244 lines
7.1 KiB
Markdown
244 lines
7.1 KiB
Markdown
# LLM API
|
||
|
||
LLM API模块提供与大语言模型交互的功能，让插件能够使用系统配置的LLM模型进行内容生成。
|
||
|
||
## 导入方式
|
||
|
||
```python
|
||
from src.plugin_system.apis import llm_api
|
||
```
|
||
|
||
## 主要功能
|
||
|
||
### 1. 模型管理
|
||
|
||
#### `get_available_models() -> Dict[str, Any]`
|
||
获取所有可用的模型配置
|
||
|
||
**返回：**
|
||
- `Dict[str, Any]`：模型配置字典，key为模型名称，value为模型配置
|
||
|
||
**示例：**
|
||
```python
|
||
models = llm_api.get_available_models()
|
||
for model_name, model_config in models.items():
|
||
print(f"模型: {model_name}")
|
||
print(f"配置: {model_config}")
|
||
```
|
||
|
||
### 2. 内容生成
|
||
|
||
#### `generate_with_model(prompt, model_config, request_type="plugin.generate", **kwargs)`
|
||
使用指定模型生成内容
|
||
|
||
**参数：**
|
||
- `prompt`：提示词
|
||
- `model_config`：模型配置（从 get_available_models 获取）
|
||
- `request_type`：请求类型标识
|
||
- `**kwargs`：其他模型特定参数，如temperature、max_tokens等
|
||
|
||
**返回：**
|
||
- `Tuple[bool, str, str, str]`：(是否成功, 生成的内容, 推理过程, 模型名称)
|
||
|
||
**示例：**
|
||
```python
|
||
models = llm_api.get_available_models()
|
||
default_model = models.get("default")
|
||
|
||
if default_model:
|
||
success, response, reasoning, model_name = await llm_api.generate_with_model(
|
||
prompt="请写一首关于春天的诗",
|
||
model_config=default_model,
|
||
temperature=0.7,
|
||
max_tokens=200
|
||
)
|
||
|
||
if success:
|
||
print(f"生成内容: {response}")
|
||
print(f"使用模型: {model_name}")
|
||
```
|
||
|
||
## 使用示例
|
||
|
||
### 1. 基础文本生成
|
||
|
||
```python
|
||
from src.plugin_system.apis import llm_api
|
||
|
||
async def generate_story(topic: str):
|
||
"""生成故事"""
|
||
models = llm_api.get_available_models()
|
||
model = models.get("default")
|
||
|
||
if not model:
|
||
return "未找到可用模型"
|
||
|
||
prompt = f"请写一个关于{topic}的短故事，大约100字左右。"
|
||
|
||
success, story, reasoning, model_name = await llm_api.generate_with_model(
|
||
prompt=prompt,
|
||
model_config=model,
|
||
request_type="story.generate",
|
||
temperature=0.8,
|
||
max_tokens=150
|
||
)
|
||
|
||
return story if success else "故事生成失败"
|
||
```
|
||
|
||
### 2. 在Action中使用LLM
|
||
|
||
```python
|
||
from src.plugin_system.base import BaseAction
|
||
|
||
class LLMAction(BaseAction):
|
||
async def execute(self, action_data, chat_stream):
|
||
# 获取用户输入
|
||
user_input = action_data.get("user_message", "")
|
||
intent = action_data.get("intent", "chat")
|
||
|
||
# 获取模型配置
|
||
models = llm_api.get_available_models()
|
||
model = models.get("default")
|
||
|
||
if not model:
|
||
return {"success": False, "error": "未配置LLM模型"}
|
||
|
||
# 构建提示词
|
||
prompt = self.build_prompt(user_input, intent)
|
||
|
||
# 生成回复
|
||
success, response, reasoning, model_name = await llm_api.generate_with_model(
|
||
prompt=prompt,
|
||
model_config=model,
|
||
request_type=f"plugin.{self.plugin_name}",
|
||
temperature=0.7
|
||
)
|
||
|
||
if success:
|
||
return {
|
||
"success": True,
|
||
"response": response,
|
||
"model_used": model_name,
|
||
"reasoning": reasoning
|
||
}
|
||
|
||
return {"success": False, "error": response}
|
||
|
||
def build_prompt(self, user_input: str, intent: str) -> str:
|
||
"""构建提示词"""
|
||
base_prompt = "你是一个友善的AI助手。"
|
||
|
||
if intent == "question":
|
||
return f"{base_prompt}\n\n用户问题：{user_input}\n\n请提供准确、有用的回答："
|
||
elif intent == "chat":
|
||
return f"{base_prompt}\n\n用户说：{user_input}\n\n请进行自然的对话："
|
||
else:
|
||
return f"{base_prompt}\n\n用户输入：{user_input}\n\n请回复："
|
||
```
|
||
|
||
### 3. 多模型对比
|
||
|
||
```python
|
||
async def compare_models(prompt: str):
|
||
"""使用多个模型生成内容并对比"""
|
||
models = llm_api.get_available_models()
|
||
results = {}
|
||
|
||
for model_name, model_config in models.items():
|
||
success, response, reasoning, actual_model = await llm_api.generate_with_model(
|
||
prompt=prompt,
|
||
model_config=model_config,
|
||
request_type="comparison.test"
|
||
)
|
||
|
||
results[model_name] = {
|
||
"success": success,
|
||
"response": response,
|
||
"model": actual_model,
|
||
"reasoning": reasoning
|
||
}
|
||
|
||
return results
|
||
```
|
||
|
||
### 4. 智能对话插件
|
||
|
||
```python
|
||
class ChatbotPlugin(BasePlugin):
|
||
async def handle_action(self, action_data, chat_stream):
|
||
user_message = action_data.get("message", "")
|
||
|
||
# 获取历史对话上下文
|
||
context = self.get_conversation_context(chat_stream)
|
||
|
||
# 构建对话提示词
|
||
prompt = self.build_conversation_prompt(user_message, context)
|
||
|
||
# 获取模型配置
|
||
models = llm_api.get_available_models()
|
||
chat_model = models.get("chat", models.get("default"))
|
||
|
||
if not chat_model:
|
||
return {"success": False, "message": "聊天模型未配置"}
|
||
|
||
# 生成回复
|
||
success, response, reasoning, model_name = await llm_api.generate_with_model(
|
||
prompt=prompt,
|
||
model_config=chat_model,
|
||
request_type="chat.conversation",
|
||
temperature=0.8,
|
||
max_tokens=500
|
||
)
|
||
|
||
if success:
|
||
# 保存对话历史
|
||
self.save_conversation(chat_stream, user_message, response)
|
||
|
||
return {
|
||
"success": True,
|
||
"reply": response,
|
||
"model": model_name
|
||
}
|
||
|
||
return {"success": False, "message": "回复生成失败"}
|
||
|
||
def build_conversation_prompt(self, user_message: str, context: list) -> str:
|
||
"""构建对话提示词"""
|
||
prompt = "你是一个有趣、友善的聊天机器人。请自然地回复用户的消息。\n\n"
|
||
|
||
# 添加历史对话
|
||
if context:
|
||
prompt += "对话历史：\n"
|
||
for msg in context[-5:]: # 只保留最近5条
|
||
prompt += f"用户: {msg['user']}\n机器人: {msg['bot']}\n"
|
||
prompt += "\n"
|
||
|
||
prompt += f"用户: {user_message}\n机器人: "
|
||
return prompt
|
||
```
|
||
|
||
## 模型配置说明
|
||
|
||
### 常用模型类型
|
||
- `default`：默认模型
|
||
- `chat`：聊天专用模型
|
||
- `creative`：创意生成模型
|
||
- `code`：代码生成模型
|
||
|
||
### 配置参数
|
||
LLM模型支持的常用参数：
|
||
- `temperature`：控制输出随机性（0.0-1.0）
|
||
- `max_tokens`：最大生成长度
|
||
- `top_p`：核采样参数
|
||
- `frequency_penalty`：频率惩罚
|
||
- `presence_penalty`：存在惩罚
|
||
|
||
## 注意事项
|
||
|
||
1. **异步操作**：LLM生成是异步的，必须使用`await`
|
||
2. **错误处理**：生成失败时返回False和错误信息
|
||
3. **配置依赖**：需要正确配置模型才能使用
|
||
4. **请求类型**：建议为不同用途设置不同的request_type
|
||
5. **性能考虑**：LLM调用可能较慢，考虑超时和缓存
|
||
6. **成本控制**：注意控制max_tokens以控制成本 |