
     
-    # 6. 创建DataSource
+    # 9. 使用原始DataFrame创建DataSource
-    dataset = DataSource(identifier, combined_df, metadata)
+    dataset = DataSource(identifier, df_for_datasource, metadata)
     
-    # 7. 获取静态文件路径
+    # 10. 获取静态文件路径
     static_path = str(
         (pathlib.Path(embedding_atlas.__file__).parent / "static").resolve()
     )
     
-    # 8. 创建并启动服务器
+    # 11. 创建并启动服务器 - 使用默认的duckdb_uri参数(即"wasm")
     app = make_server(dataset, static_path=static_path)
     import uvicorn
     print(f"Starting interactive viewer on http://{host}:{port}")
@@ -298,14 +322,35 @@ def visualize_csv_comparison(
     
     print(f"Visualization saved to {output_path}")
     
+    # Keep a copy of the original DataFrame for the interactive viewer
+    # The viewer expects Python objects, not JSON strings
+    df_for_viewer = combined_df.copy()
+    
+    # Convert __neighbors column to JSON strings before saving
+    # This is needed because DuckDB-WASM expects JSON strings for complex data types
+    if '__neighbors' in combined_df.columns:
+        import json
+        # 为保存到 parquet 文件的数据转换为 JSON 字符串
+        combined_df_for_save = combined_df.copy()
+        combined_df_for_save['__neighbors'] = combined_df['__neighbors'].apply(
+            lambda x: json.dumps({
+                'ids': x['ids'].tolist() if hasattr(x['ids'], 'tolist') else list(x['ids']),
+                'distances': x['distances'].tolist() if hasattr(x['distances'], 'tolist') else list(x['distances'])
+            }) if x is not None else None
+        )
+    
     # Save dataset for interactive viewing
     parquet_path = "combined_dataset.parquet"
-    combined_df.to_parquet(parquet_path, index=False)
+    # 保存转换为 JSON 字符串的 DataFrame
+    if '__neighbors' in combined_df.columns:
+        combined_df_for_save.to_parquet(parquet_path, index=False)
+    else:
+        combined_df.to_parquet(parquet_path, index=False)
     print(f"Dataset saved to {parquet_path}")
     
-    # Launch interactive viewer
+    # Launch interactive viewer with original DataFrame (not converted to JSON)
     if launch_interactive:
-        launch_interactive_viewer(combined_df, text_column, port, host)
+        launch_interactive_viewer(df_for_viewer, text_column, port, host)
         print(f"Interactive viewer available at http://{host}:{port}")