# OmniShopAgent 项目技术实现报告

## 一、项目概述

OmniShopAgent 是一个基于 **LangGraph** 和 **ReAct 模式** 的自主多模态时尚购物智能体。系统能够自主决定调用哪些工具、维护对话状态、判断何时回复，实现智能化的商品发现与推荐。

### 核心特性

- **自主工具选择与执行**：Agent 根据用户意图自主选择并调用工具
- **多模态搜索**：支持文本搜索 + 图像搜索
- **对话上下文感知**：多轮对话中保持上下文记忆
- **实时视觉分析**：基于 VLM 的图片风格分析

---

## 二、技术栈

| 组件 | 技术选型 |
|------|----------|
| 运行环境 | Python 3.12 |
| Agent 框架 | LangGraph 1.x |
| LLM 框架 | LangChain 1.x（支持任意 LLM，默认 gpt-4o-mini） |
| 文本向量 | text-embedding-3-small |
| 图像向量 | CLIP ViT-B/32 |
| 向量数据库 | Milvus |
| 前端 | Streamlit |
| 数据集 | Kaggle Fashion Products |

---

## 三、系统架构

### 3.1 整体架构图

```
┌─────────────────────────────────────────────────────────────────┐
│                     Streamlit 前端 (app.py)                       │
└─────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────┐
│              ShoppingAgent (shopping_agent.py)                    │
│  ┌───────────────────────────────────────────────────────────┐  │
│  │  LangGraph StateGraph + ReAct Pattern                      │  │
│  │  START → Agent → [Has tool_calls?] → Tools → Agent → END   │  │
│  └───────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
        │                    │                    │
        ▼                    ▼                    ▼
┌──────────────┐   ┌──────────────────┐   ┌─────────────────────┐
│ search_      │   │ search_by_image   │   │ analyze_image_style  │
│ products     │   │                   │   │ (OpenAI Vision)      │
└──────┬───────┘   └────────┬─────────┘   └──────────┬───────────┘
       │                    │                        │
       ▼                    ▼                        ▼
┌─────────────────────────────────────────────────────────────────┐
│           EmbeddingService (embedding_service.py)                 │
│   OpenAI API (文本)  │  CLIP Server (图像)                       │
└─────────────────────────────────────────────────────────────────┘
                                    │
                                    ▼
┌─────────────────────────────────────────────────────────────────┐
│              MilvusService (milvus_service.py)                    │
│   text_embeddings 集合  │  image_embeddings 集合                 │
└─────────────────────────────────────────────────────────────────┘
```

### 3.2 Agent 流程图（LangGraph）

```mermaid
graph LR
    START --> Agent
    Agent -->|Has tool_calls| Tools
    Agent -->|No tool_calls| END
    Tools --> Agent
```

---

## 四、关键代码实现

### 4.1 Agent 核心实现（shopping_agent.py）

#### 4.1.1 状态定义

```python
from typing_extensions import Annotated, TypedDict
from langgraph.graph.message import add_messages

class AgentState(TypedDict):
    """State for the shopping agent with message accumulation"""
    messages: Annotated[Sequence[BaseMessage], add_messages]
    current_image_path: Optional[str]  # Track uploaded image
```

- `messages` 使用 `add_messages` 实现消息累加，支持多轮对话
- `current_image_path` 存储当前上传的图片路径供工具使用

#### 4.1.2 LangGraph 图构建

```python
def _build_graph(self):
    """Build the LangGraph StateGraph"""
    
    def agent_node(state: AgentState):
        """Agent decision node - decides which tools to call or when to respond"""
        messages = state["messages"]
        if not any(isinstance(m, SystemMessage) for m in messages):
            messages = [SystemMessage(content=system_prompt)] + list(messages)
        response = self.llm_with_tools.invoke(messages)
        return {"messages": [response]}

    tool_node = ToolNode(self.tools)

    def should_continue(state: AgentState):
        """Determine if agent should continue or end"""
        last_message = state["messages"][-1]
        if hasattr(last_message, "tool_calls") and last_message.tool_calls:
            return "tools"
        return END

    workflow = StateGraph(AgentState)
    workflow.add_node("agent", agent_node)
    workflow.add_node("tools", tool_node)
    workflow.add_edge(START, "agent")
    workflow.add_conditional_edges("agent", should_continue, ["tools", END])
    workflow.add_edge("tools", "agent")

    checkpointer = MemorySaver()
    return workflow.compile(checkpointer=checkpointer)
```

关键点：
- **agent_node**：将消息传入 LLM，由 LLM 决定是否调用工具
- **should_continue**：若有 `tool_calls` 则进入工具节点，否则结束
- **MemorySaver**：按 `thread_id` 持久化对话状态

#### 4.1.3 System Prompt 设计

```python
system_prompt = """You are an intelligent fashion shopping assistant. You can:
1. Search for products by text description (use search_products)
2. Find visually similar products from images (use search_by_image)
3. Analyze image style and attributes (use analyze_image_style)

When a user asks about products:
- For text queries: use search_products directly
- For image uploads: decide if you need to analyze_image_style first, then search
- You can call multiple tools in sequence if needed
- Always provide helpful, friendly responses

CRITICAL FORMATTING RULES:
When presenting product results, you MUST use this EXACT format for EACH product:
1. [Product Name]
   ID: [Product ID Number]
   Category: [Category]
   Color: [Color]
   Gender: [Gender]
   (Include Season, Usage, Relevance if available)
..."""
```

通过 system prompt 约束工具使用和输出格式，保证前端可正确解析产品信息。

#### 4.1.4 对话入口与流式处理

```python
def chat(self, query: str, image_path: Optional[str] = None) -> dict:
    # Build input message
    message_content = query
    if image_path:
        message_content = f"{query}\n[User uploaded image: {image_path}]"

    config = {"configurable": {"thread_id": self.session_id}}
    input_state = {
        "messages": [HumanMessage(content=message_content)],
        "current_image_path": image_path,
    }

    tool_calls = []
    for event in self.graph.stream(input_state, config=config):
        if "agent" in event:
            for msg in event["agent"].get("messages", []):
                if hasattr(msg, "tool_calls") and msg.tool_calls:
                    for tc in msg.tool_calls:
                        tool_calls.append({"name": tc["name"], "args": tc.get("args", {})})
        if "tools" in event:
            # 记录工具执行结果
            ...

    final_state = self.graph.get_state(config)
    response_text = final_state.values["messages"][-1].content

    return {"response": response_text, "tool_calls": tool_calls, "error": False}
```

---

### 4.2 搜索工具实现（search_tools.py）

#### 4.2.1 文本语义搜索

```python
@tool
def search_products(query: str, limit: int = 5) -> str:
    """Search for fashion products using natural language descriptions."""
    try:
        embedding_service = get_embedding_service()
        milvus_service = get_milvus_service()

        query_embedding = embedding_service.get_text_embedding(query)

        results = milvus_service.search_similar_text(
            query_embedding=query_embedding,
            limit=min(limit, 20),
            filters=None,
            output_fields=[
                "id", "productDisplayName", "gender", "masterCategory",
                "subCategory", "articleType", "baseColour", "season", "usage",
            ],
        )

        if not results:
            return "No products found matching your search."

        output = f"Found {len(results)} product(s):\n\n"
        for idx, product in enumerate(results, 1):
            output += f"{idx}. {product.get('productDisplayName', 'Unknown Product')}\n"
            output += f"   ID: {product.get('id', 'N/A')}\n"
            output += f"   Category: {product.get('masterCategory')} > {product.get('subCategory')} > {product.get('articleType')}\n"
            output += f"   Color: {product.get('baseColour')}\n"
            output += f"   Gender: {product.get('gender')}\n"
            if "distance" in product:
                similarity = 1 - product["distance"]
                output += f"   Relevance: {similarity:.2%}\n"
            output += "\n"

        return output.strip()
    except Exception as e:
        return f"Error searching products: {str(e)}"
```

#### 4.2.2 图像相似度搜索

```python
@tool
def search_by_image(image_path: str, limit: int = 5) -> str:
    """Find similar fashion products using an image."""
    if not Path(image_path).exists():
        return f"Error: Image file not found at '{image_path}'"

    embedding_service = get_embedding_service()
    milvus_service = get_milvus_service()

    if not embedding_service.clip_client:
        embedding_service.connect_clip()

    image_embedding = embedding_service.get_image_embedding(image_path)

    results = milvus_service.search_similar_images(
        query_embedding=image_embedding,
        limit=min(limit + 1, 21),
        output_fields=[...],
    )

    # 过滤掉查询图像本身（如上传的是商品库中的图）
    query_id = Path(image_path).stem
    filtered_results = [r for r in results if Path(r.get("image_path", "")).stem != query_id]
    filtered_results = filtered_results[:limit]


```

#### 4.2.3 视觉分析（VLM）

```python
@tool
def analyze_image_style(image_path: str) -> str:
    """Analyze a fashion product image using AI vision to extract detailed style information."""
    with open(img_path, "rb") as image_file:
        image_data = base64.b64encode(image_file.read()).decode("utf-8")

    prompt = """Analyze this fashion product image and provide a detailed description.
Include:
- Product type (e.g., shirt, dress, shoes, pants, bag)
- Primary colors
- Style/design (e.g., casual, formal, sporty, vintage, modern)
- Pattern or texture (e.g., plain, striped, checked, floral)
- Key features (e.g., collar type, sleeve length, fit)
- Material appearance (if obvious, e.g., denim, cotton, leather)
- Suitable occasion (e.g., office wear, party, casual, sports)
Provide a comprehensive yet concise description (3-4 sentences)."""

    client = get_openai_client()
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{
            "role": "user",
            "content": [
                {"type": "text", "text": prompt},
                {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{image_data}", "detail": "high"}},
            ],
        }],
        max_tokens=500,
        temperature=0.3,
    )

    return response.choices[0].message.content.strip()
```

---

### 4.3 向量服务实现

#### 4.3.1 EmbeddingService（embedding_service.py）

```python
class EmbeddingService:
    def get_text_embedding(self, text: str) -> List[float]:
        """OpenAI text-embedding-3-small"""
        response = self.openai_client.embeddings.create(
            input=text, model=self.text_embedding_model
        )
        return response.data[0].embedding

    def get_image_embedding(self, image_path: Union[str, Path]) -> List[float]:
        """CLIP 图像向量"""
        if not self.clip_client:
            raise RuntimeError("CLIP client not connected. Call connect_clip() first.")
        result = self.clip_client.encode([str(image_path)])
        if isinstance(result, np.ndarray):
            embedding = result[0].tolist() if len(result.shape) > 1 else result.tolist()
        else:
            embedding = result[0].embedding.tolist()
        return embedding

    def get_text_embeddings_batch(self, texts: List[str], batch_size: int = 100) -> List[List[float]]:
        """批量文本嵌入，用于索引"""
        for i in range(0, len(texts), batch_size):
            batch = texts[i : i + batch_size]
            response = self.openai_client.embeddings.create(input=batch, ...)
            embeddings = [item.embedding for item in response.data]
            all_embeddings.extend(embeddings)
        return all_embeddings
```

#### 4.3.2 MilvusService（milvus_service.py）

**文本集合 Schema：**

```python
schema = MilvusClient.create_schema(auto_id=False, enable_dynamic_field=True)
schema.add_field(field_name="id", datatype=DataType.INT64, is_primary=True)
schema.add_field(field_name="text", datatype=DataType.VARCHAR, max_length=2000)
schema.add_field(field_name="embedding", datatype=DataType.FLOAT_VECTOR, dim=self.text_dim)  # 1536
schema.add_field(field_name="productDisplayName", datatype=DataType.VARCHAR, max_length=500)
schema.add_field(field_name="gender", datatype=DataType.VARCHAR, max_length=50)
schema.add_field(field_name="masterCategory", datatype=DataType.VARCHAR, max_length=100)
# ... 更多元数据字段
```

**图像集合 Schema：**

```python
schema.add_field(field_name="id", datatype=DataType.INT64, is_primary=True)
schema.add_field(field_name="image_path", datatype=DataType.VARCHAR, max_length=500)
schema.add_field(field_name="embedding", datatype=DataType.FLOAT_VECTOR, dim=self.image_dim)  # 512
# ... 产品元数据
```

**相似度搜索：**

```python
def search_similar_text(self, query_embedding, limit=10, output_fields=None):
    results = self.client.search(
        collection_name=self.text_collection_name,
        data=[query_embedding],
        limit=limit,
        output_fields=output_fields,
    )
    formatted_results = []
    for hit in results[0]:
        result = {"id": hit.get("id"), "distance": hit.get("distance")}
        entity = hit.get("entity", {})
        for field in output_fields:
            if field in entity:
                result[field] = entity.get(field)
        formatted_results.append(result)
    return formatted_results
```

---

### 4.4 数据索引脚本（index_data.py）

#### 4.4.1 产品数据加载

```python
def _load_products_from_csv(self) -> Dict[int, Dict[str, Any]]:
    products = {}
    # 加载 images.csv 映射
    with open(self.images_csv, "r") as f:
        images_dict = {int(row["filename"].split(".")[0]): row["link"] for row in csv.DictReader(f)}

    # 加载 styles.csv
    with open(self.styles_csv, "r") as f:
        for row in csv.DictReader(f):
            product_id = int(row["id"])
            products[product_id] = {
                "id": product_id,
                "gender": row.get("gender", ""),
                "masterCategory": row.get("masterCategory", ""),
                "subCategory": row.get("subCategory", ""),
                "articleType": row.get("articleType", ""),
                "baseColour": row.get("baseColour", ""),
                "season": row.get("season", ""),
                "usage": row.get("usage", ""),
                "productDisplayName": row.get("productDisplayName", ""),
                "imagePath": f"{product_id}.jpg",
            }
    return products
```

#### 4.4.2 文本索引

```python
def _create_product_text(self, product: Dict[str, Any]) -> str:
    """构造产品文本用于 embedding"""
    parts = [
        product.get("productDisplayName", ""),
        f"Gender: {product.get('gender', '')}",
        f"Category: {product.get('masterCategory', '')} > {product.get('subCategory', '')}",
        f"Type: {product.get('articleType', '')}",
        f"Color: {product.get('baseColour', '')}",
        f"Season: {product.get('season', '')}",
        f"Usage: {product.get('usage', '')}",
    ]
    return " | ".join([p for p in parts if p and p != "Gender: " and p != "Color: "])
```

#### 4.4.3 批量索引流程

```python
# 文本索引
texts = [self._create_product_text(p) for p in products]
embeddings = self.embedding_service.get_text_embeddings_batch(texts, batch_size=50)
milvus_data = [{
    "id": product_id,
    "text": text[:2000],
    "embedding": embedding,
    "productDisplayName": product["productDisplayName"][:500],
    "gender": product["gender"][:50],
    # ... 其他元数据
} for product_id, text, embedding in zip(...)]
self.milvus_service.insert_text_embeddings(milvus_data)

# 图像索引
image_paths = [self.image_dir / p["imagePath"] for p in products]
embeddings = self.embedding_service.get_image_embeddings_batch(image_paths, batch_size=32)
# 类似插入 image_embeddings 集合
```

---

### 4.5 Streamlit 前端（app.py）

#### 4.5.1 会话与 Agent 初始化

```python
def initialize_session():
    if "session_id" not in st.session_state:
        st.session_state.session_id = str(uuid.uuid4())
    if "shopping_agent" not in st.session_state:
        st.session_state.shopping_agent = ShoppingAgent(session_id=st.session_state.session_id)
    if "messages" not in st.session_state:
        st.session_state.messages = []
    if "uploaded_image" not in st.session_state:
        st.session_state.uploaded_image = None
```

#### 4.5.2 产品信息解析

```python
def extract_products_from_response(response: str) -> list:
    """从 Agent 回复中解析产品信息"""
    products = []
    for line in response.split("\n"):
        if re.match(r"^\*?\*?\d+\.\s+", line):
            if current_product:
                products.append(current_product)
            current_product = {"name": re.sub(r"^\*?\*?\d+\.\s+", "", line).replace("**", "").strip()}
        elif "ID:" in line:
            id_match = re.search(r"(?:ID|id):\s*(\d+)", line)
            if id_match:
                current_product["id"] = id_match.group(1)
        elif "Category:" in line:
            cat_match = re.search(r"Category:\s*(.+?)(?:\n|$)", line)
            if cat_match:
                current_product["category"] = cat_match.group(1).strip()
        # ... Color, Gender, Season, Usage, Similarity/Relevance
    return products
```

#### 4.5.3 多轮对话中的图片引用

```python
# 用户输入 "make them formal" 时，若上一条消息有图片，则引用该图片
if any(ref in query_lower for ref in ["this", "that", "the image", "it"]):
    for msg in reversed(st.session_state.messages):
        if msg.get("role") == "user" and msg.get("image_path"):
            image_path = msg["image_path"]
            break
```

---

### 4.6 配置管理（config.py）

```python
class Settings(BaseSettings):
    openai_api_key: str
    openai_model: str = "gpt-4o-mini"
    openai_embedding_model: str = "text-embedding-3-small"
    clip_server_url: str = "grpc://localhost:51000"
    milvus_uri: str = "http://localhost:19530"
    text_collection_name: str = "text_embeddings"
    image_collection_name: str = "image_embeddings"
    text_dim: int = 1536
    image_dim: int = 512

    @property
    def milvus_uri_absolute(self) -> str:
        """支持 Milvus Standalone 和 Milvus Lite"""
        if self.milvus_uri.startswith(("http://", "https://")):
            return self.milvus_uri
        if self.milvus_uri.startswith("./"):
            return os.path.join(base_dir, self.milvus_uri[2:])
        return self.milvus_uri

    class Config:
        env_file = ".env"
```

---

## 五、部署与运行

### 5.1 依赖服务

```yaml
# docker-compose.yml 提供
- etcd: 元数据存储
- minio: 对象存储
- milvus-standalone: 向量数据库
- attu: Milvus 管理界面
```

### 5.2 启动流程

```bash
# 1. 环境
pip install -r requirements.txt
cp .env.example .env  # 配置 OPENAI_API_KEY

# 2. 下载数据
python scripts/download_dataset.py  # Kaggle Fashion Product Images Dataset

# 3. 启动 CLIP 服务（需单独运行）
python -m clip_server

# 4. 启动 Milvus
docker-compose up

# 5. 索引数据
python scripts/index_data.py

# 6. 启动应用
streamlit run app.py
```

---

## 六、典型交互流程

| 场景 | 用户输入 | Agent 行为 | 工具调用 |
|------|----------|------------|----------|
| 文本搜索 | "winter coats for women" | 直接文本搜索 | `search_products("winter coats women")` |
| 图像搜索 | [上传图片] "find similar" | 图像相似度搜索 | `search_by_image(path)` |
| 风格分析+搜索 | [上传复古夹克] "what style? find matching pants" | 先分析风格再搜索 | `analyze_image_style(path)` → `search_products("vintage pants casual")` |
| 多轮上下文 | [第1轮] "show me red dresses"<br>[第2轮] "make them formal" | 结合上下文 | `search_products("red formal dresses")` |

---

## 七、设计要点总结

1. **ReAct 模式**：Agent 自主决定何时调用工具、调用哪些工具、是否继续调用。
2. **LangGraph 状态图**：`START → Agent → [条件] → Tools → Agent → END`，支持多轮工具调用。
3. **多模态**：文本 + 图像 + VLM 分析，覆盖文本搜索、以图搜图、风格理解。
4. **双向量集合**：Milvus 中 text_embeddings / image_embeddings 分别存储，支持不同模态的检索。
5. **会话持久化**：`MemorySaver` + `thread_id` 实现多轮对话记忆。
6. **格式约束**：System prompt 严格限制产品输出格式，便于前端解析和展示。

---

## 八、附录：项目结构

```
OmniShopAgent/
├── app/
│   ├── agents/
│   │   └── shopping_agent.py
│   ├── config.py
│   ├── services/
│   │   ├── embedding_service.py
│   │   └── milvus_service.py
│   └── tools/
│       └── search_tools.py
├── scripts/
│   ├── download_dataset.py
│   └── index_data.py
├── app.py
├── docker-compose.yml
└── requirements.txt
```