Feat: Support tool calling in Generate component (#7572)

### What problem does this PR solve?

Hello, our use case requires LLM agent to invoke some tools, so I made a
simple implementation here.

This PR does two things:

1. A simple plugin mechanism based on `pluginlib`:

This mechanism lives in the `plugin` directory. It will only load
plugins from `plugin/embedded_plugins` for now.

A sample plugin `bad_calculator.py` is placed in
`plugin/embedded_plugins/llm_tools`, it accepts two numbers `a` and `b`,
then give a wrong result `a + b + 100`.

In the future, it can load plugins from external location with little
code change.

Plugins are divided into different types. The only plugin type supported
in this PR is `llm_tools`, which must implement the `LLMToolPlugin`
class in the `plugin/llm_tool_plugin.py`.
More plugin types can be added in the future.

2. A tool selector in the `Generate` component:

Added a tool selector to select one or more tools for LLM:


![image](https://github.com/user-attachments/assets/74a21fdf-9333-4175-991b-43df6524c5dc)

And with the `bad_calculator` tool, it results this with the `qwen-max`
model:


![image](https://github.com/user-attachments/assets/93aff9c4-8550-414a-90a2-1a15a5249d94)


### Type of change

- [ ] Bug Fix (non-breaking change which fixes an issue)
- [x] New Feature (non-breaking change which adds functionality)
- [ ] Documentation Update
- [ ] Refactoring
- [ ] Performance Improvement
- [ ] Other (please describe):

Co-authored-by: Yingfeng <yingfeng.zhang@gmail.com>

Dockerfile

@@ -199,6 +199,7 @@ COPY graphrag graphrag
 COPY agentic_reasoning agentic_reasoning
 COPY pyproject.toml uv.lock ./
 COPY mcp mcp
+COPY plugin plugin
 
 COPY docker/service_conf.yaml.template ./conf/service_conf.yaml.template
 COPY docker/entrypoint.sh ./

Dockerfile.scratch.oc9

@@ -33,6 +33,7 @@ ADD ./rag ./rag
 ADD ./requirements.txt ./requirements.txt
 ADD ./agent ./agent
 ADD ./graphrag ./graphrag
+ADD ./plugin ./plugin
 
 RUN dnf install -y openmpi openmpi-devel python3-openmpi
 ENV C_INCLUDE_PATH /usr/include/openmpi-x86_64:$C_INCLUDE_PATH

agent/component/generate.py

@@ -16,15 +16,29 @@
 import json
 import re
 from functools import partial
+from typing import Any
 import pandas as pd
 from api.db import LLMType
 from api.db.services.conversation_service import structure_answer
 from api.db.services.llm_service import LLMBundle
 from api import settings
 from agent.component.base import ComponentBase, ComponentParamBase
+from plugin import GlobalPluginManager
+from plugin.llm_tool_plugin import llm_tool_metadata_to_openai_tool
+from rag.llm.chat_model import ToolCallSession
 from rag.prompts import message_fit_in
 
 
+class LLMToolPluginCallSession(ToolCallSession):
+    def tool_call(self, name: str, arguments: dict[str, Any]) -> str:
+        tool = GlobalPluginManager.get_llm_tool_by_name(name)
+
+        if tool is None:
+            raise ValueError(f"LLM tool {name} does not exist")
+
+        return tool().invoke(**arguments)
+
+
 class GenerateParam(ComponentParamBase):
     """
     Define the Generate component parameters.
@@ -41,6 +55,7 @@ class GenerateParam(ComponentParamBase):
         self.frequency_penalty = 0
         self.cite = True
         self.parameters = []
+        self.llm_enabled_tools = []
 
     def check(self):
         self.check_decimal_float(self.temperature, "[Generate] Temperature")
@@ -133,6 +148,15 @@ class Generate(ComponentBase):
 
     def _run(self, history, **kwargs):
         chat_mdl = LLMBundle(self._canvas.get_tenant_id(), LLMType.CHAT, self._param.llm_id)
+
+        if len(self._param.llm_enabled_tools) > 0:
+            tools = GlobalPluginManager.get_llm_tools_by_names(self._param.llm_enabled_tools)
+
+            chat_mdl.bind_tools(
+                LLMToolPluginCallSession(),
+                [llm_tool_metadata_to_openai_tool(t.get_metadata()) for t in tools]
+            )
+
         prompt = self._param.prompt
 
         retrieval_res = []

api/apps/plugin_app.py

@@ -0,0 +1,12 @@
+from flask import Response
+from flask_login import login_required
+from api.utils.api_utils import get_json_result
+from plugin import GlobalPluginManager
+
+@manager.route('/llm_tools', methods=['GET'])  # noqa: F821
+@login_required
+def llm_tools() -> Response:
+    tools = GlobalPluginManager.get_llm_tools()
+    tools_metadata = [t.get_metadata() for t in tools]
+
+    return get_json_result(data=tools_metadata)

api/db/services/llm_service.py

@@ -226,6 +226,7 @@ class LLMBundle:
 
     def bind_tools(self, toolcall_session, tools):
         if not self.is_tools:
+            logging.warning(f"Model {self.llm_name} does not support tool call, but you have assigned one or more tools to it!")
             return
         self.mdl.bind_tools(toolcall_session, tools)
 

api/ragflow_server.py

@@ -19,6 +19,7 @@
 # beartype_all(conf=BeartypeConf(violation_type=UserWarning))    # <-- emit warnings from all code
 
 from api.utils.log_utils import initRootLogger
+from plugin import GlobalPluginManager
 initRootLogger("ragflow_server")
 
 import logging
@@ -119,6 +120,8 @@ if __name__ == '__main__':
     RuntimeConfig.init_env()
     RuntimeConfig.init_config(JOB_SERVER_HOST=settings.HOST_IP, HTTP_PORT=settings.HOST_PORT)
 
+    GlobalPluginManager.load_plugins()
+
     signal.signal(signal.SIGINT, signal_handler)
     signal.signal(signal.SIGTERM, signal_handler)
 

plugin/README.md

@@ -0,0 +1,97 @@
+# Plugins
+
+This directory contains the plugin mechanism for RAGFlow.
+
+RAGFlow will load plugins from `embedded_plugins` subdirectory recursively.
+
+## Supported plugin types
+
+Currently, the only supported plugin type is `llm_tools`.
+
+- `llm_tools`: A tool for LLM to call.
+
+## How to add a plugin
+
+Add a LLM tool plugin is simple: create a plugin file, put a class inherits the `LLMToolPlugin` class in it, then implement the `get_metadata` and the `invoke` methods.
+
+- `get_metadata` method: This method returns a `LLMToolMetadata` object, which contains the description of this tool.
+The description will be provided to LLM, and the RAGFlow web frontend for displaying.
+
+- `invoke` method: This method accepts parameters generated by LLM, and return a `str` containing the tool execution result.
+All the execution logic of this tool should go into this method.
+
+When you start RAGFlow, you can see your plugin was loaded in the log:
+
+```
+2025-05-15 19:29:08,959 INFO     34670 Recursively importing plugins from path `/some-path/ragflow/plugin/embedded_plugins`
+2025-05-15 19:29:08,960 INFO     34670 Loaded llm_tools plugin BadCalculatorPlugin version 1.0.0
+```
+
+Or it may contain some errors for you to fix your plugin.
+
+### Demo
+
+We will demonstrate how to add a plugin with a calculator tool which will give wrong answers.
+
+First, create a plugin file `bad_calculator.py` under the `embedded_plugins/llm_tools` directory.
+
+Then, we create a `BadCalculatorPlugin` class, extending the `LLMToolPlugin` base class:
+
+```python
+class BadCalculatorPlugin(LLMToolPlugin):
+    _version_ = "1.0.0"
+```
+
+The `_version_` field is required, which specifies the version of the plugin.
+
+Our calculator has two numbers `a` and `b` as inputs, so we add a `invoke` method to our `BadCalculatorPlugin` class:
+
+```python
+def invoke(self, a: int, b: int) -> str:
+    return str(a + b + 100)
+```
+
+The `invoke` method will be called by LLM. It can have many parameters, but the return type must be a `str`.
+
+Finally, we have to add a `get_metadata` method, to tell LLM how to use our `bad_calculator`:
+
+```python
+@classmethod
+def get_metadata(cls) -> LLMToolMetadata:
+    return {
+        # Name of this tool, providing to LLM
+        "name": "bad_calculator",
+        # Display name of this tool, providing to RAGFlow frontend
+        "displayName": "$t:bad_calculator.name",
+        # Description of the usage of this tool, providing to LLM
+        "description": "A tool to calculate the sum of two numbers (will give wrong answer)",
+        # Description of this tool, providing to RAGFlow frontend
+        "displayDescription": "$t:bad_calculator.description",
+        # Parameters of this tool
+        "parameters": {
+            # The first parameter - a
+            "a": {
+                # Parameter type, options are: number, string, or whatever the LLM can recognise
+                "type": "number",
+                # Description of this parameter, providing to LLM
+                "description": "The first number",
+                # Description of this parameter, provding to RAGFlow frontend
+                "displayDescription": "$t:bad_calculator.params.a",
+                # Whether this parameter is required
+                "required": True
+            },
+            # The second parameter - b
+            "b": {
+                "type": "number",
+                "description": "The second number",
+                "displayDescription": "$t:bad_calculator.params.b",
+                "required": True
+            }
+        }
+```
+
+The `get_metadata` method is a `classmethod`. It will provide the description of this tool to LLM.
+
+The fields starts with `display` can use a special notation: `$t:xxx`, which will use the i18n mechanism in the RAGFlow frontend, getting text from the `llmTools` category. The frontend will display what you put here if you don't use this notation.
+
+Now our tool is ready. You can select it in the `Generate` component and try it out.

plugin/README_zh.md

@@ -0,0 +1,98 @@
+# 插件
+
+这个文件夹包含了RAGFlow的插件机制。
+
+RAGFlow将会从`embedded_plugins`子文件夹中递归加载所有的插件。
+
+## 支持的插件类型
+
+目前，唯一支持的插件类型是`llm_tools`。
+
+- `llm_tools`：用于供LLM进行调用的工具。
+
+## 如何添加一个插件
+
+添加一个LLM工具插件是很简单的：创建一个插件文件，向其中放一个继承自`LLMToolPlugin`的类，再实现它的`get_metadata`和`invoke`方法即可。
+
+- `get_metadata`方法：这个方法返回一个`LLMToolMetadata`对象，其中包含了对这个工具的描述。
+这些描述信息将被提供给LLM进行调用，和RAGFlow的Web前端用作展示。
+
+- `invoke`方法：这个方法接受LLM生成的参数，并且返回一个`str`对象，其中包含了这个工具的执行结果。
+这个工具的所有执行逻辑都应当放到这个方法里。
+
+当你启动RAGFlow时，你会在日志中看见你的插件被加载了：
+
+```
+2025-05-15 19:29:08,959 INFO     34670 Recursively importing plugins from path `/some-path/ragflow/plugin/embedded_plugins`
+2025-05-15 19:29:08,960 INFO     34670 Loaded llm_tools plugin BadCalculatorPlugin version 1.0.0
+```
+
+也可能会报错，这时就需要根据报错对你的插件进行修复。
+
+### 示例
+
+我们将会添加一个会给出错误答案的计算器工具，来演示添加插件的过程。
+
+首先，在`embedded_plugins/llm_tools`文件夹下创建一个插件文件`bad_calculator.py`。
+
+接下来，我们创建一个`BadCalculatorPlugin`类，继承基类`LLMToolPlugin`：
+
+```python
+class BadCalculatorPlugin(LLMToolPlugin):
+    _version_ = "1.0.0"
+```
+
+`_version_`字段是必填的，用于指定这个插件的版本号。
+
+我们的计算器拥有两个输入字段`a`和`b`，所以我们添加如下的`invoke`方法到`BadCalculatorPlugin`类中：
+
+```python
+def invoke(self, a: int, b: int) -> str:
+    return str(a + b + 100)
+```
+
+`invoke`方法将会被LLM所调用。这个方法可以有许多参数，但它必须返回一个`str`。
+
+最后，我们需要添加一个`get_metadata`方法，来告诉LLM怎样使用我们的`bad_calculator`工具：
+
+```python
+@classmethod
+def get_metadata(cls) -> LLMToolMetadata:
+    return {
+        # 这个工具的名称，会提供给LLM
+        "name": "bad_calculator",
+        # 这个工具的展示名称，会提供给RAGFlow的Web前端
+        "displayName": "$t:bad_calculator.name",
+        # 这个工具的用法描述，会提供给LLM
+        "description": "A tool to calculate the sum of two numbers (will give wrong answer)",
+        # 这个工具的描述，会提供给RAGFlow的Web前端
+        "displayDescription": "$t:bad_calculator.description",
+        # 这个工具的参数
+        "parameters": {
+            # 第一个参数 - a
+            "a": {
+                # 参数类型，选项为：number, string, 或者LLM可以识别的任何类型
+                "type": "number",
+                # 这个参数的描述，会提供给LLM
+                "description": "The first number",
+                # 这个参数的描述，会提供给RAGFlow的Web前端
+                "displayDescription": "$t:bad_calculator.params.a",
+                # 这个参数是否是必填的
+                "required": True
+            },
+            # 第二个参数 - b
+            "b": {
+                "type": "number",
+                "description": "The second number",
+                "displayDescription": "$t:bad_calculator.params.b",
+                "required": True
+            }
+        }
+```
+
+`get_metadata`方法是一个`classmethod`。它会把这个工具的描述提供给LLM。
+
+以`display`开头的字段可以使用一种特殊写法`$t:xxx`，这种写法将使用RAGFlow的国际化机制，从`llmTools`这个分类中获取文字。如果你不使用这种写法，那么前端将会显示此处的原始内容。
+
+现在，我们的工具已经做好了，你可以在`生成回答`组件中选择这个工具来尝试一下。
+

plugin/__init__.py

@@ -0,0 +1,3 @@
+from .plugin_manager import PluginManager
+
+GlobalPluginManager = PluginManager()

plugin/common.py

@@ -0,0 +1 @@
+PLUGIN_TYPE_LLM_TOOLS = "llm_tools"

plugin/embedded_plugins/llm_tools/bad_calculator.py

@@ -0,0 +1,37 @@
+import logging
+from plugin.llm_tool_plugin import LLMToolMetadata, LLMToolPlugin
+
+
+class BadCalculatorPlugin(LLMToolPlugin):
+    """
+    A sample LLM tool plugin, will add two numbers with 100.
+    It only present for demo purpose. Do not use it in production.
+    """
+    _version_ = "1.0.0"
+
+    @classmethod
+    def get_metadata(cls) -> LLMToolMetadata:
+        return {
+            "name": "bad_calculator",
+            "displayName": "$t:bad_calculator.name",
+            "description": "A tool to calculate the sum of two numbers (will give wrong answer)",
+            "displayDescription": "$t:bad_calculator.description",
+            "parameters": {
+                "a": {
+                    "type": "number",
+                    "description": "The first number",
+                    "displayDescription": "$t:bad_calculator.params.a",
+                    "required": True
+                },
+                "b": {
+                    "type": "number",
+                    "description": "The second number",
+                    "displayDescription": "$t:bad_calculator.params.b",
+                    "required": True
+                }
+            }
+        }
+
+    def invoke(self, a: int, b: int) -> str:
+        logging.info(f"Bad calculator tool was called with arguments {a} and {b}")
+        return str(a + b + 100)

plugin/llm_tool_plugin.py

@@ -0,0 +1,51 @@
+from typing import Any, TypedDict
+import pluginlib
+
+from .common import PLUGIN_TYPE_LLM_TOOLS
+
+
+class LLMToolParameter(TypedDict):
+    type: str
+    description: str
+    displayDescription: str
+    required: bool
+
+
+class LLMToolMetadata(TypedDict):
+    name: str
+    displayName: str
+    description: str
+    displayDescription: str
+    parameters: dict[str, LLMToolParameter]
+
+
+@pluginlib.Parent(PLUGIN_TYPE_LLM_TOOLS)
+class LLMToolPlugin:
+    @classmethod
+    @pluginlib.abstractmethod
+    def get_metadata(cls) -> LLMToolMetadata:
+        pass
+
+    def invoke(self, **kwargs) -> str:
+        raise NotImplementedError
+
+
+def llm_tool_metadata_to_openai_tool(llm_tool_metadata: LLMToolMetadata) -> dict[str, Any]:
+    return {
+        "type": "function",
+        "function": {
+            "name": llm_tool_metadata["name"],
+            "description": llm_tool_metadata["description"],
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    k: {
+                        "type": p["type"],
+                        "description": p["description"]
+                    }
+                    for k, p in llm_tool_metadata["parameters"].items()
+                },
+                "required": [k for k, p in llm_tool_metadata["parameters"].items() if p["required"]]
+            }
+        }
+    }

plugin/plugin_manager.py

@@ -0,0 +1,45 @@
+import logging
+import os
+from pathlib import Path
+import pluginlib
+
+from .common import PLUGIN_TYPE_LLM_TOOLS
+
+from .llm_tool_plugin import LLMToolPlugin
+
+
+class PluginManager:
+    _llm_tool_plugins: dict[str, LLMToolPlugin]
+
+    def __init__(self) -> None:
+        self._llm_tool_plugins = {}
+
+    def load_plugins(self) -> None:
+        loader = pluginlib.PluginLoader(
+            paths=[str(Path(os.path.dirname(__file__), "embedded_plugins"))]
+        )
+        
+        for type, plugins in loader.plugins.items():
+            for name, plugin in plugins.items():
+                logging.info(f"Loaded {type} plugin {name} version {plugin.version}")
+
+                if type == PLUGIN_TYPE_LLM_TOOLS:
+                    metadata = plugin.get_metadata()
+                    self._llm_tool_plugins[metadata["name"]] = plugin
+
+    def get_llm_tools(self) -> list[LLMToolPlugin]:
+        return list(self._llm_tool_plugins.values())
+
+    def get_llm_tool_by_name(self, name: str) -> LLMToolPlugin | None:
+        return self._llm_tool_plugins.get(name)
+
+    def get_llm_tools_by_names(self, tool_names: list[str]) -> list[LLMToolPlugin]:
+        results = []
+
+        for name in tool_names:
+            plugin = self._llm_tool_plugins.get(name)
+
+            if plugin is not None:
+                results.append(plugin)
+
+        return results

pyproject.toml

@@ -125,7 +125,8 @@ dependencies = [
     "langfuse>=2.60.0",
     "debugpy>=1.8.13",
     "mcp>=1.6.0",
-    "opensearch-py==2.7.1"
+    "opensearch-py==2.7.1",
+    "pluginlib==0.9.4",
 ]
 
 [project.optional-dependencies]

rag/llm/chat_model.py

@@ -21,6 +21,7 @@ import random
 import re
 import time
 from abc import ABC
+from typing import Any, Protocol
 
 import openai
 import requests
@@ -51,6 +52,10 @@ LENGTH_NOTIFICATION_CN = "······\n由于大模型的上下文窗口大小
 LENGTH_NOTIFICATION_EN = "...\nThe answer is truncated by your chosen LLM due to its limitation on context length."
 
 
+class ToolCallSession(Protocol):
+    def tool_call(self, name: str, arguments: dict[str, Any]) -> str: ...
+
+
 class Base(ABC):
     def __init__(self, key, model_name, base_url):
         timeout = int(os.environ.get("LM_TIMEOUT_SECONDS", 600))
@@ -251,10 +256,8 @@ class Base(ABC):
 
                             if index not in final_tool_calls:
                                 final_tool_calls[index] = tool_call
-
+                            else:
-                            final_tool_calls[index].function.arguments += tool_call.function.arguments
+                                final_tool_calls[index].function.arguments += tool_call.function.arguments
-                        if resp.choices[0].finish_reason != "stop":
-                            continue
                     else:
                         if not resp.choices:
                             continue
@@ -276,58 +279,57 @@ class Base(ABC):
                         else:
                             total_tokens += tol
 
-                        finish_reason = resp.choices[0].finish_reason
+                    finish_reason = resp.choices[0].finish_reason
-                        if finish_reason == "tool_calls" and final_tool_calls:
+                    if finish_reason == "tool_calls" and final_tool_calls:
-                            for tool_call in final_tool_calls.values():
+                        for tool_call in final_tool_calls.values():
-                                name = tool_call.function.name
+                            name = tool_call.function.name
-                                try:
+                            try:
-                                    if name == "get_current_weather":
+                                args = json.loads(tool_call.function.arguments)
-                                        args = json.loads('{"location":"Shanghai"}')
+                            except Exception as e:
-                                    else:
+                                logging.exception(msg=f"Wrong JSON argument format in LLM tool call response: {tool_call}")
-                                        args = json.loads(tool_call.function.arguments)
+                                yield ans + "\n**ERROR**: " + str(e)
-                                except Exception:
+                                finish_completion = True
-                                    continue
+                                break
-                                # args = json.loads(tool_call.function.arguments)
+
-                                tool_response = self.toolcall_session.tool_call(name, args)
+                            tool_response = self.toolcall_session.tool_call(name, args)
-                                history.append(
+                            history.append(
-                                    {
+                                {
-                                        "role": "assistant",
+                                    "role": "assistant",
-                                        "refusal": "",
+                                    "tool_calls": [
-                                        "content": "",
+                                        {
-                                        "audio": "",
+                                            "index": tool_call.index,
-                                        "function_call": "",
+                                            "id": tool_call.id,
-                                        "tool_calls": [
+                                            "function": {
-                                            {
+                                                "name": tool_call.function.name,
-                                                "index": tool_call.index,
+                                                "arguments": tool_call.function.arguments,
-                                                "id": tool_call.id,
-                                                "function": tool_call.function,
-                                                "type": "function",
                                             },
-                                        ],
+                                            "type": "function",
-                                    }
+                                        },
-                                )
+                                    ],
-                                # if tool_response.choices[0].finish_reason == "length":
+                                }
-                                #     if is_chinese(ans):
+                            )
-                                #         ans += LENGTH_NOTIFICATION_CN
+                            # if tool_response.choices[0].finish_reason == "length":
-                                #     else:
+                            #     if is_chinese(ans):
-                                #         ans += LENGTH_NOTIFICATION_EN
+                            #         ans += LENGTH_NOTIFICATION_CN
-                                #     return ans, total_tokens + self.total_token_count(tool_response)
+                            #     else:
-                                history.append({"role": "tool", "tool_call_id": tool_call.id, "content": str(tool_response)})
+                            #         ans += LENGTH_NOTIFICATION_EN
-                            final_tool_calls = {}
+                            #     return ans, total_tokens + self.total_token_count(tool_response)
-                            response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, tools=tools, **gen_conf)
+                            history.append({"role": "tool", "tool_call_id": tool_call.id, "content": str(tool_response)})
-                            continue
+                        final_tool_calls = {}
-                        if finish_reason == "length":
+                        response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, tools=tools, **gen_conf)
-                            if is_chinese(ans):
-                                ans += LENGTH_NOTIFICATION_CN
-                            else:
-                                ans += LENGTH_NOTIFICATION_EN
-                            return ans, total_tokens + self.total_token_count(resp)
-                        if finish_reason == "stop":
-                            finish_completion = True
-                            yield ans
-                            break
-                        yield ans
                         continue
+                    if finish_reason == "length":
+                        if is_chinese(ans):
+                            ans += LENGTH_NOTIFICATION_CN
+                        else:
+                            ans += LENGTH_NOTIFICATION_EN
+                        return ans, total_tokens
+                    if finish_reason == "stop":
+                        finish_completion = True
+                        yield ans
+                        break
+                    yield ans
+                    continue
 
         except openai.APIError as e:
             yield ans + "\n**ERROR**: " + str(e)
@@ -854,6 +856,14 @@ class ZhipuChat(Base):
         except Exception as e:
             return "**ERROR**: " + str(e), 0
 
+    def chat_with_tools(self, system: str, history: list, gen_conf: dict):
+        if "presence_penalty" in gen_conf:
+            del gen_conf["presence_penalty"]
+        if "frequency_penalty" in gen_conf:
+            del gen_conf["frequency_penalty"]
+
+        return super().chat_with_tools(system, history, gen_conf)
+
     def chat_streamly(self, system, history, gen_conf):
         if system:
             history.insert(0, {"role": "system", "content": system})
@@ -886,6 +896,14 @@ class ZhipuChat(Base):
 
         yield tk_count
 
+    def chat_streamly_with_tools(self, system: str, history: list, gen_conf: dict):
+        if "presence_penalty" in gen_conf:
+            del gen_conf["presence_penalty"]
+        if "frequency_penalty" in gen_conf:
+            del gen_conf["frequency_penalty"]
+
+        return super().chat_streamly_with_tools(system, history, gen_conf)
+
 
 class OllamaChat(Base):
     def __init__(self, key, model_name, **kwargs):

uv.lock

@@ -3952,6 +3952,18 @@ wheels = [
     { url = "https://mirrors.aliyun.com/pypi/packages/88/5f/e351af9a41f866ac3f1fac4ca0613908d9a41741cfcf2228f4ad853b697d/pluggy-1.5.0-py3-none-any.whl", hash = "sha256:44e1ad92c8ca002de6377e165f3e0f1be63266ab4d554740532335b9d75ea669" },
 ]
 
+[[package]]
+name = "pluginlib"
+version = "0.9.4"
+source = { registry = "https://mirrors.aliyun.com/pypi/simple" }
+dependencies = [
+    { name = "setuptools" },
+]
+sdist = { url = "https://mirrors.aliyun.com/pypi/packages/58/38/ca974ba2d8ccc7954d8ccb0394cce184ac6269bd1fbfe06f70a0da3c8946/pluginlib-0.9.4.tar.gz", hash = "sha256:88727037138f759a3952f6391ae3751536f04ad8be6023607620ea49695a3a83" }
+wheels = [
+    { url = "https://mirrors.aliyun.com/pypi/packages/b0/b5/c869b3d2ed1613afeb02c635be11f5d35fa5b2b665f4d059cfe5b8e82941/pluginlib-0.9.4-py2.py3-none-any.whl", hash = "sha256:d4cfb7d74a6d2454e256b6512fbc4bc2dd7620cb7764feb67331ef56ce4b33f2" },
+]
+
 [[package]]
 name = "polars-lts-cpu"
 version = "1.9.0"
@@ -4872,6 +4884,7 @@ dependencies = [
     { name = "pdfplumber" },
     { name = "peewee" },
     { name = "pillow" },
+    { name = "pluginlib" },
     { name = "protobuf" },
     { name = "psycopg2-binary" },
     { name = "pyclipper" },
@@ -5009,6 +5022,7 @@ requires-dist = [
     { name = "pdfplumber", specifier = "==0.10.4" },
     { name = "peewee", specifier = "==3.17.1" },
     { name = "pillow", specifier = "==10.4.0" },
+    { name = "pluginlib", specifier = "==0.9.4" },
     { name = "protobuf", specifier = "==5.27.2" },
     { name = "psycopg2-binary", specifier = "==2.9.9" },
     { name = "pyclipper", specifier = "==1.3.0.post5" },

web/src/components/llm-select/index.tsx

@@ -11,19 +11,31 @@ import { Select, SelectTrigger, SelectValue } from '../ui/select';
 interface IProps {
   id?: string;
   value?: string;
-  onChange?: (value: string) => void;
+  onInitialValue?: (value: string, option: any) => void;
+  onChange?: (value: string, option: any) => void;
   disabled?: boolean;
 }
 
-const LLMSelect = ({ id, value, onChange, disabled }: IProps) => {
+const LLMSelect = ({ id, value, onInitialValue, onChange, disabled }: IProps) => {
   const modelOptions = useComposeLlmOptionsByModelTypes([
     LlmModelType.Chat,
     LlmModelType.Image2text,
   ]);
 
+  if (onInitialValue && value) {
+    for (const modelOption of modelOptions) {
+      for (const option of modelOption.options) {
+        if (option.value === value) {
+          onInitialValue(value, option);
+          break;
+        }
+      }
+    }
+  }  
+
   const content = (
     <div style={{ width: 400 }}>
-      <LlmSettingItems
+      <LlmSettingItems onChange={onChange}
         formItemLayout={{ labelCol: { span: 10 }, wrapperCol: { span: 14 } }}
       ></LlmSettingItems>
     </div>

web/src/components/llm-setting-items/index.tsx

@@ -16,9 +16,10 @@ interface IProps {
   prefix?: string;
   formItemLayout?: any;
   handleParametersChange?(value: ModelVariableType): void;
+  onChange?(value: string, option: any): void;
 }
 
-const LlmSettingItems = ({ prefix, formItemLayout = {} }: IProps) => {
+const LlmSettingItems = ({ prefix, formItemLayout = {}, onChange }: IProps) => {
   const form = Form.useFormInstance();
   const { t } = useTranslate('chat');
   const parameterOptions = Object.values(ModelVariableType).map((x) => ({
@@ -58,6 +59,7 @@ const LlmSettingItems = ({ prefix, formItemLayout = {} }: IProps) => {
           options={modelOptions}
           showSearch
           popupMatchSelectWidth={false}
+          onChange={onChange}
         />
       </Form.Item>
       <div className="border rounded-md">

web/src/components/llm-tools-select.tsx

@@ -0,0 +1,51 @@
+import { useTranslate } from '@/hooks/common-hooks';
+import { useLlmToolsList } from '@/hooks/plugin-hooks';
+import { Select, Space } from 'antd';
+
+interface IProps {
+  value?: string;
+  onChange?: (value: string) => void;
+  disabled?: boolean;
+}
+
+const LLMToolsSelect = ({ value, onChange, disabled }: IProps) => {
+  const { t } = useTranslate("llmTools");
+  const tools = useLlmToolsList();
+
+  function wrapTranslation(text: string): string {
+    if (!text) {
+      return text;
+    }
+
+    if (text.startsWith("$t:")) {
+      return t(text.substring(3));
+    }
+
+    return text;
+  }
+
+  const toolOptions = tools.map(t => ({
+    label: wrapTranslation(t.displayName),
+    description: wrapTranslation(t.displayDescription),
+    value: t.name,
+    title: wrapTranslation(t.displayDescription),
+  }));
+
+  return (
+    <Select
+      mode="multiple"
+      options={toolOptions}
+      optionRender={option => (
+        <Space size="large">
+          {option.label}
+          {option.data.description}
+        </Space>
+      )}
+      onChange={onChange}
+      value={value}
+      disabled={disabled}
+    ></Select>
+  );
+};
+
+export default LLMToolsSelect;

web/src/hooks/llm-hooks.tsx

@@ -71,6 +71,7 @@ function buildLlmOptionsWithIcon(x: IThirdOAIModel) {
     ),
     value: `${x.llm_name}@${x.fid}`,
     disabled: !x.available,
+    is_tools: x.is_tools,
   };
 }
 
@@ -142,7 +143,7 @@ export const useComposeLlmOptionsByModelTypes = (
 
   return modelTypes.reduce<
     (DefaultOptionType & {
-      options: { label: JSX.Element; value: string; disabled: boolean }[];
+      options: { label: JSX.Element; value: string; disabled: boolean; is_tools: boolean }[];
     })[]
   >((pre, cur) => {
     const options = allOptions[cur];

web/src/hooks/plugin-hooks.tsx

@@ -0,0 +1,17 @@
+import { ILLMTools } from '@/interfaces/database/plugin';
+import pluginService from '@/services/plugin-service';
+import { useQuery } from '@tanstack/react-query';
+
+export const useLlmToolsList = (): ILLMTools => {
+  const { data } = useQuery({
+    queryKey: ['llmTools'],
+    initialData: [],
+    queryFn: async () => {
+      const { data } = await pluginService.getLlmTools();
+
+      return data?.data ?? [];
+    },
+  });
+
+  return data;
+};

web/src/interfaces/database/llm.ts

@@ -13,6 +13,7 @@ export interface IThirdOAIModel {
   update_time: number;
   tenant_id?: string;
   tenant_name?: string;
+  is_tools: boolean;
 }
 
 export type IThirdOAIModelCollection = Record<string, IThirdOAIModel[]>;

web/src/interfaces/database/plugin.ts

@@ -0,0 +1,13 @@
+export type ILLMTools = ILLMToolMetadata[];
+
+export interface ILLMToolMetadata {
+    name: string;
+    displayName: string;
+    displayDescription: string;
+    parameters: Map<string, ILLMToolParameter>;
+}
+
+export interface ILLMToolParameter {
+    type: string;
+    displayDescription: string;
+}

web/src/locales/en.ts

@@ -454,6 +454,8 @@ This auto-tagging feature enhances retrieval by adding another layer of domain-s
       model: 'Model',
       modelTip: 'Large language chat model',
       modelMessage: 'Please select!',
+      modelEnabledTools: 'Enabled tools',
+      modelEnabledToolsTip: 'Please select one or more tools for the chat model to use. It takes no effect for models not supporting tool call.',
       freedom: 'Freedom',
       improvise: 'Improvise',
       precise: 'Precise',
@@ -1267,5 +1269,15 @@ This delimiter is used to split the input text into several text pieces echo of
       inputVariables: 'Input variables',
       runningHintText: 'is running...🕞',
     },
+    llmTools: {
+      bad_calculator: {
+        name: "Calculator",
+        description: "A tool to calculate the sum of two numbers (will give wrong answer)",
+        params: {
+          a: "The first number",
+          b: "The second number",
+        },
+      },
+    },
   },
 };

web/src/locales/zh.ts

@@ -461,6 +461,8 @@ General：实体和关系提取提示来自 GitHub - microsoft/graphrag：基于
       model: '模型',
       modelTip: '大语言聊天模型',
       modelMessage: '请选择',
+      modelEnabledTools: '可用的工具',
+      modelEnabledToolsTip: '请选择一个或多个可供该模型所使用的工具。仅对支持工具调用的模型生效。',
       freedom: '自由度',
       improvise: '即兴创作',
       precise: '精确',
@@ -1231,5 +1233,15 @@ General：实体和关系提取提示来自 GitHub - microsoft/graphrag：基于
       knowledge: 'knowledge',
       chat: 'chat',
     },
+    llmTools: {
+      bad_calculator: {
+        name: "计算器",
+        description: "用于计算两个数的和的工具（会给出错误答案）",
+        params: {
+          a: "第一个数",
+          b: "第二个数",
+        },
+      },
+    },
   },
 };

web/src/pages/flow/form/generate-form/index.tsx

@@ -4,10 +4,18 @@ import { PromptEditor } from '@/components/prompt-editor';
 import { useTranslate } from '@/hooks/common-hooks';
 import { Form, Switch } from 'antd';
 import { IOperatorForm } from '../../interface';
+import LLMToolsSelect from '@/components/llm-tools-select';
+import { useState } from 'react';
 
 const GenerateForm = ({ onValuesChange, form }: IOperatorForm) => {
   const { t } = useTranslate('flow');
 
+  const [isCurrentLlmSupportTools, setCurrentLlmSupportTools] = useState(false);
+
+  const onLlmSelectChanged = (_: string, option: any) => {
+    setCurrentLlmSupportTools(option.is_tools);
+  };
+
   return (
     <Form
       name="basic"
@@ -21,7 +29,7 @@ const GenerateForm = ({ onValuesChange, form }: IOperatorForm) => {
         label={t('model', { keyPrefix: 'chat' })}
         tooltip={t('modelTip', { keyPrefix: 'chat' })}
       >
-        <LLMSelect></LLMSelect>
+        <LLMSelect onInitialValue={onLlmSelectChanged} onChange={onLlmSelectChanged}></LLMSelect>
       </Form.Item>
       <Form.Item
         name={['prompt']}
@@ -38,6 +46,13 @@ const GenerateForm = ({ onValuesChange, form }: IOperatorForm) => {
         {/* <Input.TextArea rows={8}></Input.TextArea> */}
         <PromptEditor></PromptEditor>
       </Form.Item>
+      <Form.Item
+        name={'llm_enabled_tools'}
+        label={t('modelEnabledTools', { keyPrefix: 'chat' })}
+        tooltip={t('modelEnabledToolsTip', { keyPrefix: 'chat' })}
+      >
+        <LLMToolsSelect disabled={!isCurrentLlmSupportTools}></LLMToolsSelect>
+      </Form.Item>
       <Form.Item
         name={['cite']}
         label={t('cite')}

web/src/services/plugin-service.ts

@@ -0,0 +1,18 @@
+import api from '@/utils/api';
+import registerServer from '@/utils/register-server';
+import request from '@/utils/request';
+
+const {
+  llm_tools
+} = api;
+
+const methods = {
+  getLlmTools: {
+    url: llm_tools,
+    method: 'get',
+  },
+} as const;
+
+const pluginService = registerServer<keyof typeof methods>(methods, request);
+
+export default pluginService;

web/src/utils/api.ts

@@ -32,6 +32,9 @@ export default {
   delete_llm: `${api_host}/llm/delete_llm`,
   deleteFactory: `${api_host}/llm/delete_factory`,
 
+  // plugin
+  llm_tools: `${api_host}/plugin/llm_tools`,
+
   // knowledge base
   kb_list: `${api_host}/kb/list`,
   create_kb: `${api_host}/kb/create`,