feat: multimodal RAG with PDF image extraction and display

Extract images from PDFs using pymupdf, generate descriptions via
Qwen3-VL-8B, store in ChromaDB alongside text chunks, and render
images in chat answers. Includes image proxy rewrite, force re-process
endpoint, and VLM API timeout.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/src/llm/siliconflow.py

@@ -1,7 +1,8 @@
 """
-大模型API集成 — 支持 SiliconFlow 和 DeepSeek 官方
+大模型API集成 — 支持 SiliconFlow 和 DeepSeek 官方 + 视觉模型
 """
 import os
+import base64
 from typing import List, Dict, Any, Optional, AsyncGenerator, Tuple
 from langchain_openai import ChatOpenAI
 from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, SystemMessage
@@ -13,6 +14,17 @@ from ..core.config import get_settings
 
 settings = get_settings()
 
+# 图片描述提示词
+IMAGE_DESCRIPTION_PROMPT = """你是一个国土空间规划专家。请详细描述这张PDF文档中的图片内容。
+
+图片周围文字上下文（来自PDF页面）：{context_text}
+
+要求：
+1. 说明图片类型（地图/规划图/图表/流程图/示意图/照片等）
+2. 描述图片中的关键信息、数据和空间关系
+3. 提取图中所有文字标注
+4. 描述控制在200-300字"""
+
 # DeepSeek 官方模型 ID 前缀（用于自动路由）
 DEEPSEEK_OFFICIAL_MODELS = {
     "deepseek-chat",
@@ -150,6 +162,61 @@ class SiliconFlowLLM:
         return messages
 
 
+    async def describe_image(self, image_path: str, context_text: str = "") -> str:
+        """使用VLM模型描述图片内容
+
+        Args:
+            image_path: 图片文件路径
+            context_text: 图片周围的PDF文本上下文
+
+        Returns:
+            图片的文字描述
+        """
+        import asyncio
+        import time
+
+        # 读取图片并编码为base64
+        with open(image_path, "rb") as f:
+            image_data = base64.b64encode(f.read()).decode("utf-8")
+
+        # 检测图片格式
+        ext = os.path.splitext(image_path)[1].lower()
+        mime_map = {".png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg", ".gif": "image/gif", ".webp": "image/webp"}
+        mime_type = mime_map.get(ext, "image/png")
+
+        prompt = IMAGE_DESCRIPTION_PROMPT.format(context_text=context_text[:600])
+
+        vision_model = "Qwen/Qwen3-VL-8B-Instruct"
+        api_key, base_url, _ = _resolve_provider(vision_model)
+
+        client = openai.AsyncOpenAI(api_key=api_key, base_url=base_url)
+
+        max_retries = 3
+        for attempt in range(max_retries):
+            try:
+                response = await client.chat.completions.create(
+                    model=vision_model,
+                    messages=[{
+                        "role": "user",
+                        "content": [
+                            {"type": "text", "text": prompt},
+                            {"type": "image_url", "image_url": {"url": f"data:{mime_type};base64,{image_data}"}},
+                        ],
+                    }],
+                    max_tokens=600,
+                    temperature=0.3,
+                    timeout=90.0,
+                )
+                return response.choices[0].message.content or ""
+
+            except Exception as e:
+                print(f"[VLM] 描述失败 attempt={attempt+1}: {e}")
+                if attempt < max_retries - 1:
+                    await asyncio.sleep(2 ** attempt)
+
+        return ""
+
+
 # 全局LLM实例（使用默认模型）
 llm_client = SiliconFlowLLM()