|
| 1 | +--- |
| 2 | +slug: multimodal-support |
| 3 | +title: 多模态支持 |
| 4 | +description: Ms-Agent 多模态对话使用指南:图片理解、分析功能配置与使用方法。 |
| 5 | +--- |
| 6 | + |
| 7 | +# 多模态支持 |
| 8 | + |
| 9 | +本文档介绍如何使用 ms-agent 进行多模态对话,包括图片理解和分析功能。 |
| 10 | + |
| 11 | +## 概述 |
| 12 | + |
| 13 | +ms-agent 已经支持多模态模型,如阿里云的 `qwen3.5-plus` 模型。多模态模型能够: |
| 14 | +- 分析图片内容 |
| 15 | +- 识别图片中的对象、场景和文字 |
| 16 | +- 结合图片内容进行对话 |
| 17 | + |
| 18 | +## 前置要求 |
| 19 | + |
| 20 | +### 1. 安装依赖 |
| 21 | + |
| 22 | +确保已安装必要的依赖包: |
| 23 | + |
| 24 | +```bash |
| 25 | +pip install openai |
| 26 | +``` |
| 27 | + |
| 28 | +### 2. 配置 API Key |
| 29 | + |
| 30 | +(以 qwen3.5-plus 为例)获取 DashScope API Key 并设置环境变量: |
| 31 | + |
| 32 | +```bash |
| 33 | +export DASHSCOPE_API_KEY='your-dashscope-api-key' |
| 34 | +``` |
| 35 | + |
| 36 | +或者在配置文件中直接设置 `dashscope_api_key`。 |
| 37 | + |
| 38 | +## 配置多模态模型 |
| 39 | + |
| 40 | +多模态功能主要取决于两点: |
| 41 | +1. **选择支持多模态的模型**(如 `qwen3.5-plus`) |
| 42 | +2. **使用正确的消息格式**(包含 `image_url` 块) |
| 43 | + |
| 44 | +你可以在现有配置基础上,通过代码动态修改模型配置: |
| 45 | + |
| 46 | +```python |
| 47 | +from ms_agent.config import Config |
| 48 | +from ms_agent import LLMAgent |
| 49 | +import os |
| 50 | + |
| 51 | +# 使用现有配置文件(如 ms_agent/agent/agent.yaml) |
| 52 | +config = Config.from_task('ms_agent/agent/agent.yaml') |
| 53 | + |
| 54 | +# 覆盖配置为多模态模型 |
| 55 | +config.llm.model = 'qwen3.5-plus' |
| 56 | +config.llm.service = 'dashscope' |
| 57 | +config.llm.dashscope_api_key = os.environ.get('DASHSCOPE_API_KEY', '') |
| 58 | +config.llm.modelscope_base_url = 'https://dashscope.aliyuncs.com/compatible-mode/v1' |
| 59 | + |
| 60 | +# 创建 LLMAgent |
| 61 | +agent = LLMAgent(config=config) |
| 62 | +``` |
| 63 | + |
| 64 | +## 使用 LLMAgent 进行多模态对话 |
| 65 | + |
| 66 | +推荐使用 `LLMAgent` 来进行多模态对话,它提供了更完整的功能,包括记忆管理、工具调用和回调支持。 |
| 67 | + |
| 68 | +### 基本用法 |
| 69 | + |
| 70 | +```python |
| 71 | +import asyncio |
| 72 | +import os |
| 73 | +from ms_agent import LLMAgent |
| 74 | +from ms_agent.config import Config |
| 75 | +from ms_agent.llm.utils import Message |
| 76 | + |
| 77 | +async def multimodal_chat(): |
| 78 | + # 创建配置 |
| 79 | + config = Config.from_task('ms_agent/agent/agent.yaml') |
| 80 | + config.llm.model = 'qwen3.5-plus' |
| 81 | + config.llm.service = 'dashscope' |
| 82 | + config.llm.dashscope_api_key = os.environ.get('DASHSCOPE_API_KEY', '') |
| 83 | + config.llm.modelscope_base_url = 'https://dashscope.aliyuncs.com/compatible-mode/v1' |
| 84 | + |
| 85 | + # 创建 LLMAgent |
| 86 | + agent = LLMAgent(config=config) |
| 87 | + |
| 88 | + # 构建多模态消息 |
| 89 | + multimodal_content = [ |
| 90 | + {"type": "text", "text": "请描述这张图片。"}, |
| 91 | + {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}} |
| 92 | + ] |
| 93 | + |
| 94 | + # 调用 agent |
| 95 | + response = await agent.run(messages=[Message(role="user", content=multimodal_content)]) |
| 96 | + print(response[-1].content) |
| 97 | + |
| 98 | +asyncio.run(multimodal_chat()) |
| 99 | +``` |
| 100 | + |
| 101 | +### 非 Stream 模式 |
| 102 | + |
| 103 | +```python |
| 104 | +# 配置中禁用 stream |
| 105 | +config.generation_config.stream = False |
| 106 | + |
| 107 | +agent = LLMAgent(config=config) |
| 108 | + |
| 109 | +multimodal_content = [ |
| 110 | + {"type": "text", "text": "请描述这张图片。"}, |
| 111 | + {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}} |
| 112 | +] |
| 113 | + |
| 114 | +# 非 stream 模式:直接返回完整响应 |
| 115 | +response = await agent.run(messages=[Message(role="user", content=multimodal_content)]) |
| 116 | +print(f"[回复] {response[-1].content}") |
| 117 | +print(f"[Token使用] 输入: {response[-1].prompt_tokens}, 输出: {response[-1].completion_tokens}") |
| 118 | +``` |
| 119 | + |
| 120 | +### Stream 模式 |
| 121 | + |
| 122 | +```python |
| 123 | +# 配置中启用 stream |
| 124 | +config.generation_config.stream = True |
| 125 | + |
| 126 | +agent = LLMAgent(config=config) |
| 127 | + |
| 128 | +multimodal_content = [ |
| 129 | + {"type": "text", "text": "请描述这张图片。"}, |
| 130 | + {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}} |
| 131 | +] |
| 132 | + |
| 133 | +# stream 模式:返回生成器 |
| 134 | +generator = await agent.run( |
| 135 | + messages=[Message(role="user", content=multimodal_content)], |
| 136 | + stream=True |
| 137 | +) |
| 138 | + |
| 139 | +full_response = "" |
| 140 | +async for response_chunk in generator: |
| 141 | + if response_chunk and len(response_chunk) > 0: |
| 142 | + last_msg = response_chunk[-1] |
| 143 | + if last_msg.content: |
| 144 | + # 流式输出新增内容 |
| 145 | + print(last_msg.content[len(full_response):], end='', flush=True) |
| 146 | + full_response = last_msg.content |
| 147 | + |
| 148 | +print(f"\n[完整回复] {full_response}") |
| 149 | +``` |
| 150 | + |
| 151 | +### 多轮对话 |
| 152 | + |
| 153 | +LLMAgent 支持多轮对话,可以在对话中混合使用图片和文本: |
| 154 | + |
| 155 | +```python |
| 156 | +agent = LLMAgent(config=config, tag="multimodal_conversation") |
| 157 | + |
| 158 | +# 第一轮:发送图片 |
| 159 | +multimodal_content = [ |
| 160 | + {"type": "text", "text": "这张图片里有几个人?"}, |
| 161 | + {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}} |
| 162 | +] |
| 163 | + |
| 164 | +messages = [Message(role="user", content=multimodal_content)] |
| 165 | +response = await agent.run(messages=messages) |
| 166 | +print(f"[第一轮回复] {response[-1].content}") |
| 167 | + |
| 168 | +# 第二轮:继续追问(纯文本,保留上下文) |
| 169 | +messages = response # 使用上一轮的回复作为上下文 |
| 170 | +messages.append(Message(role="user", content="他们在做什么?")) |
| 171 | +response = await agent.run(messages=messages) |
| 172 | +print(f"[第二轮回复] {response[-1].content}") |
| 173 | +``` |
| 174 | + |
| 175 | +## 多模态消息格式 |
| 176 | + |
| 177 | +ms-agent 使用 OpenAI 兼容的多模态消息格式。图片可以通过以下三种方式提供: |
| 178 | + |
| 179 | +### 1. 图片 URL |
| 180 | + |
| 181 | +```python |
| 182 | +from ms_agent.llm.utils import Message |
| 183 | + |
| 184 | +multimodal_content = [ |
| 185 | + {"type": "text", "text": "请描述这张图片。"}, |
| 186 | + {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}} |
| 187 | +] |
| 188 | + |
| 189 | +messages = [ |
| 190 | + Message(role="user", content=multimodal_content) |
| 191 | +] |
| 192 | + |
| 193 | +response = llm.generate(messages=messages) |
| 194 | +``` |
| 195 | + |
| 196 | +### 2. Base64 编码 |
| 197 | + |
| 198 | +```python |
| 199 | +import base64 |
| 200 | + |
| 201 | +# 读取并编码图片 |
| 202 | +with open('image.jpg', 'rb') as f: |
| 203 | + image_data = base64.b64encode(f.read()).decode('utf-8') |
| 204 | + |
| 205 | +multimodal_content = [ |
| 206 | + {"type": "text", "text": "这是什么?"}, |
| 207 | + { |
| 208 | + "type": "image_url", |
| 209 | + "image_url": { |
| 210 | + "url": f"data:image/jpeg;base64,{image_data}" |
| 211 | + } |
| 212 | + } |
| 213 | +] |
| 214 | + |
| 215 | +messages = [Message(role="user", content=multimodal_content)] |
| 216 | +response = llm.generate(messages=messages) |
| 217 | +``` |
| 218 | + |
| 219 | +### 3. 本地文件路径 |
| 220 | + |
| 221 | +```python |
| 222 | +import base64 |
| 223 | +import os |
| 224 | + |
| 225 | +image_path = 'path/to/image.png' |
| 226 | + |
| 227 | +# 获取 MIME 类型 |
| 228 | +ext = os.path.splitext(image_path)[1].lower() |
| 229 | +mime_type = { |
| 230 | + '.png': 'image/png', |
| 231 | + '.jpg': 'image/jpeg', |
| 232 | + '.jpeg': 'image/jpeg', |
| 233 | + '.gif': 'image/gif', |
| 234 | + '.webp': 'image/webp' |
| 235 | +}.get(ext, 'image/png') |
| 236 | + |
| 237 | +# 读取并编码 |
| 238 | +with open(image_path, 'rb') as f: |
| 239 | + image_data = base64.b64encode(f.read()).decode('utf-8') |
| 240 | + |
| 241 | +multimodal_content = [ |
| 242 | + {"type": "text", "text": "描述这张图片。"}, |
| 243 | + { |
| 244 | + "type": "image_url", |
| 245 | + "image_url": { |
| 246 | + "url": f"data:{mime_type};base64,{image_data}" |
| 247 | + } |
| 248 | + } |
| 249 | +] |
| 250 | + |
| 251 | +messages = [Message(role="user", content=multimodal_content)] |
| 252 | +response = llm.generate(messages=messages) |
| 253 | +``` |
| 254 | + |
| 255 | +## 运行示例 |
| 256 | + |
| 257 | +### 运行 Agent 示例 |
| 258 | + |
| 259 | +```bash |
| 260 | +# 运行完整测试套件(包括 stream 和非 stream 模式) |
| 261 | +python examples/agent/test_llm_agent_multimodal.py |
| 262 | +``` |
| 263 | + |
| 264 | +## 常见问题 |
| 265 | + |
| 266 | +### Q: 图片大小有限制吗? |
| 267 | + |
| 268 | +A: 是的,不同模型有不同的限制: |
| 269 | +- qwen3.5-plus: 推荐图片大小不超过 4MB |
| 270 | +- 分辨率建议不超过 2048x2048 |
| 271 | + |
| 272 | +### Q: 支持哪些图片格式? |
| 273 | + |
| 274 | +A: 通常支持: |
| 275 | +- JPEG / JPG |
| 276 | +- PNG |
| 277 | +- GIF |
| 278 | +- WebP |
| 279 | + |
| 280 | +### Q: 可以一次发送多张图片吗? |
| 281 | + |
| 282 | +A: 是的,可以在消息中添加多个 `image_url` 块: |
| 283 | + |
| 284 | +```python |
| 285 | +multimodal_content = [ |
| 286 | + {"type": "text", "text": "比较这两张图片。"}, |
| 287 | + {"type": "image_url", "image_url": {"url": "https://example.com/img1.jpg"}}, |
| 288 | + {"type": "image_url", "image_url": {"url": "https://example.com/img2.jpg"}} |
| 289 | +] |
| 290 | +``` |
| 291 | + |
| 292 | +### Q: 流式输出支持吗? |
| 293 | + |
| 294 | +A: 是的,多模态对话支持流式输出。设置 `stream: true` 即可: |
| 295 | + |
| 296 | +```python |
| 297 | +config.generation_config.stream = True |
| 298 | +response = llm.generate(messages=messages, stream=True) |
| 299 | +``` |
0 commit comments