Merge branch 'main' of https://github.com/xerrors/Yuxi-Know into main

Author: xerrors

docs/latest/changelog/roadmap.md

@@ -15,12 +15,12 @@
 - 同名文件处理逻辑：遇到同名文件则在上传区域提示，是否删除旧文件
 - conversation 待修改为异步的版本
 - DBManager 需要将数据库修改为异步的aiosqlite或者异步mysql，缓存使用Redis存储
-- agent 状态中的文件区域，新增可以下载
 
 ### Bugs
 - 部分异常状态下，智能体的模型名称出现重叠[#279](https://github.com/xerrors/Yuxi-Know/issues/279)
 - DeepSeek 官方接口适配会出现问题
 - 目前的知识库的图片存在公开访问风险
+- 深度分析智能体需要考虑上下文超限的问题
 
 ### 新增
 - 优化知识库详情页面，更加简洁清晰
@@ -34,6 +34,8 @@
 - 新增自定义模型支持、新增 dashscope rerank/embeddings 模型的支持
 - 新增文档解析的图片支持，已支持 MinerU Officical、Docs、Markdown Zip格式
 - 新增暗色模式支持并调整整体 UI（[#343](https://github.com/xerrors/Yuxi-Know/pull/343)）
+- agent 状态中的文件区域，新增可以下载
+- 移除 Chroma 的支持，当前版本标记为移除
 
 ### 修复
 - 修复重排序模型实际未生效的问题

server/routers/knowledge_router.py

@@ -87,10 +87,12 @@ async def create_database(
     """创建知识库"""
     logger.debug(
         f"Create database {database_name} with kb_type {kb_type}, "
-        f"additional_params {additional_params}, llm_info {llm_info}"
+        f"additional_params {additional_params}, llm_info {llm_info}, "
+        f"embed_model_name {embed_model_name}"
     )
     try:
         additional_params = {**(additional_params or {})}
+        additional_params["auto_generate_questions"] = False  # 默认不生成问题
 
         def normalize_reranker_config(kb: str, params: dict) -> None:
             reranker_cfg = params.get("reranker_config")
@@ -112,12 +114,12 @@ def normalize_reranker_config(kb: str, params: dict) -> None:
             if not isinstance(reranker_cfg, Mapping):
                 raise HTTPException(status_code=400, detail="reranker_config must be an object")
 
-            enabled = bool(reranker_cfg.get("enabled", False))
+            reranker_enabled = bool(reranker_cfg.get("enabled", False))
             model = (reranker_cfg.get("model") or "").strip()
             recall_top_k = max(1, int(reranker_cfg.get("recall_top_k", 50)))
             final_top_k = max(1, int(reranker_cfg.get("final_top_k", 10)))
 
-            if enabled:
+            if reranker_enabled:
                 if not model:
                     raise HTTPException(status_code=400, detail="reranker_config.model is required when enabled")
                 if model not in config.reranker_names:
@@ -132,7 +134,7 @@ def normalize_reranker_config(kb: str, params: dict) -> None:
                 model = model if model in config.reranker_names else ""
 
             params["reranker_config"] = {
-                "enabled": enabled,
+                "enabled": reranker_enabled,
                 "model": model,
                 "recall_top_k": recall_top_k,
                 "final_top_k": final_top_k,

src/agents/common/mcp_repos/arxiv-mcp-server

@@ -0,0 +1 @@
+Subproject commit 057e2000be7b56823239815b0fe7c7fc0dbced96

src/agents/common/mcp_repos/mcp-server-mysql

@@ -0,0 +1 @@
+Subproject commit 6a0367834ea0fb5e5c94b9711e3e2756966789ea

src/config/static/models.py

@@ -29,7 +29,7 @@ class EmbedModelInfo(BaseModel):
     dimension: int = Field(..., description="向量维度")
     base_url: str = Field(..., description="API 基础 URL")
     api_key: str = Field(..., description="API Key 或环境变量名")
-
+    model_id: str | None = Field(None, description="可选的模型 ID")
 
 class RerankerInfo(BaseModel):
     """重排序模型配置"""
@@ -158,42 +158,49 @@ class RerankerInfo(BaseModel):
 
 DEFAULT_EMBED_MODELS: dict[str, EmbedModelInfo] = {
     "siliconflow/BAAI/bge-m3": EmbedModelInfo(
+        model_id="siliconflow/BAAI/bge-m3",
         name="BAAI/bge-m3",
         dimension=1024,
         base_url="https://api.siliconflow.cn/v1/embeddings",
         api_key="SILICONFLOW_API_KEY",
     ),
     "siliconflow/Pro/BAAI/bge-m3": EmbedModelInfo(
+        model_id="siliconflow/Pro/BAAI/bge-m3",
         name="Pro/BAAI/bge-m3",
         dimension=1024,
         base_url="https://api.siliconflow.cn/v1/embeddings",
         api_key="SILICONFLOW_API_KEY",
     ),
     "siliconflow/Qwen/Qwen3-Embedding-0.6B": EmbedModelInfo(
+        model_id="siliconflow/Qwen/Qwen3-Embedding-0.6B",
         name="Qwen/Qwen3-Embedding-0.6B",
         dimension=1024,
         base_url="https://api.siliconflow.cn/v1/embeddings",
         api_key="SILICONFLOW_API_KEY",
     ),
     "vllm/Qwen/Qwen3-Embedding-0.6B": EmbedModelInfo(
+        model_id="vllm/Qwen/Qwen3-Embedding-0.6B",
         name="Qwen3-Embedding-0.6B",
         dimension=1024,
         base_url="http://localhost:8000/v1/embeddings",
         api_key="no_api_key",
     ),
     "ollama/nomic-embed-text": EmbedModelInfo(
+        model_id="ollama/nomic-embed-text",
         name="nomic-embed-text",
         dimension=768,
         base_url="http://localhost:11434/api/embed",
         api_key="no_api_key",
     ),
     "ollama/bge-m3": EmbedModelInfo(
+        model_id="ollama/bge-m3",
         name="bge-m3",
         dimension=1024,
         base_url="http://localhost:11434/api/embed",
         api_key="no_api_key",
     ),
     "dashscope/text-embedding-v4": EmbedModelInfo(
+        model_id="dashscope/text-embedding-v4",
         name="text-embedding-v4",
         dimension=1024,
         base_url="https://dashscope.aliyuncs.com/compatible-mode/v1/embeddings",

src/knowledge/implementations/milvus.py

@@ -7,6 +7,7 @@
 
 from pymilvus import Collection, CollectionSchema, DataType, FieldSchema, connections, db, utility
 
+from src import config
 from src.knowledge.base import KnowledgeBase
 from src.knowledge.indexing import process_file_to_markdown
 from src.knowledge.utils.kb_utils import (
@@ -91,10 +92,14 @@ async def _create_kb_instance(self, db_id: str, kb_config: dict) -> Any:
         """创建 Milvus 集合"""
         logger.info(f"Creating Milvus collection for {db_id}")
 
-        if db_id not in self.databases_meta:
+        if not (metadata := self.databases_meta.get(db_id)):
             raise ValueError(f"Database {db_id} not found")
 
-        embed_info = self.databases_meta[db_id].get("embed_info", {})
+        # embed_info = metadata.get("embed_info", {})
+        if not (embed_info := metadata.get("embed_info")):
+            logger.error(f"Embedding info not found for database {db_id}, using default model")
+            embed_info = config.embed_model_names[config.embed_model]
+
         collection_name = db_id
 
         try:
@@ -117,8 +122,8 @@ async def _create_kb_instance(self, db_id: str, kb_config: dict) -> Any:
 
         except Exception:
             # 创建新集合
-            embedding_dim = getattr(embed_info, "dimension", 1024) if embed_info else 1024
-            model_name = getattr(embed_info, "name", "default") if embed_info else "default"
+            embedding_dim = embed_info.get("dimension", 1024)
+            model_name = embed_info.get("name", "default")
 
             # 定义集合Schema
             fields = [
@@ -142,7 +147,7 @@ async def _create_kb_instance(self, db_id: str, kb_config: dict) -> Any:
             index_params = {"metric_type": "COSINE", "index_type": "IVF_FLAT", "params": {"nlist": 1024}}
             collection.create_index("embedding", index_params)
 
-            logger.info(f"Created new Milvus collection: {collection_name}")
+            logger.info(f"Created new Milvus collection: {collection_name}: {model_name=}, {embedding_dim=}")
 
         return collection
 
@@ -154,25 +159,29 @@ async def _initialize_kb_instance(self, instance: Any) -> None:
         except Exception as e:
             logger.warning(f"Failed to load collection into memory: {e}")
 
-    def _get_async_embedding_function(self, embed_info: dict):
+    def _get_async_embedding(self, embed_info: dict):
         """获取 embedding 函数"""
+        # 检查是否有 model_id 字段，优先使用 select_embedding_model
+        if embed_info and "model_id" in embed_info:
+            from src.models.embed import select_embedding_model
+            return select_embedding_model(embed_info["model_id"])
+
+        # 使用原有的逻辑（兼容模式））
         config_dict = get_embedding_config(embed_info)
-        embedding_model = OtherEmbedding(
+        return OtherEmbedding(
             model=config_dict.get("model"),
             base_url=config_dict.get("base_url"),
             api_key=config_dict.get("api_key"),
         )
 
+    def _get_async_embedding_function(self, embed_info: dict):
+        """获取 embedding 函数"""
+        embedding_model = self._get_async_embedding(embed_info)
         return partial(embedding_model.abatch_encode, batch_size=40)
 
     def _get_embedding_function(self, embed_info: dict):
         """获取 embedding 函数"""
-        config_dict = get_embedding_config(embed_info)
-        embedding_model = OtherEmbedding(
-            model=config_dict.get("model"),
-            base_url=config_dict.get("base_url"),
-            api_key=config_dict.get("api_key"),
-        )
+        embedding_model = self._get_async_embedding(embed_info)
 
         return partial(embedding_model.batch_encode, batch_size=40)
 

src/knowledge/manager.py

@@ -246,16 +246,12 @@ async def create_database(
         db_id = db_info["db_id"]
 
         async with self._metadata_lock:
-            # 准备 additional_params，包含 auto_generate_questions
-            saved_params = kwargs.copy()
-            saved_params["auto_generate_questions"] = False
-
             self.global_databases_meta[db_id] = {
                 "name": database_name,
                 "description": description,
                 "kb_type": kb_type,
                 "created_at": utc_isoformat(),
-                "additional_params": saved_params,
+                "additional_params": kwargs.copy(),
             }
             self._save_global_metadata()
 

src/knowledge/utils/kb_utils.py

@@ -247,15 +247,23 @@ def get_embedding_config(embed_info: dict) -> dict:
 
     try:
         if embed_info:
-            # 处理 embed_info 可能是字典或 EmbedModelInfo 对象的情况
-            if hasattr(embed_info, "name"):
+            # 优先检查是否有 model_id 字段
+            if "model_id" in embed_info:
+                from src.models.embed import select_embedding_model
+
+                model = select_embedding_model(embed_info["model_id"])
+                config_dict["model"] = model.model
+                config_dict["api_key"] = model.api_key
+                config_dict["base_url"] = model.base_url
+                config_dict["dimension"] = getattr(model, "dimension", 1024)
+            elif hasattr(embed_info, "name"):
                 # EmbedModelInfo 对象
                 config_dict["model"] = embed_info.name
                 config_dict["api_key"] = os.getenv(embed_info.api_key) or embed_info.api_key
                 config_dict["base_url"] = embed_info.base_url
                 config_dict["dimension"] = embed_info.dimension
             else:
-                # 字典形式
+                # 字典形式（保持向后兼容）
                 config_dict["model"] = embed_info["name"]
                 config_dict["api_key"] = os.getenv(embed_info["api_key"]) or embed_info["api_key"]
                 config_dict["base_url"] = embed_info["base_url"]

src/models/embed.py

@@ -11,7 +11,7 @@
 
 
 class BaseEmbeddingModel(ABC):
-    def __init__(self, model=None, name=None, dimension=None, url=None, base_url=None, api_key=None):
+    def __init__(self, model=None, name=None, dimension=None, url=None, base_url=None, api_key=None, model_id=None):
         """
         Args:
             model: 模型名称，冗余设计，同name
@@ -140,6 +140,7 @@ async def aencode(self, message: list[str] | str) -> list[list[float]]:
         payload = {"model": self.model, "input": message}
         async with httpx.AsyncClient() as client:
             try:
+                print(f"\n\n\nOllama Embedding request: {payload}\n\n\n")
                 response = await client.post(self.base_url, json=payload, timeout=60)
                 response.raise_for_status()
                 result = response.json()