流程跑通

tangwang
1 parent bad17b15
Showing 4 changed files with 239 additions and 62 deletions Show diff stats
.env.example
app.py
app/agents/shopping_agent.py
cli_debug.py
@@ -19,7 +19,7 @@ SIMILARITY_THRESHOLD=0.6
 # Search API (see docs/搜索API对接指南.md)
 SEARCH_API_BASE_URL=http://120.76.41.98:6002
-SEARCH_API_TENANT_ID=162
+SEARCH_API_TENANT_ID=170
 # ====================
 # Application Configuration
@@ -245,6 +245,10 @@ def initialize_session():
     if "show_image_upload" not in st.session_state:
         st.session_state.show_image_upload = False
+    # Debug panel toggle
+    if "show_debug" not in st.session_state:
+        st.session_state.show_debug = False
+
 def save_uploaded_image(uploaded_file) -> Optional[str]:
     """Save uploaded image to temp directory"""
@@ -394,6 +398,7 @@ def display_message(message: dict):
     content = message["content"]
     image_path = message.get("image_path")
     tool_calls = message.get("tool_calls", [])
+    debug_steps = message.get("debug_steps", [])
     if role == "user":
         st.markdown('<div class="message user-message">', unsafe_allow_html=True)
@@ -411,9 +416,42 @@ def display_message(message: dict):
     else:  # assistant
         # Display tool calls horizontally - only tool names
         if tool_calls:
-            tool_names = [tc['name'] for tc in tool_calls]
+            tool_names = [tc["name"] for tc in tool_calls]
             st.caption(" → ".join(tool_names))
             st.markdown("")
+
+        # Optional: detailed debug panel (reasoning + tool details)
+        if debug_steps and st.session_state.get("show_debug"):
+            with st.expander("思考 & 工具调用详细过程", expanded=False):
+                for idx, step in enumerate(debug_steps, 1):
+                    node = step.get("node", "unknown")
+                    st.markdown(f"**Step {idx} – {node}**")
+
+                    if node == "agent":
+                        msgs = step.get("messages", [])
+                        if msgs:
+                            st.markdown("**Agent Messages**")
+                            for m in msgs:
+                                role = m.get("type", "assistant")
+                                content = m.get("content", "")
+                                st.markdown(f"- `{role}`: {content}")
+
+                        tcs = step.get("tool_calls", [])
+                        if tcs:
+                            st.markdown("**Planned Tool Calls**")
+                            for j, tc in enumerate(tcs, 1):
+                                st.markdown(f"- **{j}. {tc.get('name')}**")
+                                st.code(tc.get("args", {}), language="json")
+
+                    elif node == "tools":
+                        results = step.get("results", [])
+                        if results:
+                            st.markdown("**Tool Results**")
+                            for j, r in enumerate(results, 1):
+                                st.markdown(f"- **Result {j}:**")
+                                st.code(r.get("content", ""), language="text")
+
+                    st.markdown("---")
         # Extract and display products if any
         products = extract_products_from_response(content)
@@ -486,9 +524,9 @@ def display_welcome():
         st.markdown(
             """
             <div class="feature-card">
-                <div class="feature-icon">💬</div>
-                <div class="feature-title">Text Search</div>
-                <div>Describe what you want</div>
+                <div class="feature-icon">💗</div>
+                <div class="feature-title">懂你</div>
+                <div>能记住你的偏好，给你推荐适合的</div>
             </div>
             """,
             unsafe_allow_html=True,
@@ -498,9 +536,9 @@ def display_welcome():
         st.markdown(
             """
             <div class="feature-card">
-                <div class="feature-icon">📸</div>
-                <div class="feature-title">Image Search</div>
-                <div>Upload product photos</div>
+                <div class="feature-icon">🛍️</div>
+                <div class="feature-title">懂商品</div>
+                <div>深度理解店铺内所有商品，智能匹配你的需求</div>
             </div>
             """,
             unsafe_allow_html=True,
@@ -510,9 +548,9 @@ def display_welcome():
         st.markdown(
             """
             <div class="feature-card">
-                <div class="feature-icon">🔍</div>
-                <div class="feature-title">Visual Analysis</div>
-                <div>AI analyzes prodcut style</div>
+                <div class="feature-icon">💭</div>
+                <div class="feature-title">贴心</div>
+                <div>任意聊</div>
             </div>
             """,
             unsafe_allow_html=True,
@@ -522,9 +560,9 @@ def display_welcome():
         st.markdown(
             """
             <div class="feature-card">
-                <div class="feature-icon">💭</div>
-                <div class="feature-title">Conversational</div>
-                <div>Remembers context</div>
+                <div class="feature-icon">👗</div>
+                <div class="feature-title">懂时尚</div>
+                <div>穿搭顾问 + 轻松对比</div>
             </div>
             """,
             unsafe_allow_html=True,
@@ -559,6 +597,14 @@ def main():
             st.session_state.uploaded_image = None
             st.rerun()
+        # Debug toggle
+        st.markdown("---")
+        st.checkbox(
+            "显示调试过程 (debug)",
+            key="show_debug",
+            help="展开后可查看中间思考过程及工具调用详情",
+        )
+
         st.markdown("---")
         st.caption(f"Session: `{st.session_state.session_id[:8]}...`")
@@ -696,6 +742,7 @@ What are you looking for today?&quot;&quot;&quot;
                     "role": "assistant",
                     "content": response,
                     "tool_calls": tool_calls,
+                    "debug_steps": result.get("debug_steps", []),
                 }
             )
@@ -78,38 +78,34 @@ class ShoppingAgent:
         """Build the LangGraph StateGraph"""
         # System prompt for the agent
-        system_prompt = """You are an intelligent fashion shopping assistant. You can:
-1. Search for products by text description (use search_products)
-2. 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: use analyze_image_style first to understand the product, then use search_products with the extracted description
-- 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)
-
-Example:
+        system_prompt = """你是一位智能时尚购物助手，你可以：
+1. 根据文字描述搜索商品（使用 search_products）
+2. 分析图片风格和属性（使用 analyze_image_style）
+
+当用户咨询商品时：
+- 文字提问：直接使用 search_products 搜索
+- 图片上传：先用 analyze_image_style 理解商品，再用提取的描述调用 search_products 搜索
+- 可按需连续调用多个工具
+- 始终保持有用、友好的回复风格
+
+关键格式规则：
+展示商品结果时，每个商品必须严格按以下格式输出：
+
+1. [标题 title]
+   ID: [商品ID]
+   分类: [category_path]
+   中文名: [title_cn]（如有）
+   标签: [tags]（如有）
+
+示例：
 1. Puma Men White 3/4 Length Pants
    ID: 12345
-   Category: Apparel > Bottomwear > Track Pants
-   Color: White
-   Gender: Men
-   Season: Summer
-   Usage: Sports
-   Relevance: 95.2%
+   分类: 服饰 > 裤装 > 运动裤
+   中文名: 彪马男士白色九分运动裤
+   标签: 运动,夏季,白色
-DO NOT skip the ID field! It is essential for displaying product images.
-Be conversational in your introduction, but preserve the exact product format."""
+不可省略 ID 字段！它是展示商品图片的关键。
+介绍要口语化，但必须保持上述商品格式。"""
         def agent_node(state: AgentState):
             """Agent decision node - decides which tools to call or when to respond"""
@@ -165,7 +161,9 @@ Be conversational in your introduction, but preserve the exact product format.&quot;&quot;
             image_path: Optional path to uploaded image
         Returns:
-            Dict with response and metadata
+            Dict with response and metadata, including:
+                - tool_calls: list of tool calls with args and (truncated) results
+                - debug_steps: detailed intermediate reasoning & tool execution steps
         """
         try:
             logger.info(
@@ -191,32 +189,74 @@ Be conversational in your introduction, but preserve the exact product format.&quot;&quot;
                 "current_image_path": image_path,
             }
-            # Track tool calls
+            # Track tool calls (high-level) and detailed debug steps
             tool_calls = []
+            debug_steps = []
-            # Stream events to capture tool calls
+            # Stream events to capture tool calls and intermediate reasoning
             for event in self.graph.stream(input_state, config=config):
                 logger.info(f"Event: {event}")
-                
-                # Check for agent node (tool calls)
+
+                # Agent node: LLM reasoning & tool decisions
                 if "agent" in event:
                     agent_output = event["agent"]
-                    if "messages" in agent_output:
-                        for msg in agent_output["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", {}),
-                                    })
-                
-                # Check for tool node (tool results)
+                    messages = agent_output.get("messages", [])
+
+                    step_messages = []
+                    step_tool_calls = []
+
+                    for msg in messages:
+                        msg_text = _extract_message_text(msg)
+                        msg_entry = {
+                            "type": getattr(msg, "type", "assistant"),
+                            "content": msg_text[:500],  # truncate for safety
+                        }
+                        step_messages.append(msg_entry)
+
+                        # Capture tool calls from this agent message
+                        if hasattr(msg, "tool_calls") and msg.tool_calls:
+                            for tc in msg.tool_calls:
+                                tc_entry = {
+                                    "name": tc.get("name"),
+                                    "args": tc.get("args", {}),
+                                }
+                                tool_calls.append(tc_entry)
+                                step_tool_calls.append(tc_entry)
+
+                    debug_steps.append(
+                        {
+                            "node": "agent",
+                            "messages": step_messages,
+                            "tool_calls": step_tool_calls,
+                        }
+                    )
+
+                # Tool node: actual tool execution results
                 if "tools" in event:
                     tools_output = event["tools"]
-                    if "messages" in tools_output:
-                        for i, msg in enumerate(tools_output["messages"]):
-                            if i < len(tool_calls):
-                                tool_calls[i]["result"] = str(msg.content)[:200] + "..."
+                    messages = tools_output.get("messages", [])
+
+                    step_tool_results = []
+
+                    for i, msg in enumerate(messages):
+                        content_text = _extract_message_text(msg)
+                        result_preview = content_text[:500] + ("..." if len(content_text) > 500 else "")
+
+                        if i < len(tool_calls):
+                            tool_calls[i]["result"] = result_preview
+
+                        step_tool_results.append(
+                            {
+                                "content": result_preview,
+                            }
+                        )
+
+                    debug_steps.append(
+                        {
+                            "node": "tools",
+                            "results": step_tool_results,
+                        }
+                    )
             # Get final state
             final_state = self.graph.get_state(config)
@@ -228,6 +268,7 @@ Be conversational in your introduction, but preserve the exact product format.&quot;&quot;
             return {
                 "response": response_text,
                 "tool_calls": tool_calls,
+                "debug_steps": debug_steps,
                 "error": False,
             }
@@ -0,0 +1,89 @@
+import json
+from typing import Optional
+
+from app.agents.shopping_agent import ShoppingAgent
+
+
+def run_once(agent: ShoppingAgent, query: str, image_path: Optional[str] = None) -> None:
+    """Run a single query through the agent and pretty-print details."""
+    result = agent.chat(query=query, image_path=image_path)
+
+    print("\n=== Assistant Response ===")
+    print(result.get("response", ""))
+
+    print("\n=== Tool Calls ===")
+    tool_calls = result.get("tool_calls", []) or []
+    if not tool_calls:
+        print("(no tool calls)")
+    else:
+        for i, tc in enumerate(tool_calls, 1):
+            print(f"[{i}] {tc.get('name')}")
+            print("  args:")
+            print("  " + json.dumps(tc.get("args", {}), ensure_ascii=False, indent=2).replace("\n", "\n  "))
+            if "result" in tc:
+                print("  result (truncated):")
+                print("  " + str(tc.get("result")).replace("\n", "\n  "))
+
+    print("\n=== Debug Steps ===")
+    debug_steps = result.get("debug_steps", []) or []
+    if not debug_steps:
+        print("(no debug steps)")
+    else:
+        for idx, step in enumerate(debug_steps, 1):
+            node = step.get("node", "unknown")
+            print(f"\n--- Step {idx} [{node}] ---")
+
+            if node == "agent":
+                msgs = step.get("messages", []) or []
+                if msgs:
+                    print("  Agent messages:")
+                    for m in msgs:
+                        role = m.get("type", "assistant")
+                        content = m.get("content", "")
+                        print(f"    - {role}: {content}")
+
+                tcs = step.get("tool_calls", []) or []
+                if tcs:
+                    print("  Planned tool calls:")
+                    for j, tc in enumerate(tcs, 1):
+                        print(f"    [{j}] {tc.get('name')}")
+                        print(
+                            "      args: "
+                            + json.dumps(tc.get("args", {}), ensure_ascii=False)
+                        )
+
+            elif node == "tools":
+                results = step.get("results", []) or []
+                if results:
+                    print("  Tool results:")
+                    for j, r in enumerate(results, 1):
+                        content = r.get("content", "")
+                        print(f"    [{j}] {content}")
+
+
+def main() -> None:
+    """Simple CLI debugger to inspect agent reasoning and tool usage."""
+    agent = ShoppingAgent(session_id="cli-debug")
+    print("ShopAgent CLI Debugger")
+    print("输入你的问题，或者输入 `exit` 退出。\n")
+
+    while True:
+        try:
+            query = input("你：").strip()
+        except (EOFError, KeyboardInterrupt):
+            print("\n再见 👋")
+            break
+
+        if not query:
+            continue
+        if query.lower() in {"exit", "quit"}:
+            print("再见 👋")
+            break
+
+        run_once(agent, query=query)
+        print("\n" + "=" * 60 + "\n")
+
+
+if __name__ == "__main__":
+    main()
+